s42chen commited on
Commit
cb8a01d
·
verified ·
1 Parent(s): 539ea65

Add files using upload-large-folder tool

Browse files
Files changed (50) hide show
  1. README.md +78 -29
  2. shared/package-files/06/0613f491d3a8fe4e53b8d84ca3a42bea37b110745c0e98be5308e924fb3db0dc/fused-score.log +22 -0
  3. shared/package-files/06/0632869a92c7cdd74d3ff751604c24a2d0fd252d866be4143125b2e424992c3f/eval-score-N5-recur-ms10-p96.log +41 -0
  4. shared/package-files/06/0674003ec6726a71037cf4cb852155f0c2b4a244d97f1f205e90edc3a8657464/model_config.json +1108 -0
  5. shared/package-files/0b/0b257defdfda2eed651ae468799ba3c865da5f9b1e998c87f12202905d332c67/consensus_subset.py +70 -0
  6. shared/package-files/0b/0b2938300f3af759ecaf7b1bc27fa153b05d5eff63eaca867867c5359e8cdeb5/train.py +238 -0
  7. shared/package-files/0b/0b543c06ef5862083df44fe801d381850637714742e2f1a3807c4424e0201cdf/e74-na-as-h.sh +7 -0
  8. shared/package-files/0b/0b721b94afb29ec79be5ec0608f2cc2328babf9a1e904f7ce7ca02e50cf8b68e/fuselab.py +105 -0
  9. shared/package-files/0b/0be5bef1b6f13ac83bf2b7f4e0e46ef560c4abe029983b41a2a3f426fbc257d8/cycle.py +334 -0
  10. shared/package-files/26/267e0693fa268f2bfb165485332ee52eaf290e5da9280077acc80b5a34ff97d2/netprop-eval-cap08.log +6 -0
  11. shared/package-files/2c/2c1d1d1556dcc6cd38eabf319f9a059e6b4f0a73b84d64cfe6ea5b09eb9c385d/eval-score-I30-on-merged-10-freq.log +41 -0
  12. shared/package-files/2c/2c31cd541a7add3868008214c8eaadc44684525c7d58849163797d9a6317c66f/rank_keep.py +76 -0
  13. shared/package-files/2c/2cb8e5cf99660ebdb7e2de530ffa2b0355e1d28cfdc9a5ff91f0dca005adfef3/B45a-anneal6k.log +10 -0
  14. shared/package-files/38/38a8667ef1302bb086a90b575ea6e9db125097721da1b0ff0210243969fedf30/fullfold-5Ms-I9b-L.report.json +123 -0
  15. shared/package-files/38/38ed421583fc85afbabc6ad85c91cec0d699f9c2c327af4236588876936bac6c/eval-I3-50k-100k.log +4 -0
  16. shared/package-files/3a/3a37da7fa50fabb4093d8876f075fcf95fb76c68b4cc2c5c3435c592f4016061/fullfold-5Ms-I9b-L.sh +13 -0
  17. shared/package-files/3a/3a3b4478d48856473c3dbd30b55c98a2c401226853ab7182d40bd9387deb7f41/report-ind.json +107 -0
  18. shared/package-files/3a/3a94f027d9c7dc0bff6c6f3f083e12827079f68fe14936be6322b7dbae19fed3/swap-I16+I18+I20-w1.0_0.05_0.5.log +8 -0
  19. shared/package-files/3c/3c02d5172d26f1587777ace6f7dacea73daab5aaeb92cea81a5670c43cde5cef/netprop-released-sub-0_25-2.log +4 -0
  20. shared/package-files/3c/3c140bc0e96a66db3c0d7451b4fda9cde77161967797935f6fb8dd197679a5a9/eval-score-I12.log +41 -0
  21. shared/package-files/3c/3ca1aa2268d6a4235b90b917bd4ec950f2712c826e93d39952b9ba05e725c5db/library-reverse-small-znw_8.log +10 -0
  22. shared/package-files/56/5606ac500db6e746b0790eaf33948de692ddb40e206e2a256478df7b39fe0160/validation.jsonl +10 -0
  23. shared/package-files/56/56382d924fa83294fbf0719351015cf3cd417b347da8c0754849c9ed485ed27d/stack_ceiling.py +130 -0
  24. shared/package-files/56/565d1106b459cf127e3695d8351dab7b2c257a2e849af804956e688275602ac4/report-np075-val-sub-replay.json +107 -0
  25. shared/package-files/56/56d8f62123d1d469f6fba33cb1d2528e433075d82052bd9ce09172670c665e98/matched-B3-30k.log +8 -0
  26. shared/package-files/75/751915c72e387c626ac9a076693ea766f19ade5559e92addc37cec166f8c5f9d/train.py +371 -0
  27. shared/package-files/75/75a4f93f95b85ff5388ef21c1eeaa788e8ece72ef666c843530028f7c82eabad/e73-store-real-validation.log +20 -0
  28. shared/package-files/7c/7cf300e8d128316f3531a381d6e0cf81481efe9217b1c01a28a078b8628e2611/eval-score-N5-recur-mop-10.log +41 -0
  29. shared/package-files/93/93355d7b360441f5ecd9b3cb24b5c389ef9a69d901c4f0f60c5cd3ca863ace0d/inv-pool-strata.log +12 -0
  30. shared/package-files/93/93356a643cbdd9fba86186125a89b55bda768140821198c6378c5aef321ac800/eval-score-I16+I17.log +41 -0
  31. shared/package-files/93/934b804bda1e0387c4be73816a4bfcef9dda4e1f50a3af300746b49a84e4c801/dataset_all.py +77 -0
  32. shared/package-files/93/9352c4e7d7af5238ccfb70c06597c1e117aa2489a28a318dccc5e459fe864cfc/replay-B20k+B30k-I1-50k-val-sub.log +18 -0
  33. shared/package-files/93/93e3bd183a2c73f60c7d014cc0ba3f6c0e33353d079873b4416ef58a826a3aee/ce-jitter-fused-B30k.log +13 -0
  34. shared/package-files/a5/a535d77b8ebc5ee1cd2b998a0eae5cd64daea829ff52d893e59d012178cc519d/publish.py +226 -0
  35. shared/package-files/b6/b63740294886c81680e5a0189d2abc271cd9d7726978ea72eb962bad61eb5d32/model_config.json +610 -0
  36. shared/package-files/b6/b665f6a2fef711b478e6d456fccce87c974a50619a8d892d36877cff76f550fb/anchor-B2-30k-v1.log +16 -0
  37. shared/package-files/b6/b672973a2274bbae420e6ec878372a79c2cb4916874eb9a6436bc7dd9292db77/eval-A10k-sub.log +11 -0
  38. shared/package-files/b6/b6da4868833e77578579a777da141035ab34a9f6aa4814e9233c49b7697c721d/netprop-oracle-rerank.log +5 -0
  39. shared/package-files/b9/b9052f01862eff09dd56dcd554fe689280543be8842bdb1fbbfc31cd4651f67d/consensus-subset-5Ms-N5u48-L-ion-net2.log +14 -0
  40. shared/package-files/b9/b90cf50c87342676bbe09c8ae51cac8f1e1bdf71c9feefd8e3b49ed81e3d8435/eval-B3-20k30k-full.log +46 -0
  41. shared/package-files/b9/b9694748013b9646c614eb704e331f78559c87cce1b9ed50225c4135bcc38eb9/e74-term-union-hyb-both.log +12 -0
  42. shared/package-files/b9/b98c577aafa80ccc1c381a527016ca758c6239da8627aba8bfb5a6bde98a3876/na_ion_types.py +56 -0
  43. shared/package-files/b9/b9f06e0872fa67848191be9925014bb305879bf9c389875d28f68a13f145a45a/recur-mof-5.log +4 -0
  44. shared/package-files/cb/cb1d041f6c7ed56984943cbd6808c4525c639a0326a92a31c24b585fe260ee53/eval-B10k-sub-hce10.log +7 -0
  45. shared/package-files/e6/e614f3e10a43efafcb939209fd868bbdc2c1c38b8b8cb0ec8082fa0750438a87/eval-score-I9b+I13.log +41 -0
  46. shared/package-files/e6/e648ff9e2e19f94d3b1ab306363508326ffb239560531b291ee3c5e1d6e65b62/train-I1.log +17 -0
  47. shared/package-files/e6/e6a6e6659a649411fa75007439d647bdf23583d2c21903147bb73d17300d2669/model_manifest.json +13 -0
  48. shared/package-files/ef/ef251257843f98236987449f06d2bcc25c03935762f8f88980d317dc4dfb85af/swap-recompare-N5-recur-hyb-s10-f5-p48-w0.75_0.1_0.5.log +8 -0
  49. shared/package-files/ef/efbf894e84d19de3b41b6f12aa3ad68b9e2229115175b9403d0f89821c7b034d/eval_blocks.py +105 -0
  50. shared/package-files/ef/efed6ca2099a4feb065029ab8f1e0c6676887395b891b8cfc8858597e71ca195/eval-score-I16.log +41 -0
README.md CHANGED
@@ -1,45 +1,65 @@
1
  ---
2
- pretty_name: Science Speedrun Astra research artifacts
3
  tags:
4
  - scientific-research
5
  - agent-evaluation
6
  - research-artifacts
7
  ---
8
 
9
- # Science Speedrun: 21 completed Astra task histories
10
 
11
- Reproducible submitted solutions, native evaluation reports, and screened Codex logs from
12
- GPT-6 Astra with xhigh reasoning. These are file artifacts, not a tabular training dataset.
13
- The histories contain 540 inference-distinct candidates; each has a completed
 
 
14
  sealed evaluation. Existing additional measurements of duplicate packages are retained.
15
  `reviews/fairness-20261031/` holds the operator-side fairness review of every first-round
16
  trajectory (per-task verdicts and the cross-cutting infrastructure findings); each first-round
17
  index carries its verdict under `operator_fairness_review`, and a fresh trial's index names
18
  the first trial it repeats under `fresh_trial_of`.
19
 
20
- | Task | Distinct candidates | Artifacts |
21
- |---|---:|---|
22
- | atomistic-materials | 3 | [Files](tasks/atomistic-materials/atomistic-materials-astra-xhigh-1h100-48h-20261020b) |
23
- | brain-computer-interface-decoding | 127 | [Files](tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a) |
24
- | crystal-structure-generation | 20 | [Files](tasks/crystal-structure-generation/crystal-structure-generation-astra-xhigh-1h100-48h-20261020b) |
25
- | earth-observation-pretraining-transfer | 11 | [Files](tasks/earth-observation-pretraining-transfer/earth-astra-xhigh-48h-20261020a) |
26
- | fluid-flow-control | 11 | [Files](tasks/fluid-flow-control/fluid-flow-control-astra-xhigh-1h100-48h-20261022a) |
27
- | fruit-fly-neuron-reconstruction | 7 | [Files](tasks/fruit-fly-neuron-reconstruction/fruit-fly-neuron-reconstruction-astra-xhigh-1h100-48h-20261022a) |
28
- | genome-sequence-to-function | 33 | [Files](tasks/genome-sequence-to-function/genome-sequence-to-function-astra-xhigh-1h100-48h-20261022a) |
29
- | mass-spectrometry-molecule-retrieval | 21 | [Files](tasks/mass-spectrometry-molecule-retrieval/mass-spectrometry-molecule-retrieval-astra-xhigh-1h100-48h-20261020b) |
30
- | molecular-pretraining-transfer | 14 | [Files](tasks/molecular-pretraining-transfer/molecular-pretraining-transfer-astra-xhigh-1h100-48h-20261023a) |
31
- | natural-disease-progression | 13 | [Files](tasks/natural-disease-progression/ndp-astra-xhigh-48h-20261020a) |
32
- | pde-surrogate | 60 | [Files](tasks/pde-surrogate/pde-surrogate-astra-xhigh-1h100-48h-20261020b) |
33
- | probabilistic-global-weather | 9 | [Files](tasks/probabilistic-global-weather/probabilistic-global-weather-astra-xhigh-1h100-48h-20261024f) |
34
- | protein-design | 1 | [Files](tasks/protein-design/protein-design-astra-xhigh-1h100-48h-20261022a) · superseded by the corrected trial |
35
- | protein-design | 3 | [Files](tasks/protein-design/protein-design-astra-xhigh-1h100-48h-20261027a) |
36
- | protein-folding-prediction | 18 | [Files](tasks/protein-folding-prediction/protein-folding-prediction-astra-xhigh-1h100-48h-20261020b) |
37
- | protein-sequence-modeling | 58 | [Files](tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261022a) |
38
- | protein-sequence-modeling | 9 | [Files](tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a) · fresh trial after the fairness review |
39
- | single-cell-perturbation-prediction | 17 | [Files](tasks/single-cell-perturbation-prediction/single-cell-perturbation-prediction-astra-xhigh-1h100-48h-20261020b) |
40
- | surface-code-decoding | 45 | [Files](tasks/surface-code-decoding/surface-code-decoding-astra-xhigh-1h100-48h-20261020b) |
41
- | timeseries-forecasting | 53 | [Files](tasks/timeseries-forecasting/timeseries-forecasting-astra-xhigh-1h100-48h-20261022a) |
42
- | tokamak-magnetic-control | 7 | [Files](tasks/tokamak-magnetic-control/tokamak-astra-xhigh-48h-20261020a) |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43
 
44
  ## Layout
45
 
@@ -110,10 +130,39 @@ implemented equation operators declared in its manifest; this is permitted by th
110
  contract and the scores stand, but from that point the reward mostly measures re-derivation
111
  of the governing equations rather than a learned surrogate.
112
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
113
  ## Provenance and excluded material
114
 
115
  Submissions were produced by the evaluated agent; reports by the native scientific
116
- scorers under the recorded adapters; logs by Codex and the harness. Organization,
117
  indexes, plots, and the reconstruction helper were created by the operator.
118
  Logs are screened for credentials, with redaction counts and original/exported hashes
119
  recorded in the indexes. Authentication stores, private test inputs/labels, raw research
 
1
  ---
2
+ pretty_name: Science Speedrun research artifacts
3
  tags:
4
  - scientific-research
5
  - agent-evaluation
6
  - research-artifacts
7
  ---
8
 
9
+ # Science Speedrun: 39 completed task histories
10
 
11
+ Reproducible submitted solutions, native evaluation reports, and screened agent logs from
12
+ GPT-6 Astra (Codex, xhigh reasoning; 21 histories) and from Claude Code
13
+ (Fable 5.2 requested, xhigh effort; 18 histories, see "Claude Code runs" below). These are
14
+ file artifacts, not a tabular training dataset.
15
+ The histories contain 832 inference-distinct candidates; each has a completed
16
  sealed evaluation. Existing additional measurements of duplicate packages are retained.
17
  `reviews/fairness-20261031/` holds the operator-side fairness review of every first-round
18
  trajectory (per-task verdicts and the cross-cutting infrastructure findings); each first-round
19
  index carries its verdict under `operator_fairness_review`, and a fresh trial's index names
20
  the first trial it repeats under `fresh_trial_of`.
21
 
22
+ | Task | Agent · model | Distinct candidates | Artifacts |
23
+ |---|---|---:|---|
24
+ | atomistic-materials | Codex · gpt-6-astra | 3 | [Files](tasks/atomistic-materials/atomistic-materials-astra-xhigh-1h100-48h-20261020b) |
25
+ | atomistic-materials | Claude Code · claude-fable-5-2 | 11 | [Files](tasks/atomistic-materials/atomistic-materials-fable52-xhigh-1h100-48h-20261110a) |
26
+ | brain-computer-interface-decoding | Codex · gpt-6-astra | 127 | [Files](tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a) |
27
+ | brain-computer-interface-decoding | Claude Code · claude-fable-5-2 | 14 | [Files](tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-fable52-xhigh-1h100-48h-20261110a) |
28
+ | crystal-structure-generation | Codex · gpt-6-astra | 20 | [Files](tasks/crystal-structure-generation/crystal-structure-generation-astra-xhigh-1h100-48h-20261020b) |
29
+ | earth-observation-pretraining-transfer | Codex · gpt-6-astra | 11 | [Files](tasks/earth-observation-pretraining-transfer/earth-astra-xhigh-48h-20261020a) |
30
+ | earth-observation-pretraining-transfer | Claude Code · claude-fable-5-2 | 5 | [Files](tasks/earth-observation-pretraining-transfer/earth-observation-pretraining-transfer-fable52-xhigh-1h100-48h-20261110b) |
31
+ | fluid-flow-control | Codex · gpt-6-astra | 11 | [Files](tasks/fluid-flow-control/fluid-flow-control-astra-xhigh-1h100-48h-20261022a) |
32
+ | fluid-flow-control | Claude Code · claude-fable-5-2 | 12 | [Files](tasks/fluid-flow-control/fluid-flow-control-fable52-xhigh-1h100-48h-20261110c) |
33
+ | fruit-fly-neuron-reconstruction | Codex · gpt-6-astra | 7 | [Files](tasks/fruit-fly-neuron-reconstruction/fruit-fly-neuron-reconstruction-astra-xhigh-1h100-48h-20261022a) |
34
+ | fruit-fly-neuron-reconstruction | Claude Code · claude-fable-5-2 | 8 | [Files](tasks/fruit-fly-neuron-reconstruction/fruit-fly-neuron-reconstruction-fable52-xhigh-1h100-48h-20261110a) |
35
+ | genome-sequence-to-function | Codex · gpt-6-astra | 33 | [Files](tasks/genome-sequence-to-function/genome-sequence-to-function-astra-xhigh-1h100-48h-20261022a) |
36
+ | genome-sequence-to-function | Claude Code · claude-fable-5-2 | 26 | [Files](tasks/genome-sequence-to-function/genome-sequence-to-function-fable52-xhigh-1h100-48h-20261110a) |
37
+ | mass-spectrometry-molecule-retrieval | Codex · gpt-6-astra | 21 | [Files](tasks/mass-spectrometry-molecule-retrieval/mass-spectrometry-molecule-retrieval-astra-xhigh-1h100-48h-20261020b) |
38
+ | mass-spectrometry-molecule-retrieval | Claude Code · claude-fable-5-2 | 17 | [Files](tasks/mass-spectrometry-molecule-retrieval/mass-spectrometry-molecule-retrieval-fable52-xhigh-1h100-48h-20261110a) |
39
+ | molecular-pretraining-transfer | Codex · gpt-6-astra | 14 | [Files](tasks/molecular-pretraining-transfer/molecular-pretraining-transfer-astra-xhigh-1h100-48h-20261023a) |
40
+ | molecular-pretraining-transfer | Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (105 / 504 requests) | 15 | [Files](tasks/molecular-pretraining-transfer/molecular-pretraining-transfer-fable52-xhigh-1h100-48h-20261110a) |
41
+ | natural-disease-progression | Claude Code · claude-fable-5-2 | 29 | [Files](tasks/natural-disease-progression/natural-disease-progression-fable52-xhigh-1h100-48h-20261110b) |
42
+ | natural-disease-progression | Codex · gpt-6-astra | 13 | [Files](tasks/natural-disease-progression/ndp-astra-xhigh-48h-20261020a) |
43
+ | pde-surrogate | Codex · gpt-6-astra | 60 | [Files](tasks/pde-surrogate/pde-surrogate-astra-xhigh-1h100-48h-20261020b) |
44
+ | pde-surrogate | Claude Code · claude-fable-5-2 | 24 | [Files](tasks/pde-surrogate/pde-surrogate-fable52-xhigh-1h100-48h-20261110a) |
45
+ | probabilistic-global-weather | Codex · gpt-6-astra | 9 | [Files](tasks/probabilistic-global-weather/probabilistic-global-weather-astra-xhigh-1h100-48h-20261024f) |
46
+ | probabilistic-global-weather | Claude Code · claude-fable-5-2 | 4 | [Files](tasks/probabilistic-global-weather/probabilistic-global-weather-fable52-xhigh-1h100-48h-20261110a) |
47
+ | protein-design | Codex · gpt-6-astra | 1 | [Files](tasks/protein-design/protein-design-astra-xhigh-1h100-48h-20261022a) · superseded by the corrected trial |
48
+ | protein-design | Codex · gpt-6-astra | 3 | [Files](tasks/protein-design/protein-design-astra-xhigh-1h100-48h-20261027a) |
49
+ | protein-folding-prediction | Codex · gpt-6-astra | 18 | [Files](tasks/protein-folding-prediction/protein-folding-prediction-astra-xhigh-1h100-48h-20261020b) |
50
+ | protein-folding-prediction | Claude Code · claude-fable-5-2 | 12 | [Files](tasks/protein-folding-prediction/protein-folding-prediction-fable52-xhigh-1h100-48h-20261110a) |
51
+ | protein-sequence-modeling | Codex · gpt-6-astra | 58 | [Files](tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261022a) |
52
+ | protein-sequence-modeling | Codex · gpt-6-astra | 9 | [Files](tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a) · fresh trial after the fairness review |
53
+ | protein-sequence-modeling | Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (789 / 421 requests) | 16 | [Files](tasks/protein-sequence-modeling/protein-sequence-modeling-fable52-xhigh-1h100-48h-20261110a) |
54
+ | single-cell-perturbation-prediction | Codex · gpt-6-astra | 17 | [Files](tasks/single-cell-perturbation-prediction/single-cell-perturbation-prediction-astra-xhigh-1h100-48h-20261020b) |
55
+ | single-cell-perturbation-prediction | Claude Code · claude-fable-5-2 | 17 | [Files](tasks/single-cell-perturbation-prediction/single-cell-perturbation-prediction-fable52-xhigh-1h100-48h-20261110a) |
56
+ | surface-code-decoding | Codex · gpt-6-astra | 45 | [Files](tasks/surface-code-decoding/surface-code-decoding-astra-xhigh-1h100-48h-20261020b) |
57
+ | surface-code-decoding | Claude Code · claude-fable-5-2 | 16 | [Files](tasks/surface-code-decoding/surface-code-decoding-fable52-xhigh-1h100-48h-20261110a) |
58
+ | target-based-drug-screening | Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (10 / 876 requests) | 23 | [Files](tasks/target-based-drug-screening/target-based-drug-screening-fable52-xhigh-1h100-48h-20261110a) |
59
+ | timeseries-forecasting | Codex · gpt-6-astra | 53 | [Files](tasks/timeseries-forecasting/timeseries-forecasting-astra-xhigh-1h100-48h-20261022a) |
60
+ | timeseries-forecasting | Claude Code · claude-fable-5-2 | 33 | [Files](tasks/timeseries-forecasting/timeseries-forecasting-fable52-xhigh-1h100-48h-20261110a) |
61
+ | tokamak-magnetic-control | Codex · gpt-6-astra | 7 | [Files](tasks/tokamak-magnetic-control/tokamak-astra-xhigh-48h-20261020a) |
62
+ | tokamak-magnetic-control | Claude Code · claude-fable-5-2 | 10 | [Files](tasks/tokamak-magnetic-control/tokamak-magnetic-control-fable52-xhigh-1h100-48h-20261110b) |
63
 
64
  ## Layout
65
 
 
130
  contract and the scores stand, but from that point the reward mostly measures re-derivation
131
  of the governing equations rather than a learned surrogate.
132
 
133
+ ## Claude Code runs
134
+
135
+ Run folders named `<task>-fable52-xhigh-1h100-48h-2026111*` are Claude Code trials (agent
136
+ version 2.1.319, launched 2026-11-10): one fresh session per task with no prior work reused,
137
+ on the same images, data mounts and 48-hour one-H100 research budget as the Astra trials.
138
+ 18 of the 20 trials are included; a trial is added once every distinct snapshot of its
139
+ history has been scored. Target-based drug screening is the twentieth task; its Astra trial
140
+ is not in this archive yet.
141
+
142
+ The model requested was `claude-fable-5-2` at xhigh effort. In 3 included trials, API safeguard
143
+ refusals of Fable 5.2 requests made Claude Code fall back to `claude-opus-5` for the rest of
144
+ the session: Molecular Pretraining (within minutes of launch; 105 requests answered by Fable 5.2, 504 by Opus 5); Protein Sequence (about 11 hours after launch; 789 requests answered by Fable 5.2, 421 by Opus 5); Drug Screening (within minutes of launch; 10 requests answered by Fable 5.2, 876 by Opus 5). Read a run answered
145
+ mostly by Opus 5 as an Opus 5 result and the others as mixed-model runs; each index records
146
+ the counts under `model_fallback` and the table above shows them.
147
+ Each index also carries `agent`, `model_requested`, `scoring_hardware` and `caveats`.
148
+ Differences from the Astra runs are stated there: PDE surrogate snapshots were scored on one
149
+ A100 (allowed by the task; one snapshot scored on both differed by 1.4e-5); Protein Folding
150
+ and Molecular Pretraining snapshots were scored on one H100 after a matching eight-GPU
151
+ replay; Fruit Fly evaluations mount the H100 build of the trusted point-sampling library.
152
+ Two trials carry a completion correction, with the original failure records kept under
153
+ `logs/completion-correction/`: the launcher wrongly marked the finished Molecular trial
154
+ failed, and the Weather trial was stopped by its supervisor 430 s before the cutoff (its last
155
+ snapshot is accepted as final; the clock was not reset). Evaluations that were replayed
156
+ after a cluster interruption are listed per snapshot under `evaluation_replays`.
157
+ `logs/sessions/` holds Claude Code's own session transcripts, screened like the Codex logs.
158
+ `plots/claude-score-over-time.*` shows score over research time for the scored Claude Code
159
+ histories with single-score or few-metric tasks; its CSV files hold the plotted values and
160
+ the baseline references.
161
+
162
  ## Provenance and excluded material
163
 
164
  Submissions were produced by the evaluated agent; reports by the native scientific
165
+ scorers under the recorded adapters; logs by the agents (Codex, Claude Code) and the harness. Organization,
166
  indexes, plots, and the reconstruction helper were created by the operator.
167
  Logs are screened for credentials, with redaction counts and original/exported hashes
168
  recorded in the indexes. Authentication stores, private test inputs/labels, raw research
shared/package-files/06/0613f491d3a8fe4e53b8d84ca3a42bea37b110745c0e98be5308e924fb3db0dc/fused-score.log ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ released all=0.7481 dom=0.9089 non=0.5867 h5=0.8744 h20=0.9313 | diff=+0.0000 [+0.0000,+0.0000] | ind 0.6012 +0.0000 [+0.0000,+0.0000]
2
+ pools +0 -0 | macro 0.8379 +0.0000 [+0.0000,+0.0000] | without top 20 +0.0000 [+0.0000,+0.0000] | weight min(n,30) +0.0000 [+0.0000,+0.0000]
3
+ cos all=0.7481 dom=0.9089 non=0.5867 h5=0.8744 h20=0.9313 | diff=+0.0000 [+0.0000,+0.0000] | ind 0.6012 +0.0000 [+0.0000,+0.0000]
4
+ pools +0 -0 | macro 0.8379 +0.0000 [+0.0000,+0.0000] | without top 20 +0.0000 [+0.0000,+0.0000] | weight min(n,30) +0.0000 [+0.0000,+0.0000]
5
+ cos_pm all=0.7489 dom=0.9098 non=0.5874 h5=0.8740 h20=0.9313 | diff=+0.0008 [-0.0004,+0.0026] | ind 0.6023 +0.0011 [+0.0000,+0.0023]
6
+ pools +5 -2 | macro 0.8387 +0.0009 [-0.0006,+0.0024] | without top 20 +0.0009 [-0.0004,+0.0028] | weight min(n,30) +0.0010 [-0.0004,+0.0029]
7
+ pow05 all=0.7468 dom=0.9094 non=0.5836 h5=0.8738 h20=0.9333 | diff=-0.0013 [-0.0102,+0.0071] | ind 0.6052 +0.0040 [+0.0013,+0.0070]
8
+ pools +29 -27 | macro 0.8385 +0.0006 [-0.0038,+0.0050] | without top 20 -0.0014 [-0.0109,+0.0081] | weight min(n,30) -0.0012 [-0.0098,+0.0076]
9
+ pow060 all=0.7519 dom=0.9130 non=0.5902 h5=0.8724 h20=0.9300 | diff=+0.0038 [-0.0026,+0.0108] | ind 0.6059 +0.0047 [+0.0027,+0.0069]
10
+ pools +25 -15 | macro 0.8408 +0.0030 [-0.0009,+0.0068] | without top 20 +0.0043 [-0.0026,+0.0121] | weight min(n,30) +0.0040 [-0.0027,+0.0113]
11
+ pow070 all=0.7525 dom=0.9136 non=0.5907 h5=0.8728 h20=0.9292 | diff=+0.0044 [-0.0020,+0.0114] | ind 0.6053 +0.0041 [+0.0025,+0.0058]
12
+ pools +23 -11 | macro 0.8414 +0.0035 [+0.0003,+0.0071] | without top 20 +0.0049 [-0.0018,+0.0127] | weight min(n,30) +0.0046 [-0.0020,+0.0121]
13
+ pow05m all=0.7471 dom=0.9092 non=0.5843 h5=0.8728 h20=0.9315 | diff=-0.0010 [-0.0101,+0.0072] | ind 0.6044 +0.0032 [+0.0006,+0.0062]
14
+ pools +28 -23 | macro 0.8393 +0.0015 [-0.0027,+0.0056] | without top 20 -0.0012 [-0.0107,+0.0084] | weight min(n,30) -0.0009 [-0.0096,+0.0079]
15
+ pow05idf05 all=0.7508 dom=0.9123 non=0.5888 h5=0.8761 h20=0.9341 | diff=+0.0027 [-0.0070,+0.0115] | ind 0.6141 +0.0129 [+0.0096,+0.0162]
16
+ pools +41 -31 | macro 0.8408 +0.0030 [-0.0021,+0.0077] | without top 20 +0.0031 [-0.0075,+0.0131] | weight min(n,30) +0.0035 [-0.0060,+0.0130]
17
+ idf05 all=0.7462 dom=0.9102 non=0.5814 h5=0.8745 h20=0.9282 | diff=-0.0020 [-0.0096,+0.0056] | ind 0.6110 +0.0098 [+0.0063,+0.0133]
18
+ pools +23 -27 | macro 0.8367 -0.0012 [-0.0053,+0.0030] | without top 20 -0.0022 [-0.0105,+0.0063] | weight min(n,30) -0.0019 [-0.0093,+0.0059]
19
+ log05 all=0.7482 dom=0.9143 non=0.5814 h5=0.8730 h20=0.9302 | diff=+0.0001 [-0.0110,+0.0097] | ind 0.6087 +0.0075 [+0.0047,+0.0102]
20
+ pools +34 -25 | macro 0.8405 +0.0027 [-0.0018,+0.0074] | without top 20 +0.0047 [-0.0025,+0.0127] | weight min(n,30) +0.0026 [-0.0056,+0.0106]
21
+ log20 all=0.7510 dom=0.9127 non=0.5888 h5=0.8721 h20=0.9296 | diff=+0.0029 [-0.0050,+0.0099] | ind 0.6067 +0.0055 [+0.0035,+0.0076]
22
+ pools +26 -17 | macro 0.8405 +0.0027 [-0.0012,+0.0065] | without top 20 +0.0033 [-0.0051,+0.0115] | weight min(n,30) +0.0037 [-0.0037,+0.0111]
shared/package-files/06/0632869a92c7cdd74d3ff751604c24a2d0fd252d866be4143125b2e424992c3f/eval-score-N5-recur-ms10-p96.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 5 members, 442.1 s
2
+ cos pool rule all=0.6425 dom=0.8526 non=0.4316 h5=0.7435 macro=0.7478
3
+ cen pool rule all=0.6299 dom=0.8354 non=0.4235 h5=0.7426 macro=0.7227
4
+ cos_pm pool rule all=0.6465 dom=0.8533 non=0.4388 h5=0.7477 macro=0.7493
5
+ dot pool rule all=0.6526 dom=0.8366 non=0.4679 h5=0.7792 macro=0.7413
6
+ geo pool rule all=0.6324 dom=0.8468 non=0.4171 h5=0.7343 macro=0.7389
7
+ idf05 pool rule all=0.6545 dom=0.8601 non=0.4482 h5=0.7545 macro=0.7513
8
+ idf1 pool rule all=0.6587 dom=0.8591 non=0.4576 h5=0.7624 macro=0.7522
9
+ lin pool rule all=0.6148 dom=0.8227 non=0.4060 h5=0.7208 macro=0.7129
10
+ log01 pool rule all=0.6688 dom=0.8617 non=0.4752 h5=0.7809 macro=0.7472
11
+ log05 pool rule all=0.6823 dom=0.8598 non=0.5041 h5=0.7977 macro=0.7561
12
+ log20 pool rule all=0.6575 dom=0.8613 non=0.4528 h5=0.7614 macro=0.7581
13
+ nll02 pool rule all=0.5052 dom=0.7518 non=0.2577 h5=0.6442 macro=0.6190
14
+ nll10 pool rule all=0.4722 dom=0.7099 non=0.2335 h5=0.6081 macro=0.5709
15
+ poolcen pool rule all=0.6329 dom=0.8326 non=0.4323 h5=0.7640 macro=0.7230
16
+ pow020 pool rule all=0.6242 dom=0.8354 non=0.4121 h5=0.7698 macro=0.7312
17
+ pow030 pool rule all=0.6236 dom=0.8471 non=0.3992 h5=0.7780 macro=0.7357
18
+ pow040 pool rule all=0.6702 dom=0.8547 non=0.4850 h5=0.7916 macro=0.7478
19
+ pow05 pool rule all=0.6504 dom=0.8587 non=0.4413 h5=0.7645 macro=0.7531
20
+ pow05idf025 pool rule all=0.6556 dom=0.8620 non=0.4484 h5=0.7668 macro=0.7578
21
+ pow05idf05 pool rule all=0.6527 dom=0.8621 non=0.4425 h5=0.7667 macro=0.7572
22
+ pow05m pool rule all=0.6483 dom=0.8555 non=0.4403 h5=0.7561 macro=0.7519
23
+ pow05s pool rule all=0.6466 dom=0.8557 non=0.4366 h5=0.7637 macro=0.7519
24
+ pow060 pool rule all=0.6520 dom=0.8578 non=0.4454 h5=0.7642 macro=0.7558
25
+ pow060idf025 pool rule all=0.6568 dom=0.8618 non=0.4509 h5=0.7656 macro=0.7587
26
+ pow060idf050 pool rule all=0.6547 dom=0.8618 non=0.4467 h5=0.7687 macro=0.7584
27
+ pow060m pool rule all=0.6472 dom=0.8538 non=0.4397 h5=0.7585 macro=0.7531
28
+ pow065 pool rule all=0.6546 dom=0.8592 non=0.4493 h5=0.7640 macro=0.7566
29
+ pow070 pool rule all=0.6565 dom=0.8607 non=0.4516 h5=0.7624 macro=0.7575
30
+ pow070idf025 pool rule all=0.6547 dom=0.8609 non=0.4478 h5=0.7674 macro=0.7563
31
+ pow070idf050 pool rule all=0.6571 dom=0.8608 non=0.4525 h5=0.7673 macro=0.7563
32
+ pow070m pool rule all=0.6444 dom=0.8532 non=0.4348 h5=0.7549 macro=0.7528
33
+ pow075 pool rule all=0.6535 dom=0.8564 non=0.4497 h5=0.7584 macro=0.7569
34
+ pow080 pool rule all=0.6556 dom=0.8561 non=0.4543 h5=0.7556 macro=0.7569
35
+ pow080m pool rule all=0.6451 dom=0.8544 non=0.4350 h5=0.7516 macro=0.7522
36
+ pow085 pool rule all=0.6551 dom=0.8565 non=0.4528 h5=0.7560 macro=0.7563
37
+ pow085idf025 pool rule all=0.6531 dom=0.8608 non=0.4447 h5=0.7542 macro=0.7546
38
+ pow085idf050 pool rule all=0.6569 dom=0.8610 non=0.4520 h5=0.7596 macro=0.7549
39
+ pow090 pool rule all=0.6522 dom=0.8552 non=0.4484 h5=0.7482 macro=0.7540
40
+ pow2 pool rule all=0.6148 dom=0.8258 non=0.4029 h5=0.7367 macro=0.7221
41
+ tani pool rule all=0.6319 dom=0.8356 non=0.4274 h5=0.7477 macro=0.7377
shared/package-files/06/0674003ec6726a71037cf4cb852155f0c2b4a244d97f1f205e90edc3a8657464/model_config.json ADDED
@@ -0,0 +1,1108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "sparse_k": 100,
3
+ "inference_batch_size": 128,
4
+ "workers": 14,
5
+ "num_bins": 150000,
6
+ "upper_limit": 1500,
7
+ "max_observed_peaks": 50,
8
+ "observed_intensity_threshold": 0.001,
9
+ "ensemble": "prediction",
10
+ "pool_aggregation": 1.0,
11
+ "autocast": false,
12
+ "model": {
13
+ "pe_embed_k": 14,
14
+ "embed_elem_group": true,
15
+ "multi_hop_max_dist": 10
16
+ },
17
+ "checkpoint": "model.pt",
18
+ "views": [
19
+ {
20
+ "precursor_window": 0.5,
21
+ "pred_power": 0.75,
22
+ "match_ppm": 20,
23
+ "match_floor_mda": 4,
24
+ "match_by_instrument": {
25
+ "Orbitrap": [
26
+ 10,
27
+ 1
28
+ ]
29
+ }
30
+ }
31
+ ],
32
+ "inverse": {
33
+ "pooled": true,
34
+ "score": {
35
+ "members": "probability",
36
+ "power": 0.7
37
+ },
38
+ "weight_domain": 0.75,
39
+ "weight_other": 0.05,
40
+ "weight_other_single": 0.5,
41
+ "batch": 2048
42
+ },
43
+ "library": {
44
+ "kind": "sum",
45
+ "k": 20,
46
+ "top": 32,
47
+ "power": 6,
48
+ "debias": true,
49
+ "weight_domain": 0.0,
50
+ "weight_other": 0.4,
51
+ "weight_other_single": 0.0
52
+ },
53
+ "ion": {
54
+ "weight_domain": 0.05,
55
+ "weight_other": 0.01
56
+ },
57
+ "consensus": {
58
+ "k": 10,
59
+ "top": 32,
60
+ "power": 3.0,
61
+ "answers": 10,
62
+ "temperature": 0.02,
63
+ "absorb": 0.5,
64
+ "bin": 0.01,
65
+ "weight_domain": 1.0,
66
+ "weight_other": 1.0
67
+ },
68
+ "inverse_members": [
69
+ {
70
+ "source": "runs/I16-ff512x6-real08-corpus2+train3-noise05/step-0050000-ema.pt",
71
+ "sha256": "035b6418b0be40e11479e170e5c926b352bc56cd8ce823cdfa88169421c91ff1",
72
+ "step": 50000,
73
+ "class": "FormulaFormer",
74
+ "weights": "ema",
75
+ "initialization": "random",
76
+ "seed": 1,
77
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
78
+ "self_simulated_store": {
79
+ "path": "inv/selfsim/corpus2+train3.npz",
80
+ "rows": 2031301,
81
+ "manifest": {
82
+ "schema": "allcond-inverse-selfsim/v1",
83
+ "merged": [
84
+ "inv/selfsim/corpus2.npz",
85
+ "inv/selfsim/train3.npz"
86
+ ],
87
+ "parts": [
88
+ {
89
+ "schema": "allcond-inverse-selfsim/v1",
90
+ "merged": [
91
+ "inv/selfsim/sub75.npz",
92
+ "inv/selfsim/sub25.npz"
93
+ ],
94
+ "parts": [
95
+ {
96
+ "name": "sub75",
97
+ "source": "synth",
98
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
99
+ "bundle": "builds/5Ms-I1+I3x2-L",
100
+ "members": "1,3",
101
+ "structures": 435076,
102
+ "tasks": 651127,
103
+ "failed": 0,
104
+ "rows": 651127,
105
+ "seconds": 635.7592854518443,
106
+ "schema": "allcond-inverse-selfsim/v1",
107
+ "floor": 0.02,
108
+ "spectra": 651083,
109
+ "without_peak": 44,
110
+ "mean_annotated_peaks": 15.61580167198345,
111
+ "mean_candidates": 20.62178653319552,
112
+ "ppm": 20.0,
113
+ "floor_da": 0.004,
114
+ "max_peaks": 24,
115
+ "annotate_seconds": 25.62758301384747
116
+ },
117
+ {
118
+ "name": "sub25",
119
+ "source": "synth",
120
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
121
+ "bundle": "builds/5Ms-I1+I3x2-L",
122
+ "members": "1,3",
123
+ "structures": 435076,
124
+ "tasks": 217012,
125
+ "failed": 0,
126
+ "rows": 217012,
127
+ "seconds": 227.74617605702952,
128
+ "schema": "allcond-inverse-selfsim/v1",
129
+ "floor": 0.02,
130
+ "spectra": 217000,
131
+ "without_peak": 12,
132
+ "mean_annotated_peaks": 15.62026728110599,
133
+ "mean_candidates": 20.61567563084069,
134
+ "ppm": 20.0,
135
+ "floor_da": 0.004,
136
+ "max_peaks": 24,
137
+ "annotate_seconds": 10.8742212112993
138
+ }
139
+ ],
140
+ "spectra": 868083,
141
+ "structures": 434944
142
+ },
143
+ {
144
+ "name": "train3",
145
+ "source": "synth",
146
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
147
+ "bundle": "builds/5Ms-I1+I3x2-L",
148
+ "members": "1,3",
149
+ "structures": 435076,
150
+ "tasks": 1163260,
151
+ "failed": 0,
152
+ "rows": 1163260,
153
+ "seconds": 1262.926542488858,
154
+ "schema": "allcond-inverse-selfsim/v1",
155
+ "floor": 0.02,
156
+ "spectra": 1163218,
157
+ "without_peak": 42,
158
+ "mean_annotated_peaks": 17.868643710809152,
159
+ "mean_candidates": 25.376552103570997,
160
+ "ppm": 20.0,
161
+ "floor_da": 0.004,
162
+ "max_peaks": 24,
163
+ "annotate_seconds": 43.40271363919601
164
+ }
165
+ ],
166
+ "spectra": 2031301,
167
+ "structures": 434950
168
+ }
169
+ },
170
+ "stores": {
171
+ "gen-x1.manifest.json": null,
172
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
173
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
174
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
175
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
176
+ },
177
+ "recipe": {
178
+ "store": "/workspace/research/inv",
179
+ "out": "runs/I16-ff512x6-real08-corpus2+train3-noise05",
180
+ "pools": "pools-validation-input.npz",
181
+ "input": "/workspace/research/validation-input",
182
+ "steps": 50000,
183
+ "batch": 512,
184
+ "real_frac": 0.08,
185
+ "real_balance": 0.5,
186
+ "no_synth": false,
187
+ "target": "",
188
+ "heavy_boost": 1.0,
189
+ "synth_structures": 1.0,
190
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
191
+ "extra_frac": 0.5,
192
+ "heavy_lo": 450.0,
193
+ "heavy_hi": 850.0,
194
+ "lr": 0.0006,
195
+ "weight_decay": 0.0001,
196
+ "warmup": 2000,
197
+ "ema": 0.999,
198
+ "dim": 512,
199
+ "layers": 6,
200
+ "heads": 8,
201
+ "dropout": 0.1,
202
+ "peak_drop": 0.3,
203
+ "truncate": 0.3,
204
+ "min_keep": 4,
205
+ "jitter": 0.15,
206
+ "synth_energy_drop": 0.3,
207
+ "synth_instrument_drop": 0.3,
208
+ "validate_every": 5000,
209
+ "retrieve_every": 10000,
210
+ "retrieve_pools": 0,
211
+ "seed": 1,
212
+ "contrast_weight": 0.0,
213
+ "contrast_k": 31,
214
+ "contrast_tau": 0.05,
215
+ "noise_rate": 0.5,
216
+ "noise_max": 12,
217
+ "noise_scatter": 0.35,
218
+ "noise_shape": 3.0,
219
+ "init": null,
220
+ "max_steps_this_run": 0
221
+ }
222
+ },
223
+ {
224
+ "source": "runs/I18-ff512x6-real08-corpus2+train3-noise05-real025/step-0050000-ema.pt",
225
+ "sha256": "2eb34f936b8645b0eaabfcb2a483c2c88d6fb640937a89c66cda6a9def1ef267",
226
+ "step": 50000,
227
+ "class": "FormulaFormer",
228
+ "weights": "ema",
229
+ "initialization": "random",
230
+ "seed": 1,
231
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
232
+ "self_simulated_store": {
233
+ "path": "inv/selfsim/corpus2+train3.npz",
234
+ "rows": 2031301,
235
+ "manifest": {
236
+ "schema": "allcond-inverse-selfsim/v1",
237
+ "merged": [
238
+ "inv/selfsim/corpus2.npz",
239
+ "inv/selfsim/train3.npz"
240
+ ],
241
+ "parts": [
242
+ {
243
+ "schema": "allcond-inverse-selfsim/v1",
244
+ "merged": [
245
+ "inv/selfsim/sub75.npz",
246
+ "inv/selfsim/sub25.npz"
247
+ ],
248
+ "parts": [
249
+ {
250
+ "name": "sub75",
251
+ "source": "synth",
252
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
253
+ "bundle": "builds/5Ms-I1+I3x2-L",
254
+ "members": "1,3",
255
+ "structures": 435076,
256
+ "tasks": 651127,
257
+ "failed": 0,
258
+ "rows": 651127,
259
+ "seconds": 635.7592854518443,
260
+ "schema": "allcond-inverse-selfsim/v1",
261
+ "floor": 0.02,
262
+ "spectra": 651083,
263
+ "without_peak": 44,
264
+ "mean_annotated_peaks": 15.61580167198345,
265
+ "mean_candidates": 20.62178653319552,
266
+ "ppm": 20.0,
267
+ "floor_da": 0.004,
268
+ "max_peaks": 24,
269
+ "annotate_seconds": 25.62758301384747
270
+ },
271
+ {
272
+ "name": "sub25",
273
+ "source": "synth",
274
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
275
+ "bundle": "builds/5Ms-I1+I3x2-L",
276
+ "members": "1,3",
277
+ "structures": 435076,
278
+ "tasks": 217012,
279
+ "failed": 0,
280
+ "rows": 217012,
281
+ "seconds": 227.74617605702952,
282
+ "schema": "allcond-inverse-selfsim/v1",
283
+ "floor": 0.02,
284
+ "spectra": 217000,
285
+ "without_peak": 12,
286
+ "mean_annotated_peaks": 15.62026728110599,
287
+ "mean_candidates": 20.61567563084069,
288
+ "ppm": 20.0,
289
+ "floor_da": 0.004,
290
+ "max_peaks": 24,
291
+ "annotate_seconds": 10.8742212112993
292
+ }
293
+ ],
294
+ "spectra": 868083,
295
+ "structures": 434944
296
+ },
297
+ {
298
+ "name": "train3",
299
+ "source": "synth",
300
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
301
+ "bundle": "builds/5Ms-I1+I3x2-L",
302
+ "members": "1,3",
303
+ "structures": 435076,
304
+ "tasks": 1163260,
305
+ "failed": 0,
306
+ "rows": 1163260,
307
+ "seconds": 1262.926542488858,
308
+ "schema": "allcond-inverse-selfsim/v1",
309
+ "floor": 0.02,
310
+ "spectra": 1163218,
311
+ "without_peak": 42,
312
+ "mean_annotated_peaks": 17.868643710809152,
313
+ "mean_candidates": 25.376552103570997,
314
+ "ppm": 20.0,
315
+ "floor_da": 0.004,
316
+ "max_peaks": 24,
317
+ "annotate_seconds": 43.40271363919601
318
+ }
319
+ ],
320
+ "spectra": 2031301,
321
+ "structures": 434950
322
+ }
323
+ },
324
+ "stores": {
325
+ "gen-x1.manifest.json": null,
326
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
327
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
328
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
329
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
330
+ },
331
+ "recipe": {
332
+ "store": "/workspace/research/inv",
333
+ "out": "runs/I18-ff512x6-real08-corpus2+train3-noise05-real025",
334
+ "pools": "pools-validation-input.npz",
335
+ "input": "/workspace/research/validation-input",
336
+ "steps": 50000,
337
+ "batch": 512,
338
+ "real_frac": 0.08,
339
+ "real_balance": 0.5,
340
+ "no_synth": false,
341
+ "target": "",
342
+ "heavy_boost": 1.0,
343
+ "synth_structures": 1.0,
344
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
345
+ "extra_frac": 0.5,
346
+ "heavy_lo": 450.0,
347
+ "heavy_hi": 850.0,
348
+ "lr": 0.0006,
349
+ "weight_decay": 0.0001,
350
+ "warmup": 2000,
351
+ "ema": 0.999,
352
+ "dim": 512,
353
+ "layers": 6,
354
+ "heads": 8,
355
+ "dropout": 0.1,
356
+ "peak_drop": 0.3,
357
+ "truncate": 0.3,
358
+ "min_keep": 4,
359
+ "jitter": 0.15,
360
+ "synth_energy_drop": 0.3,
361
+ "synth_instrument_drop": 0.3,
362
+ "validate_every": 5000,
363
+ "retrieve_every": 10000,
364
+ "retrieve_pools": 0,
365
+ "seed": 1,
366
+ "contrast_weight": 0.0,
367
+ "contrast_k": 31,
368
+ "contrast_tau": 0.05,
369
+ "noise_rate": 0.5,
370
+ "noise_max": 12,
371
+ "noise_scatter": 0.35,
372
+ "noise_shape": 3.0,
373
+ "noise_real_rate": 0.25,
374
+ "synth_peak_drop": -1.0,
375
+ "init": null,
376
+ "max_steps_this_run": 0
377
+ }
378
+ },
379
+ {
380
+ "source": "runs/I19-ff512x6-real08-corpus2+train3-noise05-jit08/step-0050000-ema.pt",
381
+ "sha256": "7b95edd35fa25625b8ad091d901545d59b5257c95122f7912931317abbc42701",
382
+ "step": 50000,
383
+ "class": "FormulaFormer",
384
+ "weights": "ema",
385
+ "initialization": "random",
386
+ "seed": 1,
387
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
388
+ "self_simulated_store": {
389
+ "path": "inv/selfsim/corpus2+train3.npz",
390
+ "rows": 2031301,
391
+ "manifest": {
392
+ "schema": "allcond-inverse-selfsim/v1",
393
+ "merged": [
394
+ "inv/selfsim/corpus2.npz",
395
+ "inv/selfsim/train3.npz"
396
+ ],
397
+ "parts": [
398
+ {
399
+ "schema": "allcond-inverse-selfsim/v1",
400
+ "merged": [
401
+ "inv/selfsim/sub75.npz",
402
+ "inv/selfsim/sub25.npz"
403
+ ],
404
+ "parts": [
405
+ {
406
+ "name": "sub75",
407
+ "source": "synth",
408
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
409
+ "bundle": "builds/5Ms-I1+I3x2-L",
410
+ "members": "1,3",
411
+ "structures": 435076,
412
+ "tasks": 651127,
413
+ "failed": 0,
414
+ "rows": 651127,
415
+ "seconds": 635.7592854518443,
416
+ "schema": "allcond-inverse-selfsim/v1",
417
+ "floor": 0.02,
418
+ "spectra": 651083,
419
+ "without_peak": 44,
420
+ "mean_annotated_peaks": 15.61580167198345,
421
+ "mean_candidates": 20.62178653319552,
422
+ "ppm": 20.0,
423
+ "floor_da": 0.004,
424
+ "max_peaks": 24,
425
+ "annotate_seconds": 25.62758301384747
426
+ },
427
+ {
428
+ "name": "sub25",
429
+ "source": "synth",
430
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
431
+ "bundle": "builds/5Ms-I1+I3x2-L",
432
+ "members": "1,3",
433
+ "structures": 435076,
434
+ "tasks": 217012,
435
+ "failed": 0,
436
+ "rows": 217012,
437
+ "seconds": 227.74617605702952,
438
+ "schema": "allcond-inverse-selfsim/v1",
439
+ "floor": 0.02,
440
+ "spectra": 217000,
441
+ "without_peak": 12,
442
+ "mean_annotated_peaks": 15.62026728110599,
443
+ "mean_candidates": 20.61567563084069,
444
+ "ppm": 20.0,
445
+ "floor_da": 0.004,
446
+ "max_peaks": 24,
447
+ "annotate_seconds": 10.8742212112993
448
+ }
449
+ ],
450
+ "spectra": 868083,
451
+ "structures": 434944
452
+ },
453
+ {
454
+ "name": "train3",
455
+ "source": "synth",
456
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
457
+ "bundle": "builds/5Ms-I1+I3x2-L",
458
+ "members": "1,3",
459
+ "structures": 435076,
460
+ "tasks": 1163260,
461
+ "failed": 0,
462
+ "rows": 1163260,
463
+ "seconds": 1262.926542488858,
464
+ "schema": "allcond-inverse-selfsim/v1",
465
+ "floor": 0.02,
466
+ "spectra": 1163218,
467
+ "without_peak": 42,
468
+ "mean_annotated_peaks": 17.868643710809152,
469
+ "mean_candidates": 25.376552103570997,
470
+ "ppm": 20.0,
471
+ "floor_da": 0.004,
472
+ "max_peaks": 24,
473
+ "annotate_seconds": 43.40271363919601
474
+ }
475
+ ],
476
+ "spectra": 2031301,
477
+ "structures": 434950
478
+ }
479
+ },
480
+ "stores": {
481
+ "gen-x1.manifest.json": null,
482
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
483
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
484
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
485
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
486
+ },
487
+ "recipe": {
488
+ "store": "/workspace/research/inv",
489
+ "out": "runs/I19-ff512x6-real08-corpus2+train3-noise05-jit08",
490
+ "pools": "pools-validation-input.npz",
491
+ "input": "/workspace/research/validation-input",
492
+ "steps": 50000,
493
+ "batch": 512,
494
+ "real_frac": 0.08,
495
+ "real_balance": 0.5,
496
+ "no_synth": false,
497
+ "target": "",
498
+ "heavy_boost": 1.0,
499
+ "synth_structures": 1.0,
500
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
501
+ "extra_frac": 0.5,
502
+ "heavy_lo": 450.0,
503
+ "heavy_hi": 850.0,
504
+ "lr": 0.0006,
505
+ "weight_decay": 0.0001,
506
+ "warmup": 2000,
507
+ "ema": 0.999,
508
+ "dim": 512,
509
+ "layers": 6,
510
+ "heads": 8,
511
+ "dropout": 0.1,
512
+ "peak_drop": 0.3,
513
+ "truncate": 0.3,
514
+ "min_keep": 4,
515
+ "jitter": 0.15,
516
+ "synth_energy_drop": 0.3,
517
+ "synth_instrument_drop": 0.3,
518
+ "validate_every": 5000,
519
+ "retrieve_every": 10000,
520
+ "retrieve_pools": 0,
521
+ "seed": 1,
522
+ "contrast_weight": 0.0,
523
+ "contrast_k": 31,
524
+ "contrast_tau": 0.05,
525
+ "noise_rate": 0.5,
526
+ "noise_max": 12,
527
+ "noise_scatter": 0.35,
528
+ "noise_shape": 3.0,
529
+ "noise_real_rate": 0.0,
530
+ "synth_peak_drop": -1.0,
531
+ "synth_jitter": 0.8,
532
+ "noise_near": 0.0,
533
+ "noise_deltas": "inv/noise_deltas.json",
534
+ "init": null,
535
+ "max_steps_this_run": 0
536
+ }
537
+ },
538
+ {
539
+ "source": "runs/I20-ff512x6-real08-corpus2+train3-noise05-near067/step-0050000-ema.pt",
540
+ "sha256": "0e017ce27059d5215f38a8b16124408a8a2e3e61c4b86704751485e6a6f610de",
541
+ "step": 50000,
542
+ "class": "FormulaFormer",
543
+ "weights": "ema",
544
+ "initialization": "random",
545
+ "seed": 1,
546
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
547
+ "self_simulated_store": {
548
+ "path": "inv/selfsim/corpus2+train3.npz",
549
+ "rows": 2031301,
550
+ "manifest": {
551
+ "schema": "allcond-inverse-selfsim/v1",
552
+ "merged": [
553
+ "inv/selfsim/corpus2.npz",
554
+ "inv/selfsim/train3.npz"
555
+ ],
556
+ "parts": [
557
+ {
558
+ "schema": "allcond-inverse-selfsim/v1",
559
+ "merged": [
560
+ "inv/selfsim/sub75.npz",
561
+ "inv/selfsim/sub25.npz"
562
+ ],
563
+ "parts": [
564
+ {
565
+ "name": "sub75",
566
+ "source": "synth",
567
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
568
+ "bundle": "builds/5Ms-I1+I3x2-L",
569
+ "members": "1,3",
570
+ "structures": 435076,
571
+ "tasks": 651127,
572
+ "failed": 0,
573
+ "rows": 651127,
574
+ "seconds": 635.7592854518443,
575
+ "schema": "allcond-inverse-selfsim/v1",
576
+ "floor": 0.02,
577
+ "spectra": 651083,
578
+ "without_peak": 44,
579
+ "mean_annotated_peaks": 15.61580167198345,
580
+ "mean_candidates": 20.62178653319552,
581
+ "ppm": 20.0,
582
+ "floor_da": 0.004,
583
+ "max_peaks": 24,
584
+ "annotate_seconds": 25.62758301384747
585
+ },
586
+ {
587
+ "name": "sub25",
588
+ "source": "synth",
589
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
590
+ "bundle": "builds/5Ms-I1+I3x2-L",
591
+ "members": "1,3",
592
+ "structures": 435076,
593
+ "tasks": 217012,
594
+ "failed": 0,
595
+ "rows": 217012,
596
+ "seconds": 227.74617605702952,
597
+ "schema": "allcond-inverse-selfsim/v1",
598
+ "floor": 0.02,
599
+ "spectra": 217000,
600
+ "without_peak": 12,
601
+ "mean_annotated_peaks": 15.62026728110599,
602
+ "mean_candidates": 20.61567563084069,
603
+ "ppm": 20.0,
604
+ "floor_da": 0.004,
605
+ "max_peaks": 24,
606
+ "annotate_seconds": 10.8742212112993
607
+ }
608
+ ],
609
+ "spectra": 868083,
610
+ "structures": 434944
611
+ },
612
+ {
613
+ "name": "train3",
614
+ "source": "synth",
615
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
616
+ "bundle": "builds/5Ms-I1+I3x2-L",
617
+ "members": "1,3",
618
+ "structures": 435076,
619
+ "tasks": 1163260,
620
+ "failed": 0,
621
+ "rows": 1163260,
622
+ "seconds": 1262.926542488858,
623
+ "schema": "allcond-inverse-selfsim/v1",
624
+ "floor": 0.02,
625
+ "spectra": 1163218,
626
+ "without_peak": 42,
627
+ "mean_annotated_peaks": 17.868643710809152,
628
+ "mean_candidates": 25.376552103570997,
629
+ "ppm": 20.0,
630
+ "floor_da": 0.004,
631
+ "max_peaks": 24,
632
+ "annotate_seconds": 43.40271363919601
633
+ }
634
+ ],
635
+ "spectra": 2031301,
636
+ "structures": 434950
637
+ }
638
+ },
639
+ "stores": {
640
+ "gen-x1.manifest.json": null,
641
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
642
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
643
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
644
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
645
+ },
646
+ "recipe": {
647
+ "store": "/workspace/research/inv",
648
+ "out": "runs/I20-ff512x6-real08-corpus2+train3-noise05-near067",
649
+ "pools": "pools-validation-input.npz",
650
+ "input": "/workspace/research/validation-input",
651
+ "steps": 50000,
652
+ "batch": 512,
653
+ "real_frac": 0.08,
654
+ "real_balance": 0.5,
655
+ "no_synth": false,
656
+ "target": "",
657
+ "heavy_boost": 1.0,
658
+ "synth_structures": 1.0,
659
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
660
+ "extra_frac": 0.5,
661
+ "heavy_lo": 450.0,
662
+ "heavy_hi": 850.0,
663
+ "lr": 0.0006,
664
+ "weight_decay": 0.0001,
665
+ "warmup": 2000,
666
+ "ema": 0.999,
667
+ "dim": 512,
668
+ "layers": 6,
669
+ "heads": 8,
670
+ "dropout": 0.1,
671
+ "peak_drop": 0.3,
672
+ "truncate": 0.3,
673
+ "min_keep": 4,
674
+ "jitter": 0.15,
675
+ "synth_energy_drop": 0.3,
676
+ "synth_instrument_drop": 0.3,
677
+ "validate_every": 5000,
678
+ "retrieve_every": 10000,
679
+ "retrieve_pools": 0,
680
+ "seed": 1,
681
+ "contrast_weight": 0.0,
682
+ "contrast_k": 31,
683
+ "contrast_tau": 0.05,
684
+ "noise_rate": 0.5,
685
+ "noise_max": 12,
686
+ "noise_scatter": 0.35,
687
+ "noise_shape": 3.0,
688
+ "noise_real_rate": 0.0,
689
+ "synth_peak_drop": -1.0,
690
+ "synth_jitter": -1.0,
691
+ "noise_near": 0.67,
692
+ "noise_deltas": "inv/noise_deltas.json",
693
+ "init": null,
694
+ "max_steps_this_run": 0
695
+ }
696
+ },
697
+ {
698
+ "source": "runs/I21-ff512x6-real08-corpus2+train3-noise05-real05/step-0050000-ema.pt",
699
+ "sha256": "152cbfe71ab16096186443adff88b0ad14cce4dae65c655858476185167d5ebc",
700
+ "step": 50000,
701
+ "class": "FormulaFormer",
702
+ "weights": "ema",
703
+ "initialization": "random",
704
+ "seed": 1,
705
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
706
+ "self_simulated_store": {
707
+ "path": "inv/selfsim/corpus2+train3.npz",
708
+ "rows": 2031301,
709
+ "manifest": {
710
+ "schema": "allcond-inverse-selfsim/v1",
711
+ "merged": [
712
+ "inv/selfsim/corpus2.npz",
713
+ "inv/selfsim/train3.npz"
714
+ ],
715
+ "parts": [
716
+ {
717
+ "schema": "allcond-inverse-selfsim/v1",
718
+ "merged": [
719
+ "inv/selfsim/sub75.npz",
720
+ "inv/selfsim/sub25.npz"
721
+ ],
722
+ "parts": [
723
+ {
724
+ "name": "sub75",
725
+ "source": "synth",
726
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
727
+ "bundle": "builds/5Ms-I1+I3x2-L",
728
+ "members": "1,3",
729
+ "structures": 435076,
730
+ "tasks": 651127,
731
+ "failed": 0,
732
+ "rows": 651127,
733
+ "seconds": 635.7592854518443,
734
+ "schema": "allcond-inverse-selfsim/v1",
735
+ "floor": 0.02,
736
+ "spectra": 651083,
737
+ "without_peak": 44,
738
+ "mean_annotated_peaks": 15.61580167198345,
739
+ "mean_candidates": 20.62178653319552,
740
+ "ppm": 20.0,
741
+ "floor_da": 0.004,
742
+ "max_peaks": 24,
743
+ "annotate_seconds": 25.62758301384747
744
+ },
745
+ {
746
+ "name": "sub25",
747
+ "source": "synth",
748
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
749
+ "bundle": "builds/5Ms-I1+I3x2-L",
750
+ "members": "1,3",
751
+ "structures": 435076,
752
+ "tasks": 217012,
753
+ "failed": 0,
754
+ "rows": 217012,
755
+ "seconds": 227.74617605702952,
756
+ "schema": "allcond-inverse-selfsim/v1",
757
+ "floor": 0.02,
758
+ "spectra": 217000,
759
+ "without_peak": 12,
760
+ "mean_annotated_peaks": 15.62026728110599,
761
+ "mean_candidates": 20.61567563084069,
762
+ "ppm": 20.0,
763
+ "floor_da": 0.004,
764
+ "max_peaks": 24,
765
+ "annotate_seconds": 10.8742212112993
766
+ }
767
+ ],
768
+ "spectra": 868083,
769
+ "structures": 434944
770
+ },
771
+ {
772
+ "name": "train3",
773
+ "source": "synth",
774
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
775
+ "bundle": "builds/5Ms-I1+I3x2-L",
776
+ "members": "1,3",
777
+ "structures": 435076,
778
+ "tasks": 1163260,
779
+ "failed": 0,
780
+ "rows": 1163260,
781
+ "seconds": 1262.926542488858,
782
+ "schema": "allcond-inverse-selfsim/v1",
783
+ "floor": 0.02,
784
+ "spectra": 1163218,
785
+ "without_peak": 42,
786
+ "mean_annotated_peaks": 17.868643710809152,
787
+ "mean_candidates": 25.376552103570997,
788
+ "ppm": 20.0,
789
+ "floor_da": 0.004,
790
+ "max_peaks": 24,
791
+ "annotate_seconds": 43.40271363919601
792
+ }
793
+ ],
794
+ "spectra": 2031301,
795
+ "structures": 434950
796
+ }
797
+ },
798
+ "stores": {
799
+ "gen-x1.manifest.json": null,
800
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
801
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
802
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
803
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
804
+ },
805
+ "recipe": {
806
+ "store": "/workspace/research/inv",
807
+ "out": "runs/I21-ff512x6-real08-corpus2+train3-noise05-real05",
808
+ "pools": "pools-validation-input.npz",
809
+ "input": "/workspace/research/validation-input",
810
+ "steps": 50000,
811
+ "batch": 512,
812
+ "real_frac": 0.08,
813
+ "real_balance": 0.5,
814
+ "no_synth": false,
815
+ "target": "",
816
+ "heavy_boost": 1.0,
817
+ "synth_structures": 1.0,
818
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
819
+ "extra_frac": 0.5,
820
+ "heavy_lo": 450.0,
821
+ "heavy_hi": 850.0,
822
+ "lr": 0.0006,
823
+ "weight_decay": 0.0001,
824
+ "warmup": 2000,
825
+ "ema": 0.999,
826
+ "dim": 512,
827
+ "layers": 6,
828
+ "heads": 8,
829
+ "dropout": 0.1,
830
+ "peak_drop": 0.3,
831
+ "truncate": 0.3,
832
+ "min_keep": 4,
833
+ "jitter": 0.15,
834
+ "synth_energy_drop": 0.3,
835
+ "synth_instrument_drop": 0.3,
836
+ "validate_every": 5000,
837
+ "retrieve_every": 10000,
838
+ "retrieve_pools": 0,
839
+ "seed": 1,
840
+ "contrast_weight": 0.0,
841
+ "contrast_k": 31,
842
+ "contrast_tau": 0.05,
843
+ "noise_rate": 0.5,
844
+ "noise_max": 12,
845
+ "noise_scatter": 0.35,
846
+ "noise_shape": 3.0,
847
+ "noise_real_rate": 0.5,
848
+ "synth_peak_drop": -1.0,
849
+ "synth_jitter": -1.0,
850
+ "noise_near": 0.0,
851
+ "noise_deltas": "inv/noise_deltas.json",
852
+ "init": null,
853
+ "max_steps_this_run": 0
854
+ }
855
+ }
856
+ ],
857
+ "library_asset": {
858
+ "schema": "train-library/v1",
859
+ "fold": "train",
860
+ "spectra": 194119,
861
+ "molecules": 25046,
862
+ "bin": 0.01,
863
+ "prior_sample": 1024,
864
+ "prior_seed": 0,
865
+ "fragment_entries": 6765621,
866
+ "loss_entries": 6763777,
867
+ "source": "/workspace/data/train.jsonl",
868
+ "source_sha256": "de7f46e91f7b67eb5ca67468381abbdb20a8f739a0ae4b27070276798ecad0fc",
869
+ "structure_store": "/run-state/workspace/research/inv/real-train.npz",
870
+ "structure_store_sha256": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
871
+ "sha256": "22d1ddb216eef78d5040209ff6526371e5fcd8aac88e68a36823b6d5e73b3712",
872
+ "elapsed_seconds": 13.499361125999712,
873
+ "training_data": "MassSpecGym v1.5 train fold"
874
+ },
875
+ "ion_asset": {
876
+ "source": "inv/ion_error_table.json",
877
+ "sha256": "39aca3a714fd25a20cbaba4d46942764b52d730828a4b8205bf46ac5fc2a6225",
878
+ "fit": "MassSpecGym v1.5 train fold (194,119 spectra with [M+H]+ / [M+Na]+), analysis/ion_fit.py",
879
+ "training_data": "MassSpecGym v1.5 train fold",
880
+ "window_da": 0.02,
881
+ "classes": {
882
+ "domain|Orbitrap": {
883
+ "median_ppm": 0.358,
884
+ "sigma_ppm": 1.51,
885
+ "background": 0.0204
886
+ },
887
+ "domain|QTOF": {
888
+ "median_ppm": 0.027,
889
+ "sigma_ppm": 4.348,
890
+ "background": 0.0869
891
+ },
892
+ "domain|other": {
893
+ "median_ppm": 0.027,
894
+ "sigma_ppm": 4.348,
895
+ "background": 0.0869
896
+ },
897
+ "H-no-energy|Orbitrap": {
898
+ "median_ppm": -0.943,
899
+ "sigma_ppm": 2.937,
900
+ "background": 0.1196
901
+ },
902
+ "H-no-energy|QTOF": {
903
+ "median_ppm": 0.665,
904
+ "sigma_ppm": 4.58,
905
+ "background": 0.08
906
+ },
907
+ "H-no-energy|other": {
908
+ "median_ppm": 0.851,
909
+ "sigma_ppm": 4.269,
910
+ "background": 0.1021
911
+ },
912
+ "Na|Orbitrap": {
913
+ "median_ppm": -0.872,
914
+ "sigma_ppm": 2.935,
915
+ "background": 0.1834
916
+ },
917
+ "Na|QTOF": {
918
+ "median_ppm": 1.119,
919
+ "sigma_ppm": 4.103,
920
+ "background": 0.0798
921
+ },
922
+ "Na|other": {
923
+ "median_ppm": 3.003,
924
+ "sigma_ppm": 2.969,
925
+ "background": 0.0319
926
+ }
927
+ },
928
+ "fallback": {
929
+ "domain|other": "domain|QTOF"
930
+ },
931
+ "train_fold_statistics": {
932
+ "domain|Orbitrap": {
933
+ "spectra": 64738,
934
+ "present": 0.5068,
935
+ "median_ppm": 0.176,
936
+ "sigma_ppm": 1.093,
937
+ "beyond_3_sigma": 0.0763,
938
+ "ml_median_ppm": 0.358,
939
+ "ml_sigma_ppm": 1.51,
940
+ "ml_background": 0.0204,
941
+ "ml_mean_log_density": 6.2307
942
+ },
943
+ "domain|QTOF": {
944
+ "spectra": 34603,
945
+ "present": 0.5111,
946
+ "median_ppm": 0.045,
947
+ "sigma_ppm": 3.847,
948
+ "beyond_3_sigma": 0.0914,
949
+ "ml_median_ppm": 0.027,
950
+ "ml_sigma_ppm": 4.348,
951
+ "ml_background": 0.0869,
952
+ "ml_mean_log_density": 4.9233
953
+ },
954
+ "domain|other": {
955
+ "spectra": 36,
956
+ "present": 0.8611,
957
+ "median_ppm": 2.185,
958
+ "sigma_ppm": 4.203,
959
+ "beyond_3_sigma": 0.0,
960
+ "ml_median_ppm": 2.789,
961
+ "ml_sigma_ppm": 3.688,
962
+ "ml_background": 0.0,
963
+ "ml_mean_log_density": 5.6271
964
+ },
965
+ "H-no-energy|Orbitrap": {
966
+ "spectra": 51704,
967
+ "present": 0.5479,
968
+ "median_ppm": -0.989,
969
+ "sigma_ppm": 2.647,
970
+ "beyond_3_sigma": 0.1156,
971
+ "ml_median_ppm": -0.943,
972
+ "ml_sigma_ppm": 2.937,
973
+ "ml_background": 0.1196,
974
+ "ml_mean_log_density": 5.0128
975
+ },
976
+ "H-no-energy|QTOF": {
977
+ "spectra": 11931,
978
+ "present": 0.626,
979
+ "median_ppm": 0.801,
980
+ "sigma_ppm": 4.239,
981
+ "beyond_3_sigma": 0.0754,
982
+ "ml_median_ppm": 0.665,
983
+ "ml_sigma_ppm": 4.58,
984
+ "ml_background": 0.08,
985
+ "ml_mean_log_density": 4.7319
986
+ },
987
+ "H-no-energy|other": {
988
+ "spectra": 3467,
989
+ "present": 0.67,
990
+ "median_ppm": 0.61,
991
+ "sigma_ppm": 4.033,
992
+ "beyond_3_sigma": 0.0895,
993
+ "ml_median_ppm": 0.851,
994
+ "ml_sigma_ppm": 4.269,
995
+ "ml_background": 0.1021,
996
+ "ml_mean_log_density": 4.7426
997
+ },
998
+ "Na|Orbitrap": {
999
+ "spectra": 24454,
1000
+ "present": 0.4954,
1001
+ "median_ppm": -0.875,
1002
+ "sigma_ppm": 2.797,
1003
+ "beyond_3_sigma": 0.1574,
1004
+ "ml_median_ppm": -0.872,
1005
+ "ml_sigma_ppm": 2.935,
1006
+ "ml_background": 0.1834,
1007
+ "ml_mean_log_density": 4.589
1008
+ },
1009
+ "Na|QTOF": {
1010
+ "spectra": 2361,
1011
+ "present": 0.8602,
1012
+ "median_ppm": 1.126,
1013
+ "sigma_ppm": 4.211,
1014
+ "beyond_3_sigma": 0.0551,
1015
+ "ml_median_ppm": 1.119,
1016
+ "ml_sigma_ppm": 4.103,
1017
+ "ml_background": 0.0798,
1018
+ "ml_mean_log_density": 4.5016
1019
+ },
1020
+ "Na|other": {
1021
+ "spectra": 825,
1022
+ "present": 0.9309,
1023
+ "median_ppm": 3.305,
1024
+ "sigma_ppm": 3.261,
1025
+ "beyond_3_sigma": 0.026,
1026
+ "ml_median_ppm": 3.003,
1027
+ "ml_sigma_ppm": 2.969,
1028
+ "ml_background": 0.0319,
1029
+ "ml_mean_log_density": 4.9698
1030
+ }
1031
+ }
1032
+ },
1033
+ "method": "native-glacier-all-conditions-forward-simulation+formula-transformer-inverse-fusion+train-library-neighbour-fusion+measured-intact-ion-term+cross-pool-consensus",
1034
+ "inference_runtime": "/opt/glacier/bin/python",
1035
+ "members": [
1036
+ {
1037
+ "source": "runs/B-all-b32/checkpoints/step-0020000-ema.ckpt",
1038
+ "sha256": "80ec4715663e505007412e3d89dbe6020cd145713840c15affaf82604fc5765c",
1039
+ "global_step": 20000,
1040
+ "weights": "ema",
1041
+ "subset": "all",
1042
+ "seed": 1,
1043
+ "elapsed_seconds": 5136.2172065590275,
1044
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
1045
+ "initialization": "random",
1046
+ "training_data": "MassSpecGym v1.5 train fold",
1047
+ "synthetic": null
1048
+ },
1049
+ {
1050
+ "source": "runs/B-all-b32/checkpoints/step-0030000-ema.ckpt",
1051
+ "sha256": "0228e4fa0589333f02db8223e49d6ecd23cd100d1a6182ad268d9db4f8468798",
1052
+ "global_step": 30000,
1053
+ "weights": "ema",
1054
+ "subset": "all",
1055
+ "seed": 1,
1056
+ "elapsed_seconds": 8314.79353913304,
1057
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
1058
+ "initialization": "random",
1059
+ "training_data": "MassSpecGym v1.5 train fold",
1060
+ "synthetic": null
1061
+ },
1062
+ {
1063
+ "source": "runs/B2-all-b32-seed2/checkpoints/step-0020000-ema.ckpt",
1064
+ "sha256": "ac3c85cdabfb5f68ca3ca506732f1f8e79fa6e9f3d5bf5fa574b6e1f3249312c",
1065
+ "global_step": 20000,
1066
+ "weights": "ema",
1067
+ "subset": "all",
1068
+ "seed": 2,
1069
+ "elapsed_seconds": 6945.18486677701,
1070
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
1071
+ "initialization": "random",
1072
+ "training_data": "MassSpecGym v1.5 train fold",
1073
+ "synthetic": null
1074
+ },
1075
+ {
1076
+ "source": "runs/B2-all-b32-seed2/checkpoints/step-0030000-ema.ckpt",
1077
+ "sha256": "f11667b8ee56a15c82d26ab9052816ecc63f2081eaa6fd4ea82d38d1e2a36ce1",
1078
+ "global_step": 30000,
1079
+ "weights": "ema",
1080
+ "subset": "all",
1081
+ "seed": 2,
1082
+ "elapsed_seconds": 9298.063542745003,
1083
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
1084
+ "initialization": "random",
1085
+ "training_data": "MassSpecGym v1.5 train fold",
1086
+ "synthetic": null
1087
+ },
1088
+ {
1089
+ "source": "runs/S2-B20k-synth/checkpoints/step-0030000-ema.ckpt",
1090
+ "sha256": "44742fa31f337e59b543d903018780560569c2e8e50a008ba008859c8440e3a2",
1091
+ "global_step": 30000,
1092
+ "weights": "ema",
1093
+ "subset": "all",
1094
+ "seed": 1,
1095
+ "elapsed_seconds": 8906.84206694903,
1096
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
1097
+ "initialization": "own-checkpoint:6c5a43fada335e5f4e9f8265ec6cfd54a8f317eb8d2100d96ea3731809005be2",
1098
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus (auxiliary intensity term)",
1099
+ "synthetic": {
1100
+ "manifest_sha256": "431454e702f8da3194684fc9a565135d29a38a407aaa31e9399d871360c6e962",
1101
+ "per_item": 1,
1102
+ "weight": 1.0,
1103
+ "schedule": "e2e",
1104
+ "instrument": "flag"
1105
+ }
1106
+ }
1107
+ ]
1108
+ }
shared/package-files/0b/0b257defdfda2eed651ae468799ba3c865da5f9b1e998c87f12202905d332c67/consensus_subset.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Packaged consensus term on a SUBSET of the validation input, against a replay of the package on that subset.
2
+
3
+ consensus_subset.py --queries val-sub/queries.jsonl --rankings evals/replay/NAME-sub/rankings.jsonl
4
+ [--pred evals/fullfold/BASE/pred] [--build builds/NAME]
5
+
6
+ The term depends on the whole input set, so a replay on a subset cannot be compared with the full-fold run.
7
+ This check recomputes the term for the subset from the saved full-fold branch outputs (forward, inverse,
8
+ library, intact-ion of the base build; those do not depend on the other pools of the input) with the
9
+ packaged ``consensus.run`` and compares the label rank of every query with the replay's rankings.
10
+ Validation labels only evaluate.
11
+ """
12
+ import argparse
13
+ import json
14
+ import pickle
15
+ import sys
16
+ from pathlib import Path
17
+
18
+ import numpy as np
19
+
20
+ HERE = Path(__file__).resolve().parent
21
+ sys.path.insert(0, str(HERE.parent / 'analysis'))
22
+ sys.path.insert(0, str(HERE.parent / 'pkg'))
23
+ import consensus
24
+ import netprop
25
+ import scorelab
26
+
27
+ ROOT = Path('/workspace/research')
28
+
29
+
30
+ def main():
31
+ p = argparse.ArgumentParser(description=__doc__)
32
+ p.add_argument('--queries', required=True)
33
+ p.add_argument('--rankings', required=True)
34
+ p.add_argument('--pred', default='evals/fullfold/5Ms-I9b+I13-L-ion/pred')
35
+ p.add_argument('--build', default='builds/5Ms-I9b+I13-L-ion-net')
36
+ args = p.parse_args()
37
+ import torch
38
+ config = json.load(open(ROOT / args.build / 'model_config.json'))
39
+ wanted = [json.loads(line)['query_id'] for line in open(ROOT / args.queries)]
40
+ queries, pools = netprop.load()
41
+ by_id = {q['query_id']: q for q in queries}
42
+ queries = [by_id[q] for q in wanted if q in by_id] # hand-built queries of the edge input have no label
43
+ ev = scorelab.Evaluator(queries, pools)
44
+ pred = {name: pickle.load(open(ROOT / args.pred / f'{name}.pkl', 'rb')) for name in ('distances', 'inverse', 'library', 'ion')}
45
+ base = netprop.released(queries, pools, pred, config)
46
+ store = np.load(ROOT / 'inv/pools-validation-input.npz')
47
+ slot = {pid: n for n, pid in enumerate(store['pool_id'])}
48
+ data = {'pool_id': store['pool_id'], 'pool_cand_start': store['pool_cand_start'], 'cand_fps': store['cand_fps'],
49
+ 'query_pool': np.array([slot.get(q['candidate_pool_id'], -1) for q in queries], dtype=np.int32)}
50
+ terms = consensus.run(queries, pools, data, base, config['consensus'], torch.device('cuda' if torch.cuda.is_available() else 'cpu'),
51
+ lambda event, **values: print(event, json.dumps(values)))
52
+ weights = config['consensus']
53
+ fused = [b if (b is None or t is None) else b + (weights['weight_domain'] if dom else weights['weight_other']) * t
54
+ for b, t, dom in zip(base, terms, ev.domain)]
55
+ rank0, rank = ev.ranks(base, 1.0), ev.ranks(fused, 1.0)
56
+ replay = {r['query_id']: r['ranking'] for r in map(json.loads, open(ROOT / args.rankings))}
57
+ got = np.array([replay[q['query_id']].index(ev.ids[q['candidate_pool_id']][ev.target[i]]) for i, q in enumerate(queries)])
58
+ first = np.array([rows[0] for rows in ev.by_pool.values()])
59
+ d = (rank[first] == 0).astype(float) - (rank0[first] == 0).astype(float)
60
+ print(f'subset: {len(queries)} labelled queries, {len(first)} pools')
61
+ print(f'base (no term) Hit@1 {np.mean(rank0 == 0):.4f} Hit@5 {np.mean(rank0 < 5):.4f} Hit@20 {np.mean(rank0 < 20):.4f} macro {np.mean(rank0[first] == 0):.4f}')
62
+ print(f'lab, with term Hit@1 {np.mean(rank == 0):.4f} Hit@5 {np.mean(rank < 5):.4f} Hit@20 {np.mean(rank < 20):.4f} macro {np.mean(rank[first] == 0):.4f}'
63
+ f' pools +{int((d > 0).sum())} -{int((d < 0).sum())}')
64
+ print(f'replay Hit@1 {np.mean(got == 0):.4f} Hit@5 {np.mean(got < 5):.4f} Hit@20 {np.mean(got < 20):.4f} macro {np.mean(got[first] == 0):.4f}')
65
+ print(f'label rank: replay == lab for {int(np.sum(got == rank))} / {len(rank)} queries; top-1 status equal for {int(np.sum((got == 0) == (rank == 0)))}; '
66
+ f'largest rank difference {int(np.max(np.abs(got - rank)))}')
67
+
68
+
69
+ if __name__ == '__main__':
70
+ main()
shared/package-files/0b/0b2938300f3af759ecaf7b1bc27fa153b05d5eff63eaca867867c5359e8cdeb5/train.py ADDED
@@ -0,0 +1,238 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Train and atomically publish the runnable Fourier-DeepSets starter."""
3
+
4
+ import argparse
5
+ import json
6
+ import os
7
+ from pathlib import Path
8
+ import random
9
+ import shutil
10
+ import time
11
+
12
+ import numpy as np
13
+ import torch
14
+ import torch.nn.functional as F
15
+ from torch.utils.data import DataLoader
16
+
17
+ from common import LabeledSpectra, compact_json, sha256
18
+ from model import DeepSetsFingerprint
19
+ from publication import publish_directory
20
+
21
+
22
+ def cosine_loss(prediction, target):
23
+ return 1 - F.cosine_similarity(prediction, target, dim=-1).mean()
24
+
25
+
26
+ def evaluate(model, loader, device):
27
+ model.eval()
28
+ total, count = 0.0, 0
29
+ with torch.no_grad():
30
+ for spectra, targets in loader:
31
+ spectra = spectra.to(device, non_blocking=True)
32
+ targets = targets.to(device, non_blocking=True)
33
+ loss = cosine_loss(model(spectra), targets)
34
+ total += float(loss) * spectra.shape[0]
35
+ count += spectra.shape[0]
36
+ return total / count
37
+
38
+
39
+ def atomic_torch_save(value, path: Path):
40
+ temporary = path.with_name(path.name + ".tmp")
41
+ torch.save(value, temporary)
42
+ os.replace(temporary, path)
43
+
44
+
45
+ def main():
46
+ parser = argparse.ArgumentParser()
47
+ parser.add_argument("--input-root", type=Path, default=Path("/workspace/data"))
48
+ parser.add_argument("--output-dir", type=Path, default=Path.cwd())
49
+ parser.add_argument("--epochs", type=int, default=50)
50
+ parser.add_argument("--batch-size", type=int, default=256)
51
+ parser.add_argument("--workers", type=int, default=8)
52
+ parser.add_argument("--learning-rate", type=float, default=3e-4)
53
+ parser.add_argument("--hidden-size", type=int, default=128)
54
+ parser.add_argument("--fingerprint-size", type=int, default=4096)
55
+ parser.add_argument("--layers-per-mlp", type=int, default=4)
56
+ parser.add_argument("--dropout", type=float, default=0.1)
57
+ parser.add_argument("--max-train-records", type=int)
58
+ parser.add_argument("--max-validation-records", type=int)
59
+ parser.add_argument("--seed", type=int, default=0)
60
+ parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
61
+ parser.add_argument("--work-dir", type=Path, help="Optimizer checkpoints and logs, outside the submission (default: /workspace/research/<output name>-training)")
62
+ parser.add_argument("--resume", type=Path, help="Resume last.pt at its completed epoch")
63
+ parser.add_argument("--max-seconds", type=float, default=3600,
64
+ help="Cumulative fit budget, checked at epoch boundaries")
65
+ args = parser.parse_args()
66
+ if min(args.epochs, args.batch_size, args.max_seconds) <= 0 or args.workers < 0:
67
+ raise ValueError("epochs must be positive")
68
+
69
+ output = args.output_dir.resolve()
70
+ work = (args.work_dir or Path("/workspace/research") / (output.name + "-training")).resolve()
71
+ if work == output or work.is_relative_to(output):
72
+ raise ValueError("keep optimizer checkpoints outside the submission")
73
+ work.mkdir(parents=True, exist_ok=True)
74
+ started = time.monotonic()
75
+ random.seed(args.seed)
76
+ np.random.seed(args.seed)
77
+ torch.manual_seed(args.seed)
78
+ if torch.cuda.is_available():
79
+ torch.cuda.manual_seed_all(args.seed)
80
+
81
+ train_data = LabeledSpectra(
82
+ args.input_root / "train.jsonl", args.max_train_records, args.seed,
83
+ fingerprint_size=args.fingerprint_size,
84
+ )
85
+ validation_data = LabeledSpectra(
86
+ args.input_root / "validation.jsonl", args.max_validation_records, args.seed,
87
+ fingerprint_size=args.fingerprint_size,
88
+ )
89
+ loader_generator = torch.Generator().manual_seed(args.seed)
90
+ train_loader = DataLoader(
91
+ train_data, batch_size=args.batch_size, shuffle=True, generator=loader_generator,
92
+ num_workers=args.workers, pin_memory=args.device.startswith("cuda"),
93
+ persistent_workers=args.workers > 0,
94
+ )
95
+ validation_loader = DataLoader(
96
+ validation_data, batch_size=args.batch_size, shuffle=False,
97
+ num_workers=args.workers, pin_memory=args.device.startswith("cuda"),
98
+ persistent_workers=args.workers > 0,
99
+ )
100
+ config = {
101
+ "hidden_size": args.hidden_size,
102
+ "fingerprint_size": args.fingerprint_size,
103
+ "layers_per_mlp": args.layers_per_mlp,
104
+ "dropout": args.dropout,
105
+ "max_peaks": 60,
106
+ "mz_min": 10.0,
107
+ "mz_max": 1000.0,
108
+ "fingerprint_radius": 2,
109
+ }
110
+ device = torch.device(args.device)
111
+ model = DeepSetsFingerprint(
112
+ hidden_size=args.hidden_size,
113
+ fingerprint_size=args.fingerprint_size,
114
+ layers_per_mlp=args.layers_per_mlp,
115
+ dropout=args.dropout,
116
+ ).to(device)
117
+ optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
118
+ scaler = torch.cuda.amp.GradScaler(enabled=device.type == "cuda")
119
+ identity = {
120
+ "schema": "massspecgym-retrieval-fit/v1", "initialization": "random", "config": config,
121
+ "seed": args.seed, "batch_size": args.batch_size, "learning_rate": args.learning_rate,
122
+ "train_sha256": sha256(args.input_root / "train.jsonl"),
123
+ "validation_sha256": sha256(args.input_root / "validation.jsonl"),
124
+ "train_records": len(train_data), "validation_records": len(validation_data),
125
+ }
126
+ if (work / "recipe.json").exists() and (
127
+ not args.resume or json.loads((work / "recipe.json").read_text()) != identity
128
+ ):
129
+ raise ValueError("use a fresh work directory or --resume with the same recipe")
130
+ (work / "recipe.json").write_text(compact_json(identity) + "\n")
131
+ best_loss, best_epoch, best_state = float("inf"), None, None
132
+ first_epoch, prior = 1, 0.0
133
+ if args.resume:
134
+ saved = torch.load(args.resume, map_location="cpu", weights_only=False)
135
+ if saved["identity"] != identity:
136
+ raise ValueError("resume data/model/seed/batch configuration changed")
137
+ model.load_state_dict(saved["model"])
138
+ optimizer.load_state_dict(saved["optimizer"])
139
+ scaler.load_state_dict(saved["scaler"])
140
+ best_state, best_epoch, best_loss = saved["best_state"], saved["best_epoch"], saved["best_loss"]
141
+ first_epoch, prior = saved["epoch"] + 1, saved["elapsed_seconds"]
142
+ random.setstate(saved["rng"]["python"])
143
+ np.random.set_state(saved["rng"]["numpy"])
144
+ torch.set_rng_state(saved["rng"]["torch"])
145
+ loader_generator.set_state(saved["rng"]["loader"])
146
+ if torch.cuda.is_available() and saved["rng"]["cuda"]:
147
+ torch.cuda.set_rng_state_all(saved["rng"]["cuda"])
148
+ if prior >= args.max_seconds:
149
+ raise ValueError("resume has already exhausted the cumulative training budget")
150
+ source = Path(__file__).resolve().parent
151
+
152
+ def publish_best():
153
+ summary = dict(identity, selected_epoch=best_epoch, best_validation_cosine_loss=best_loss,
154
+ elapsed_seconds=prior + time.monotonic() - started)
155
+ def build(stage):
156
+ for name in ("infer.sh", "predict.py", "model.py", "common.py", "train.py", "publication.py"):
157
+ shutil.copyfile(source / name, stage / name)
158
+ (stage / "infer.sh").chmod(0o755)
159
+ checkpoint = stage / "model.pt"
160
+ torch.save({"state_dict": best_state, "model_config": config}, checkpoint)
161
+ (stage / "model_config.json").write_text(compact_json(config) + "\n")
162
+ (stage / "training_summary.json").write_text(compact_json(summary) + "\n")
163
+ (stage / "README.md").write_text(
164
+ "# Trained Fourier-DeepSets retrieval model\n\n"
165
+ "Initialized from scratch on the provided train fold; selected on validation loss. "
166
+ "See training_summary.json for the data and recipe identity.\n")
167
+ manifest = {
168
+ "schema": "massspecgym-retrieval-model/v1",
169
+ "checkpoint": {"path": "model.pt", "sha256": sha256(checkpoint), "format": "pytorch"},
170
+ "model_config": "model_config.json",
171
+ "training_provenance": {"data": "MassSpecGym v1.5 train fold", "pretrained_assets": []},
172
+ }
173
+ (stage / "model_manifest.json").write_text(compact_json(manifest) + "\n")
174
+ publish_directory(output, build)
175
+
176
+ if best_state is not None:
177
+ publish_best()
178
+ for epoch in range(first_epoch, args.epochs + 1):
179
+ model.train()
180
+ running, seen = 0.0, 0
181
+ for spectra, targets in train_loader:
182
+ spectra = spectra.to(device, non_blocking=True)
183
+ targets = targets.to(device, non_blocking=True)
184
+ optimizer.zero_grad(set_to_none=True)
185
+ with torch.autocast(
186
+ device_type=device.type,
187
+ dtype=torch.float16 if device.type == "cuda" else torch.bfloat16,
188
+ enabled=device.type == "cuda",
189
+ ):
190
+ loss = cosine_loss(model(spectra), targets)
191
+ if not torch.isfinite(loss):
192
+ raise FloatingPointError(
193
+ f"non-finite training loss at epoch {epoch}; "
194
+ "check the feature precision and input values"
195
+ )
196
+ scaler.scale(loss).backward()
197
+ scaler.step(optimizer)
198
+ scaler.update()
199
+ running += float(loss) * spectra.shape[0]
200
+ seen += spectra.shape[0]
201
+ validation_loss = evaluate(model, validation_loader, device)
202
+ if not np.isfinite(validation_loss):
203
+ raise FloatingPointError(
204
+ f"non-finite validation loss at epoch {epoch}"
205
+ )
206
+ print(compact_json({
207
+ "epoch": epoch,
208
+ "train_loss": running / seen,
209
+ "validation_loss": validation_loss,
210
+ }), flush=True)
211
+ if validation_loss < best_loss:
212
+ best_loss = validation_loss
213
+ best_epoch = epoch
214
+ best_state = {
215
+ name: value.detach().cpu().clone()
216
+ for name, value in model.state_dict().items()
217
+ }
218
+
219
+ publish_best()
220
+ elapsed = prior + time.monotonic() - started
221
+ atomic_torch_save({
222
+ "identity": identity, "epoch": epoch, "model": model.state_dict(),
223
+ "optimizer": optimizer.state_dict(), "scaler": scaler.state_dict(),
224
+ "best_state": best_state, "best_epoch": best_epoch, "best_loss": best_loss,
225
+ "elapsed_seconds": elapsed,
226
+ "rng": {"python": random.getstate(), "numpy": np.random.get_state(),
227
+ "torch": torch.get_rng_state(), "loader": loader_generator.get_state(),
228
+ "cuda": torch.cuda.get_rng_state_all() if torch.cuda.is_available() else []},
229
+ }, work / "last.pt")
230
+ with (work / "metrics.jsonl").open("a") as stream:
231
+ stream.write(compact_json({"epoch": epoch, "train_loss": running / seen,
232
+ "validation_loss": validation_loss, "elapsed_seconds": elapsed}) + "\n")
233
+ if elapsed >= args.max_seconds:
234
+ break
235
+
236
+
237
+ if __name__ == "__main__":
238
+ main()
shared/package-files/0b/0b543c06ef5862083df44fe801d381850637714742e2f1a3807c4424e0201cdf/e74-na-as-h.sh ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ # E74: [M+Na]+ spectra read as protonated by the inverse networks (view alone and mixed with the release view)
3
+ cd /workspace/research
4
+ export CUDA_MPS_PIPE_DIRECTORY=/tmp/mps GLACIER_SOURCE=/workspace/research/ms-pred/src
5
+ for w in 0.0 1.0 0.25 0.5; do
6
+ taskset -c 11-24 /opt/glacier/bin/python allcond/inv/eval_score.py evals/inv-score/e74-na-as-h-w$w.pkl runs/I16-ff512x6-real08-corpus2+train3-noise05/step-0050000-ema.pt runs/I18-ff512x6-real08-corpus2+train3-noise05-real025/step-0050000-ema.pt runs/I19-ff512x6-real08-corpus2+train3-noise05-jit08/step-0050000-ema.pt runs/I20-ff512x6-real08-corpus2+train3-noise05-near067/step-0050000-ema.pt runs/I21-ff512x6-real08-corpus2+train3-noise05-real05/step-0050000-ema.pt --pools inv/recur/pkg-hyb-p48.npz --mix inv/recur/na-as-h-hyb-p48.npz --mix-weight $w --max-peaks 48 > logs/e74-na-as-h-w$w.log 2>&1
7
+ done
shared/package-files/0b/0b721b94afb29ec79be5ec0608f2cc2328babf9a1e904f7ce7ca02e50cf8b68e/fuselab.py ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Fast pool-level fusion laboratory on cached validation distances (development only).
2
+
3
+ Loads once: simulator distances (mean over scorelab member caches for one variant), any number
4
+ of inverse-branch distance pickles, validation targets (evaluation only) and the candidate-id
5
+ tie order of the official ranking rule. Scores are combined per pool as matrices
6
+ [spectra, candidates]; a pool-level rule returns one fused vector per pool (rank of the target
7
+ is then shared by all its queries) or one vector per spectrum.
8
+ """
9
+ import hashlib
10
+ import json
11
+ import pickle
12
+ from pathlib import Path
13
+
14
+ import numpy as np
15
+
16
+
17
+ class Lab:
18
+ def __init__(self, root, variant, members, cache='/tmp/fuselab-cache.pkl'):
19
+ root = Path(root)
20
+ key = hashlib.sha256((str(root) + variant + '|'.join(members)).encode()).hexdigest()[:12]
21
+ cache = Path(str(cache).replace('.pkl', f'-{key}.pkl'))
22
+ if cache.exists():
23
+ self.__dict__.update(pickle.load(cache.open('rb')))
24
+ return
25
+ queries = [json.loads(line) for line in (root / 'queries.jsonl').open()]
26
+ labels = {}
27
+ for line in open('/workspace/data/validation.jsonl'):
28
+ row = json.loads(line)
29
+ labels[row['query_id']] = hashlib.sha256(row['target_smiles'].encode()).hexdigest()
30
+ tie, index = {}, {}
31
+ for line in (root / 'candidate_pools.jsonl').open():
32
+ pool = json.loads(line)
33
+ ids = np.array([c['candidate_id'] for c in pool['candidates']])
34
+ tie[pool['candidate_pool_id']] = np.argsort(np.argsort(ids))
35
+ index[pool['candidate_pool_id']] = {c: n for n, c in enumerate(ids)}
36
+ self.target = np.array([index[q['candidate_pool_id']][labels[q['query_id']]] for q in queries])
37
+ self.pool_ids = list(dict.fromkeys(q['candidate_pool_id'] for q in queries))
38
+ slot = {pid: n for n, pid in enumerate(self.pool_ids)}
39
+ self.query_pool = np.array([slot[q['candidate_pool_id']] for q in queries])
40
+ self.tie = [tie[pid] for pid in self.pool_ids]
41
+ self.members = [np.flatnonzero(self.query_pool == s) for s in range(len(self.pool_ids))]
42
+ self.adduct = np.array([q['adduct'] for q in queries])
43
+ self.has_ce = np.array([q['collision_energy'] is not None for q in queries])
44
+ self.instrument = np.array([q['instrument_type'] for q in queries])
45
+ self.mz = np.array([q['precursor_mz'] for q in queries])
46
+ self.peaks = np.array([len(q['mzs']) for q in queries])
47
+ self.domain = (self.adduct == '[M+H]+') & self.has_ce
48
+ dist = []
49
+ for spec in members:
50
+ results = pickle.load(open(spec.split('=', 1)[1], 'rb'))
51
+ first = results[next(iter(results))]
52
+ dist.append([d if d is not None else f for d, f in zip(results[variant], first)])
53
+ self.sim = [None if any(m[i] is None for m in dist) else np.mean([m[i] for m in dist], axis=0)
54
+ for i in range(len(queries))]
55
+ pickle.dump(self.__dict__, cache.open('wb'))
56
+
57
+ def setup(self):
58
+ rng = np.random.default_rng(0)
59
+ self.boot = rng.integers(0, len(self.members), size=(1000, len(self.members)))
60
+ self.sizes = np.array([len(m) for m in self.members], dtype=float)
61
+ self.target_pool = np.array([self.target[m[0]] for m in self.members])
62
+ return self
63
+
64
+ def matrix(self, per_query, s, fill=1.0):
65
+ """[spectra, candidates] matrix of one pool; missing rows are filled with ``fill``."""
66
+ rows = [per_query[i] for i in self.members[s]]
67
+ width = len(self.tie[s])
68
+ return np.stack([np.full(width, fill) if r is None else np.asarray(r, dtype=np.float64) for r in rows])
69
+
70
+ def rank(self, s, vector):
71
+ order = np.lexsort((self.tie[s], vector))
72
+ return int(np.where(order == self.target_pool[s])[0][0])
73
+
74
+ def pool_rule(self, rule):
75
+ """rule(s) -> fused vector [candidates]; returns the target rank per query."""
76
+ out = np.zeros(len(self.query_pool), dtype=np.int64)
77
+ for s, members in enumerate(self.members):
78
+ out[members] = self.rank(s, rule(s))
79
+ return out
80
+
81
+ def query_rule(self, rule):
82
+ """rule(s) -> matrix [spectra, candidates]; returns the target rank per query."""
83
+ out = np.zeros(len(self.query_pool), dtype=np.int64)
84
+ for s, members in enumerate(self.members):
85
+ m = rule(s)
86
+ for n, i in enumerate(members):
87
+ out[i] = self.rank(s, m[n])
88
+ return out
89
+
90
+ def paired(self, hit, base):
91
+ a = np.bincount(self.query_pool, weights=hit.astype(float), minlength=len(self.members))
92
+ b = np.bincount(self.query_pool, weights=base.astype(float), minlength=len(self.members))
93
+ n = self.sizes[self.boot].sum(1)
94
+ diff = (a[self.boot].sum(1) - b[self.boot].sum(1)) / n
95
+ return float(np.percentile(diff, 2.5)), float(np.percentile(diff, 97.5))
96
+
97
+ def line(self, name, rank, base=None):
98
+ hit = rank == 0
99
+ macro = np.mean([hit[m].mean() for m in self.members])
100
+ text = (f'{name:42s} all={hit.mean():.4f} dom={hit[self.domain].mean():.4f} non={hit[~self.domain].mean():.4f} '
101
+ f'h5={np.mean(rank < 5):.4f} h20={np.mean(rank < 20):.4f} macro={macro:.4f}')
102
+ if base is not None:
103
+ lo, hi = self.paired(hit, base == 0)
104
+ text += f' | {hit.mean() - (base == 0).mean():+.4f} [{lo:+.4f},{hi:+.4f}]'
105
+ return text
shared/package-files/0b/0be5bef1b6f13ac83bf2b7f4e0e46ef560c4abe029983b41a2a3f426fbc257d8/cycle.py ADDED
@@ -0,0 +1,334 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Cycle term: the inverse networks read the simulated spectrum of every candidate (lab, E66)
2
+
3
+ cycle.py prepare --out inv/cycle/val.npz [--members evals/B30k-full/pred,evals/B2-30k-full/pred] [--floor 0.02] [--workers 5]
4
+ cycle.py score --store inv/cycle/val.npz [--build builds/5Ms-I9b+I13-L-ion-net] [--pred evals/fullfold/5Ms-I9b+I13-L-ion/pred]
5
+ [--lams 0.1,0.25,0.5,1] [--save evals/cycle/terms.pkl]
6
+
7
+ Hypothesis: the released inverse branch compares the fingerprint the networks predict from the
8
+ measured spectra with the candidate's true bits. Many bits cannot be read from a spectrum at all;
9
+ for those the comparison rewards the candidate that agrees with the formula-only expectation. If
10
+ the same networks also read the candidate's *simulated* spectrum (the networks were trained on
11
+ spectra rendered by the same simulators, E60), the two predictions can be compared directly: bits
12
+ that no spectrum reveals drop out, and the comparison is a learned distance between the measured
13
+ and the simulated spectrum.
14
+
15
+ ``prepare`` annotates the saved full-fold simulations (mean of the listed members, as the training
16
+ rows of E60) with the subformulae of each candidate's own formula. ``score`` runs the released
17
+ inverse pair on them and scores several comparisons alone and inside the release. Labels evaluate
18
+ only; nothing is fitted.
19
+ """
20
+ import argparse
21
+ import json
22
+ import multiprocessing
23
+ import os
24
+ import pickle
25
+ import sys
26
+ import time
27
+ from pathlib import Path
28
+
29
+ import numpy as np
30
+
31
+ HERE = Path(__file__).resolve().parent
32
+ sys.path.insert(0, str(HERE))
33
+ sys.path.insert(0, str(HERE.parent / 'pkg'))
34
+ sys.path.insert(0, str(HERE.parent / 'inv'))
35
+ ROOT = Path('/workspace/research')
36
+ STATE = {}
37
+
38
+
39
+ def annotate_chunk(job):
40
+ import inverse
41
+ import invrun
42
+ from selfsim import merged_spectrum
43
+ low, high = job
44
+ if 'masses' not in STATE:
45
+ STATE['masses'] = [np.load(ROOT / m / 'masses.npy', mmap_mode='r') for m in STATE['members']]
46
+ STATE['intens'] = [np.load(ROOT / m / 'intens.npy', mmap_mode='r') for m in STATE['members']]
47
+ tasks, roots, adduct, floor = STATE['tasks'], STATE['roots'], STATE['adduct'], STATE['floor']
48
+ tables, out = {}, []
49
+ for n in range(low, high):
50
+ task = int(tasks[n])
51
+ formula = roots[n]
52
+ key = formula.tobytes()
53
+ table = tables.get(key)
54
+ if table is None:
55
+ if len(tables) > 64:
56
+ tables.clear()
57
+ table = tables[key] = inverse.SubformulaTable(formula)
58
+ masses = np.stack([np.asarray(m[0, task]) for m in STATE['masses']])
59
+ intens = np.stack([np.asarray(m[0, task]).astype(np.float32) for m in STATE['intens']])
60
+ mz, inten = merged_spectrum(masses, intens)
61
+ cut = inten >= floor
62
+ peaks, kept = inverse.annotate(table, mz[cut], inten[cut], inverse.ADDUCTS[int(adduct[n])], invrun.PPM, invrun.FLOOR)
63
+ out.append((peaks, kept))
64
+ return out
65
+
66
+
67
+ def prepare(args):
68
+ import infer
69
+ import inverse
70
+ import netprop
71
+ sys.path.insert(0, os.environ.get('GLACIER_SOURCE', '/workspace/research/ms-pred/src'))
72
+ from ms_pred import common
73
+ started = time.monotonic()
74
+ queries, pools = netprop.load(None)
75
+ inst = lambda name: common.instrument2onehot_pos.get(name or 'Unknown', common.instrument2onehot_pos['Unknown'])
76
+ plan = infer.Plan(queries, pools, args.workers, inst, set(common.ion2onehot_pos))
77
+ data = np.load(ROOT / 'inv/pools-validation-input.npz')
78
+ pool_id, fstart, cstart = data['pool_id'], data['pool_formula_start'], data['pool_cand_start']
79
+ formula, which = data['formula'], data['cand_formula']
80
+ q_adduct, q_instrument, q_energy = data['adduct'], data['instrument'], data['energy']
81
+ assert list(data['query_id']) == [q['query_id'] for q in queries]
82
+ slot = {str(pid): n for n, pid in enumerate(pool_id)}
83
+ members = args.members.split(',')
84
+ done = np.all([np.load(ROOT / m / 'done.npy') for m in members], axis=0)
85
+ total = len(plan.task_mol)
86
+ assert len(done) == total
87
+ root = np.zeros((total, formula.shape[1]), np.int16)
88
+ have = np.zeros(total, bool)
89
+ adduct = np.zeros(total, np.uint8)
90
+ instrument = np.zeros(total, np.uint8)
91
+ energy = np.full(total, np.nan, np.float32)
92
+ g_pool, g_tasks, g_queries = [], [], []
93
+ for (pid, _, _, _), group in plan.groups.items():
94
+ s = slot[pid]
95
+ tasks = group['tasks']
96
+ assert len(tasks) == cstart[s + 1] - cstart[s]
97
+ w = which[cstart[s]:cstart[s + 1]]
98
+ ok = (tasks >= 0) & (w >= 0)
99
+ q0 = group['queries'][0]
100
+ new = ok.copy()
101
+ new[ok] = ~have[tasks[ok]]
102
+ root[tasks[new]] = formula[fstart[s] + w[new]]
103
+ adduct[tasks[new]], instrument[tasks[new]], energy[tasks[new]] = q_adduct[q0], q_instrument[q0], q_energy[q0]
104
+ have[tasks[new]] = True
105
+ g_pool.append(s)
106
+ g_tasks.append(np.where(ok, tasks, -1))
107
+ g_queries.append(np.asarray(group['queries'], dtype=np.int64))
108
+ rows = np.flatnonzero(have & done)
109
+ print(f'plan: {len(plan.groups)} groups, {total} simulations, {len(rows)} with a formula table ({time.monotonic() - started:.0f} s)', flush=True)
110
+ STATE.update(members=members, tasks=rows, roots=root[rows], adduct=adduct[rows], floor=args.floor)
111
+ jobs = [(i, min(i + 4000, len(rows))) for i in range(0, len(rows), 4000)]
112
+ with multiprocessing.get_context('fork').Pool(args.workers) as pool:
113
+ got = [g for part in pool.imap(annotate_chunk, jobs, chunksize=1) for g in part]
114
+ counts = np.array([len(g[1]) for g in got], dtype=np.int64)
115
+ row_of_task = np.full(total, -1, np.int64)
116
+ row_of_task[rows] = np.arange(len(rows))
117
+ out = ROOT / args.out
118
+ out.parent.mkdir(parents=True, exist_ok=True)
119
+ np.savez(out, task=rows, row_of_task=row_of_task, root=root[rows], offset=np.concatenate([[0], np.cumsum(counts)]),
120
+ peaks=np.concatenate([g[0] for g in got]).astype(np.uint8), inten=np.concatenate([g[1] for g in got]).astype(np.float32),
121
+ adduct=adduct[rows], instrument=instrument[rows], energy=energy[rows],
122
+ group_pool=np.array(g_pool, dtype=np.int64), group_start=np.concatenate([[0], np.cumsum([len(t) for t in g_tasks])]),
123
+ group_tasks=np.concatenate(g_tasks), group_qstart=np.concatenate([[0], np.cumsum([len(q) for q in g_queries])]),
124
+ group_queries=np.concatenate(g_queries))
125
+ report = {'members': members, 'floor': args.floor, 'simulations': int(total), 'rows': int(len(rows)), 'without_peak': int((counts == 0).sum()),
126
+ 'mean_annotated_peaks': float(counts.mean()), 'seconds': time.monotonic() - started}
127
+ Path(str(out).replace('.npz', '.manifest.json')).write_text(json.dumps(report, indent=1) + '\n')
128
+ print(json.dumps(report), flush=True)
129
+
130
+
131
+ def score(args):
132
+ import torch
133
+ import inverse
134
+ import invrun
135
+ import netprop
136
+ import scorelab
137
+ device = torch.device('cuda')
138
+ started = time.monotonic()
139
+ queries, pools = netprop.load(None)
140
+ ev = scorelab.Evaluator(queries, pools)
141
+ config = json.load(open(ROOT / args.build / 'model_config.json'))
142
+ pred = {name: pickle.load(open(ROOT / args.pred / f'{name}.pkl', 'rb')) for name in ('distances', 'inverse', 'library', 'ion')}
143
+ base = netprop.released(queries, pools, pred, config)
144
+ data = dict(np.load(ROOT / 'inv/pools-validation-input.npz'))
145
+ saved = torch.load(ROOT / args.build / 'model.pt', map_location='cpu', weights_only=False)
146
+ models = [invrun.build_model(member, device) for member in saved['inverse']]
147
+ del saved
148
+ store = invrun.PoolStore(data, device)
149
+ cyc = dict(np.load(ROOT / args.store))
150
+ max_peaks = inverse.MAX_PEAKS
151
+ counts = np.diff(cyc['offset'])
152
+ row_of_task = cyc['row_of_task']
153
+ c_offset, c_peaks, c_inten = cyc['offset'], cyc['peaks'], cyc['inten']
154
+ c_root = torch.from_numpy(cyc['root'].astype(np.int16))
155
+ c_adduct, c_instrument, c_energy = (torch.from_numpy(cyc[k].astype(np.int64) if k != 'energy' else cyc[k]) for k in ('adduct', 'instrument', 'energy'))
156
+ groups_of = {}
157
+ for g, s in enumerate(cyc['group_pool']):
158
+ groups_of.setdefault(int(s), []).append(g)
159
+ g_start, g_tasks, g_qstart, g_queries = cyc['group_start'], cyc['group_tasks'], cyc['group_qstart'], cyc['group_queries']
160
+
161
+ @torch.no_grad()
162
+ def simulated(rows, batch=4096):
163
+ """Mean bit probabilities of the models for the annotated simulation rows ``rows`` (sorted unique)."""
164
+ out = torch.zeros((len(rows), inverse.FP_BITS), dtype=torch.float32, device=device)
165
+ for start in range(0, len(rows), batch):
166
+ part = rows[start:start + batch]
167
+ n = len(part)
168
+ peaks = np.zeros((n, max_peaks, c_peaks.shape[1]), np.uint8)
169
+ inten = np.zeros((n, max_peaks), np.float32)
170
+ for j, r in enumerate(part):
171
+ k = min(int(counts[r]), max_peaks)
172
+ peaks[j, :k] = c_peaks[c_offset[r]:c_offset[r] + k]
173
+ inten[j, :k] = c_inten[c_offset[r]:c_offset[r] + k]
174
+ count = torch.from_numpy(np.minimum(counts[part], max_peaks)).to(device)
175
+ mask = torch.arange(max_peaks, device=device).unsqueeze(0) < count.unsqueeze(1)
176
+ index = torch.from_numpy(part)
177
+ with torch.autocast('cuda', dtype=torch.bfloat16):
178
+ for model in models:
179
+ y = model(c_root[index].to(device).long(), torch.from_numpy(peaks).to(device).long(), torch.from_numpy(inten).to(device), mask,
180
+ c_adduct[index].to(device), c_instrument[index].to(device), c_energy[index].to(device))
181
+ out[start:start + n] += torch.sigmoid(y.float()) / len(models)
182
+ return out
183
+
184
+ def cosine(a, b):
185
+ return 1.0 - (a * b).sum(-1) / ((a * a).sum(-1).sqrt() * (b * b).sum(-1).sqrt()).clamp_min(1e-9)
186
+
187
+ names = ['inv', 'cycA', 'cycA1', 'cycG', 'cycC', 'cycD', 'cycK', 'cycS']
188
+ terms = {name: [None] * len(queries) for name in names}
189
+ covered = 0
190
+ for number, group in enumerate(store.groups(None, 2000)):
191
+ low, high = int(store.pair_start[group[0]]), int(store.pair_start[group[-1] + 1])
192
+ logits = [store.logits(model, low, high, batch=2048) for model in models]
193
+ blank = [store.logits(model, low, high, batch=2048, blank=True) for model in models]
194
+ need = np.unique(np.concatenate([g_tasks[g_start[g]:g_start[g + 1]] for s in group for g in groups_of.get(s, [])] + [np.zeros(0, np.int64)]))
195
+ need = need[need >= 0]
196
+ rows = np.unique(row_of_task[need])
197
+ rows = rows[rows >= 0]
198
+ psim_all = simulated(rows)
199
+ for s in group:
200
+ members = store.members[s]
201
+ nq, nf = len(members), int(store.formula_start[s + 1] - store.formula_start[s])
202
+ if nq == 0 or nf == 0:
203
+ continue
204
+ span = slice(int(store.pair_start[s]) - low, int(store.pair_start[s + 1]) - low)
205
+ per_query = torch.stack([torch.sigmoid(x[span].view(nq, nf, -1)) for x in logits]).mean(0) # [nq, nf, bits]
206
+ pobs = per_query.mean(0)
207
+ pblank = torch.stack([torch.sigmoid(x[span].view(nq, nf, -1)).mean(0) for x in blank]).mean(0)
208
+ cs = slice(int(store.cand_start[s]), int(store.cand_start[s + 1]))
209
+ fidx = store.cand_formula[cs]
210
+ bits = inverse.unpack_bits(store.cand_fps[cs]).float()
211
+ ok = fidx >= 0
212
+ fsafe = fidx.clamp_min(0)
213
+ ncand = len(fidx)
214
+ position = {int(q): n for n, q in enumerate(members)}
215
+ psim = torch.zeros((ncand, inverse.FP_BITS), device=device)
216
+ weight = torch.zeros(ncand, device=device)
217
+ d_group = torch.zeros(ncand, device=device)
218
+ for g in groups_of.get(s, []):
219
+ tasks = g_tasks[g_start[g]:g_start[g + 1]]
220
+ r = np.where(tasks >= 0, row_of_task[np.maximum(tasks, 0)], -1)
221
+ good = torch.from_numpy(r >= 0).to(device)
222
+ where = torch.from_numpy(np.searchsorted(rows, np.maximum(r, 0))).to(device).clamp_max(max(len(rows) - 1, 0))
223
+ part = psim_all[where] * good[:, None] if len(rows) else torch.zeros((ncand, inverse.FP_BITS), device=device)
224
+ gq = [position[int(q)] for q in g_queries[g_qstart[g]:g_qstart[g + 1]] if int(q) in position]
225
+ share = len(gq) / nq
226
+ psim += share * part
227
+ weight += share * good
228
+ pg = per_query[gq].mean(0)[fsafe]
229
+ d_group += share * torch.where(good, cosine(pg ** 0.7, part ** 0.7), torch.ones_like(weight))
230
+ seen = weight > 0
231
+ psim = psim / weight.clamp_min(1e-9)[:, None]
232
+ po, pb = pobs[fsafe], pblank[fsafe]
233
+ out = {
234
+ 'inv': cosine(po ** 0.7, bits),
235
+ 'cycA': cosine(po ** 0.7, psim ** 0.7),
236
+ 'cycA1': cosine(po, psim),
237
+ 'cycG': d_group,
238
+ 'cycC': cosine(po - pb, psim - pb),
239
+ 'cycK': (psim * (torch.log(psim.clamp_min(1e-4)) - torch.log(po.clamp_min(1e-4)))
240
+ + (1 - psim) * (torch.log((1 - psim).clamp_min(1e-4)) - torch.log((1 - po).clamp_min(1e-4)))).mean(-1) * 10.0,
241
+ 'cycS': cosine(po ** 0.7, 0.5 * (psim ** 0.7 + bits)),
242
+ }
243
+ rel = 1.0 - (bits - psim).abs()
244
+ out['cycD'] = 1.0 - (rel * po ** 0.7 * bits).sum(-1) / ((rel * po ** 1.4).sum(-1).sqrt() * (rel * bits).sum(-1).sqrt()).clamp_min(1e-9)
245
+ covered += int((seen & ok).sum())
246
+ for name, value in out.items():
247
+ usable = ok if name == 'inv' else (ok & seen)
248
+ worst = float(value[usable].max()) if bool(usable.any()) else 1.0
249
+ value = torch.where(usable, value, torch.full_like(value, worst)).cpu().numpy().astype(np.float64)
250
+ for q in members:
251
+ terms[name][int(q)] = value
252
+ del logits, blank, psim_all
253
+ if number % 40 == 0:
254
+ print(f' run {number} {time.monotonic() - started:.0f} s', flush=True)
255
+ print(f'terms in {time.monotonic() - started:.0f} s; candidates with a simulated prediction {covered}', flush=True)
256
+ if args.save:
257
+ (ROOT / args.save).parent.mkdir(parents=True, exist_ok=True)
258
+ pickle.dump(terms, open(ROOT / args.save, 'wb'))
259
+ evaluate(args, queries, ev, base, terms, names)
260
+
261
+
262
+ def evaluate(args, queries, ev, base, terms, names):
263
+ released_term = pickle.load(open(ROOT / args.released_term, 'rb'))['term']
264
+ released_term = released_term[next(iter(released_term))]
265
+ net = [b if (b is None or t is None) else b + t for b, t in zip(base, released_term)]
266
+ members = {}
267
+ for i, q in enumerate(queries):
268
+ members.setdefault(q['candidate_pool_id'], []).append(i)
269
+ plist = list(members)
270
+ first_q = np.array([members[pid][0] for pid in plist])
271
+ sizes = np.array([len(members[pid]) for pid in plist], dtype=float)
272
+ capped = np.minimum(sizes, 30.0)
273
+ boot, domain = ev.boot, ev.domain
274
+
275
+ def report(label, ref, fused):
276
+ rank, rank0 = ev.ranks(fused, 1.0), ref
277
+ hit, hit0 = rank == 0, rank0 == 0
278
+ d = hit[first_q].astype(float) - hit0[first_q].astype(float)
279
+ macro = d[boot].mean(axis=1)
280
+ per_query = np.sum(d[boot] * sizes[boot], axis=1) / np.sum(sizes[boot], axis=1)
281
+ d5 = (rank[first_q] < 5).astype(float) - (rank0[first_q] < 5).astype(float)
282
+ print(f' {label}: Hit@1 {hit.mean():.4f} ({hit.mean() - hit0.mean():+.4f} [{np.percentile(per_query, 2.5):+.4f}, {np.percentile(per_query, 97.5):+.4f}]) macro {d.mean():+.4f} '
283
+ f'[{np.percentile(macro, 2.5):+.4f}, {np.percentile(macro, 97.5):+.4f}] pools +{int((d > 0).sum())} -{int((d < 0).sum())} | cap30 {np.sum(d * capped) / capped.sum():+.4f} '
284
+ f'| Hit@5 {np.mean(rank < 5) - np.mean(rank0 < 5):+.4f} macro@5 {d5.mean():+.4f} | dom {hit[domain].mean() - hit0[domain].mean():+.4f} oth {hit[~domain].mean() - hit0[~domain].mean():+.4f}', flush=True)
285
+
286
+ base_rank, net_rank = ev.ranks(base, 1.0), ev.ranks(net, 1.0)
287
+ print(f'base (release without the consensus term) Hit@1 {np.mean(base_rank == 0):.4f} macro {np.mean(base_rank[first_q] == 0):.4f}; '
288
+ f'net (the release) Hit@1 {np.mean(net_rank == 0):.4f} macro {np.mean(net_rank[first_q] == 0):.4f}')
289
+ lams = [float(x) for x in args.lams.split(',')]
290
+ for name in names:
291
+ term = terms[name]
292
+ alone = ev.ranks(term, 1.0)
293
+ print(f'\n{name}: alone Hit@1 {np.mean(alone == 0):.4f} macro {np.mean(alone[first_q] == 0):.4f} | dom {np.mean(alone[domain] == 0):.4f} oth {np.mean(alone[~domain] == 0):.4f}')
294
+ for lam in lams:
295
+ report(f'base + {lam:g} x term', base_rank, [b if (b is None or t is None) else b + lam * t for b, t in zip(base, term)])
296
+ for lam in lams:
297
+ report(f'net + {lam:g} x term', net_rank, [b if (b is None or t is None) else b + lam * t for b, t in zip(net, term)])
298
+
299
+
300
+ def rescore(args):
301
+ import netprop
302
+ import scorelab
303
+ queries, pools = netprop.load(None)
304
+ ev = scorelab.Evaluator(queries, pools)
305
+ config = json.load(open(ROOT / args.build / 'model_config.json'))
306
+ pred = {name: pickle.load(open(ROOT / args.pred / f'{name}.pkl', 'rb')) for name in ('distances', 'inverse', 'library', 'ion')}
307
+ base = netprop.released(queries, pools, pred, config)
308
+ terms = pickle.load(open(ROOT / args.save, 'rb'))
309
+ evaluate(args, queries, ev, base, terms, [n for n in (args.names.split(',') if args.names else terms)])
310
+
311
+
312
+ def main():
313
+ parser = argparse.ArgumentParser()
314
+ sub = parser.add_subparsers(dest='command', required=True)
315
+ p = sub.add_parser('prepare')
316
+ p.add_argument('--out', required=True)
317
+ p.add_argument('--members', default='evals/B30k-full/pred,evals/B2-30k-full/pred')
318
+ p.add_argument('--floor', type=float, default=0.02)
319
+ p.add_argument('--workers', type=int, default=5)
320
+ for name in ('score', 'rescore'):
321
+ p = sub.add_parser(name)
322
+ p.add_argument('--store')
323
+ p.add_argument('--build', default='builds/5Ms-I9b+I13-L-ion-net')
324
+ p.add_argument('--pred', default='evals/fullfold/5Ms-I9b+I13-L-ion/pred')
325
+ p.add_argument('--released-term', default='evals/netprop/term-m10-half05-ties.pkl')
326
+ p.add_argument('--lams', default='0.1,0.25,0.5,1')
327
+ p.add_argument('--save')
328
+ p.add_argument('--names')
329
+ args = parser.parse_args()
330
+ {'prepare': prepare, 'score': score, 'rescore': rescore}[args.command](args)
331
+
332
+
333
+ if __name__ == '__main__':
334
+ main()
shared/package-files/26/267e0693fa268f2bfb165485332ee52eaf290e5da9280077acc80b5a34ff97d2/netprop-eval-cap08.log ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ released rule from saved branches: Hit@1 0.7713 pool-macro 0.8591 Hit@5 0.8797 Hit@20 0.9377 | twin pools 354 (macro 0.7429), others macro 0.8727
2
+
3
+ k=10 p=6.0 norm | m=1 tau=1.0 kernel=cap=0.8 debias=fold control=none shuffle=None: alone Hit@1 0.4150 (domain 0.3974 / other 0.4327) macro 0.3582
4
+ lam 0.00/0.25: Hit@1 0.7749 (+0.0036 [-0.0190, +0.0228]) macro +0.0065 [+0.0021, +0.0106] pools +40 -18 | no top 20 +0.0144 cap30 +0.0105 [-0.0012, +0.0225] | Hit@5 -0.0049 macro@5 +0.0065 cap30@5 +0.0039 | twin +11 -3 other +29 -15 | dom -0.0009 oth +0.0080
5
+ lam 0.00/0.50: Hit@1 0.7728 (+0.0014 [-0.0227, +0.0218]) macro +0.0080 [+0.0030, +0.0127] pools +51 -24 | no top 20 +0.0155 cap30 +0.0099 [-0.0028, +0.0237] | Hit@5 -0.0021 macro@5 +0.0062 cap30@5 +0.0070 | twin +15 -4 other +36 -20 | dom -0.0010 oth +0.0039
6
+ lam 0.50/0.50: Hit@1 0.7718 (+0.0005 [-0.0235, +0.0209]) macro +0.0089 [+0.0035, +0.0139] pools +62 -32 | no top 20 +0.0144 cap30 +0.0087 [-0.0053, +0.0235] | Hit@5 -0.0022 macro@5 +0.0062 cap30@5 +0.0068 | twin +19 -5 other +43 -27 | dom -0.0026 oth +0.0035
shared/package-files/2c/2c1d1d1556dcc6cd38eabf319f9a059e6b4f0a73b84d64cfe6ea5b09eb9c385d/eval-score-I30-on-merged-10-freq.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 1 members, 36.3 s
2
+ cos pool rule all=0.5875 dom=0.7836 non=0.3906 h5=0.7094 macro=0.6754
3
+ cen pool rule all=0.5618 dom=0.7458 non=0.3770 h5=0.6992 macro=0.6364
4
+ cos_pm pool rule all=0.5875 dom=0.7836 non=0.3906 h5=0.7094 macro=0.6754
5
+ dot pool rule all=0.6178 dom=0.7718 non=0.4632 h5=0.7562 macro=0.6666
6
+ geo pool rule all=0.5830 dom=0.7819 non=0.3833 h5=0.6940 macro=0.6719
7
+ idf05 pool rule all=0.5983 dom=0.7892 non=0.4066 h5=0.7224 macro=0.6837
8
+ idf1 pool rule all=0.6016 dom=0.7878 non=0.4146 h5=0.7235 macro=0.6831
9
+ lin pool rule all=0.5662 dom=0.7446 non=0.3871 h5=0.6889 macro=0.6382
10
+ log01 pool rule all=0.6284 dom=0.7940 non=0.4622 h5=0.7613 macro=0.6831
11
+ log05 pool rule all=0.6347 dom=0.7937 non=0.4750 h5=0.7599 macro=0.6887
12
+ log20 pool rule all=0.5990 dom=0.7893 non=0.4079 h5=0.7205 macro=0.6864
13
+ nll02 pool rule all=0.4388 dom=0.6620 non=0.2147 h5=0.5992 macro=0.5328
14
+ nll10 pool rule all=0.4026 dom=0.6039 non=0.2006 h5=0.5498 macro=0.4829
15
+ poolcen pool rule all=0.5984 dom=0.7397 non=0.4566 h5=0.7391 macro=0.6350
16
+ pow020 pool rule all=0.5765 dom=0.7642 non=0.3879 h5=0.7447 macro=0.6542
17
+ pow030 pool rule all=0.5923 dom=0.7827 non=0.4010 h5=0.7571 macro=0.6675
18
+ pow040 pool rule all=0.6306 dom=0.7913 non=0.4691 h5=0.7677 macro=0.6802
19
+ pow05 pool rule all=0.6321 dom=0.7952 non=0.4685 h5=0.7576 macro=0.6875
20
+ pow05idf025 pool rule all=0.6301 dom=0.7915 non=0.4682 h5=0.7616 macro=0.6846
21
+ pow05idf05 pool rule all=0.6298 dom=0.7940 non=0.4649 h5=0.7642 macro=0.6875
22
+ pow05m pool rule all=0.6321 dom=0.7952 non=0.4685 h5=0.7576 macro=0.6875
23
+ pow05s pool rule all=0.6290 dom=0.7963 non=0.4610 h5=0.7564 macro=0.6878
24
+ pow060 pool rule all=0.6315 dom=0.7908 non=0.4716 h5=0.7562 macro=0.6875
25
+ pow060idf025 pool rule all=0.6340 dom=0.7935 non=0.4739 h5=0.7583 macro=0.6911
26
+ pow060idf050 pool rule all=0.6314 dom=0.7934 non=0.4687 h5=0.7625 macro=0.6893
27
+ pow060m pool rule all=0.6315 dom=0.7908 non=0.4716 h5=0.7562 macro=0.6875
28
+ pow065 pool rule all=0.6017 dom=0.7915 non=0.4111 h5=0.7270 macro=0.6872
29
+ pow070 pool rule all=0.5993 dom=0.7910 non=0.4067 h5=0.7228 macro=0.6867
30
+ pow070idf025 pool rule all=0.6034 dom=0.7920 non=0.4141 h5=0.7252 macro=0.6881
31
+ pow070idf050 pool rule all=0.6331 dom=0.7920 non=0.4735 h5=0.7554 macro=0.6893
32
+ pow070m pool rule all=0.5993 dom=0.7910 non=0.4067 h5=0.7228 macro=0.6867
33
+ pow075 pool rule all=0.5995 dom=0.7906 non=0.4076 h5=0.7185 macro=0.6869
34
+ pow080 pool rule all=0.5967 dom=0.7886 non=0.4041 h5=0.7164 macro=0.6837
35
+ pow080m pool rule all=0.5967 dom=0.7886 non=0.4041 h5=0.7164 macro=0.6837
36
+ pow085 pool rule all=0.5962 dom=0.7875 non=0.4041 h5=0.7150 macro=0.6828
37
+ pow085idf025 pool rule all=0.5984 dom=0.7890 non=0.4070 h5=0.7189 macro=0.6849
38
+ pow085idf050 pool rule all=0.5989 dom=0.7881 non=0.4090 h5=0.7233 macro=0.6843
39
+ pow090 pool rule all=0.5920 dom=0.7852 non=0.3979 h5=0.7134 macro=0.6799
40
+ pow2 pool rule all=0.5740 dom=0.7625 non=0.3848 h5=0.7007 macro=0.6551
41
+ tani pool rule all=0.5778 dom=0.7665 non=0.3884 h5=0.7219 macro=0.6595
shared/package-files/2c/2c31cd541a7add3868008214c8eaadc44684525c7d58849163797d9a6317c66f/rank_keep.py ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Train-fold table for the rank-dependent thinning of simulated rows (inverse trainer, E70)
2
+
3
+ rank_keep.py [--sim inv/selfsim/cal2-train.npz] [--out inv/rank_keep.json]
4
+
5
+ For the self-simulated calibration rows of train-fold molecules (one of the molecule's own acquisition
6
+ conditions) and up to three measured train rows of the same molecule and condition: the share of
7
+ simulated peaks at intensity rank r (0 = strongest, inverse-model input space, 24 peaks) that the
8
+ measured row contains. Groups: 0 = recorded collision energy, 1 = no energy and [M+H]+, 2 = [M+Na]+.
9
+ Train fold only; no validation row is read.
10
+ """
11
+ import argparse
12
+ import json
13
+ import os
14
+ import sys
15
+ from pathlib import Path
16
+
17
+ import numpy as np
18
+
19
+ ROOT = Path('/workspace/research')
20
+ sys.path.insert(0, str(ROOT / 'allcond/pkg'))
21
+ import inverse
22
+
23
+ GROUPS = ('recorded energy', 'no energy [M+H]+', '[M+Na]+')
24
+
25
+
26
+ def main():
27
+ parser = argparse.ArgumentParser()
28
+ parser.add_argument('--sim', default='inv/selfsim/cal2-train.npz')
29
+ parser.add_argument('--out', default='inv/rank_keep.json')
30
+ parser.add_argument('--peaks', type=int, default=24)
31
+ args = parser.parse_args()
32
+ keys = ('offset', 'peaks', 'inten', 'adduct', 'instrument', 'energy', 'mol')
33
+ real = np.load(ROOT / 'inv/real-train.npz')
34
+ sim = np.load(ROOT / args.sim)
35
+ R = {k: real[k] for k in keys}
36
+ S = {k: sim[k] for k in keys}
37
+ code = lambda e: -1 if np.isnan(e) else int(round(float(e)))
38
+ sodium = inverse.ADDUCTS.index('[M+Na]+')
39
+ index = {}
40
+ for i in range(len(R['mol'])):
41
+ index.setdefault((int(R['mol'][i]), int(R['adduct'][i]), int(R['instrument'][i]), code(R['energy'][i])), []).append(i)
42
+ total = np.zeros((len(GROUPS), args.peaks))
43
+ found = np.zeros((len(GROUPS), args.peaks))
44
+ pairs = np.zeros(len(GROUPS), dtype=np.int64)
45
+ for j in range(len(S['mol'])):
46
+ cond = (int(S['adduct'][j]), int(S['instrument'][j]), code(S['energy'][j]))
47
+ rows = index.get((int(S['mol'][j]),) + cond)
48
+ sp = S['peaks'][S['offset'][j]:S['offset'][j + 1]][:args.peaks]
49
+ if not rows or len(sp) == 0:
50
+ continue
51
+ group = 2 if cond[0] == sodium else (1 if cond[2] < 0 else 0)
52
+ for i in rows[:3]:
53
+ rp = R['peaks'][R['offset'][i]:R['offset'][i + 1]][:args.peaks]
54
+ if len(rp) == 0:
55
+ continue
56
+ rset = set(tuple(p) for p in rp.tolist())
57
+ hit = np.array([tuple(p) in rset for p in sp.tolist()], dtype=np.float64)
58
+ total[group, :len(sp)] += 1
59
+ found[group, :len(sp)] += hit
60
+ pairs[group] += 1
61
+ keep = found / np.maximum(total, 1)
62
+ # ranks beyond the data of a group (none here) would keep the last measured value
63
+ for g, name in enumerate(GROUPS):
64
+ print(f'{name} | pairs {pairs[g]} | ' + ' '.join(f'{v:.2f}' for v in keep[g]))
65
+ table = {'schema': 'allcond-inverse-rank-keep/v1', 'groups': list(GROUPS), 'peaks': args.peaks,
66
+ 'source': f'{args.sim} against inv/real-train.npz (train fold)', 'pairs': pairs.tolist(),
67
+ 'keep': [[round(float(v), 4) for v in row] for row in keep]}
68
+ out = ROOT / args.out
69
+ with open(str(out) + '.tmp', 'w') as handle:
70
+ json.dump(table, handle, indent=1)
71
+ os.replace(str(out) + '.tmp', out)
72
+ print('wrote', out)
73
+
74
+
75
+ if __name__ == '__main__':
76
+ main()
shared/package-files/2c/2cb8e5cf99660ebdb7e2de530ffa2b0355e1d28cfdc9a5ff91f0dca005adfef3/B45a-anneal6k.log ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {"event": "resumed", "source": "/workspace/research/forks/B-step0045000-resume.ckpt", "step": 45000, "epoch": 7, "batch_in_epoch": 3644, "elapsed": 12964.771320266067, "skipped": 0}
2
+ {"parameters": 15108154, "train_rows": 189066, "validation_rows": 2048, "steps_per_epoch": 5908, "identity": "own-checkpoint:e50ac2416118a04bedebae881add432fa98fb6f5ec13dd9ad5cdaf13cf4b3ad1", "gpu": "NVIDIA H100 80GB HBM3"}
3
+ /opt/glacier/lib/python3.11/site-packages/torch/nn/modules/transformer.py:508: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)
4
+ output = torch._nested_tensor_from_mask(
5
+ {"step": 46000, "elapsed_seconds": 13259.363614329064, "validation_loss": 0.5682799583300948, "validation_loss_ema": 0.5617907987907529}
6
+ {"step": 47000, "elapsed_seconds": 13547.064763365066, "validation_loss": 0.5652839476242661, "validation_loss_ema": 0.5623764898627996}
7
+ {"step": 48000, "elapsed_seconds": 13877.90905588807, "validation_loss": 0.5649867970496416, "validation_loss_ema": 0.5618297383189201}
8
+ {"step": 49000, "elapsed_seconds": 14180.89154956107, "validation_loss": 0.562397638335824, "validation_loss_ema": 0.5612926306203008}
9
+ {"step": 50000, "elapsed_seconds": 14465.758826909063, "validation_loss": 0.5609858660027385, "validation_loss_ema": 0.5604573851451278}
10
+ {"step": 51000, "elapsed_seconds": 14707.53600471506, "validation_loss": 0.5606449376791716, "validation_loss_ema": 0.5602231975644827}
shared/package-files/38/38a8667ef1302bb086a90b575ea6e9db125097721da1b0ff0210243969fedf30/fullfold-5Ms-I9b-L.report.json ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "benchmark_score": {
3
+ "at_or_above_upper_reference": true,
4
+ "comparable_to_test_reference": false,
5
+ "direction": "maximize",
6
+ "evaluation_fold": "validation",
7
+ "lower_reference": 0.04004329004329004,
8
+ "metric": "hit_rate_at_1",
9
+ "raw_performance": 0.7590714910700499,
10
+ "saturation": "upper",
11
+ "score": 1.0,
12
+ "unclipped_score": 1.5842443646116713,
13
+ "upper_reference": 0.4939052175894281,
14
+ "version": "massspecgym-retrieval-score/v1"
15
+ },
16
+ "external_reference_context": {
17
+ "comparable_to_this_validation_report": false,
18
+ "fold": "test",
19
+ "hit_rate_at_1": 0.4939052175894281,
20
+ "name": "GLACIER no-CF with strict-clean MIST fallback"
21
+ },
22
+ "global_error": null,
23
+ "metrics": {
24
+ "hit_rate_at_1": 0.7590714910700499,
25
+ "hit_rate_at_20": 0.9363837562406712,
26
+ "hit_rate_at_5": 0.8761130269185239
27
+ },
28
+ "paired_baseline_delta": {
29
+ "confidence": 0.95,
30
+ "estimate": 0.006588089968603634,
31
+ "grouping": "supplied target_inchikey connectivity block",
32
+ "lower": -0.0017486762663796026,
33
+ "molecule_groups": 3185,
34
+ "queries": 19429,
35
+ "resamples": 2000,
36
+ "seed": 0,
37
+ "statistic": "spectrum-weighted mean; molecules sampled with replacement",
38
+ "upper": 0.014745243737668348
39
+ },
40
+ "queries": 19429,
41
+ "schema": "massspecgym-retrieval-development-report/v1",
42
+ "strata": {
43
+ "adduct": {
44
+ "[M+H]+": {
45
+ "hit_rate_at_1": 0.8283419548053362,
46
+ "queries": 14692
47
+ },
48
+ "[M+Na]+": {
49
+ "hit_rate_at_1": 0.5442263035676589,
50
+ "queries": 4737
51
+ }
52
+ },
53
+ "candidate_pool_size": {
54
+ "129-256": {
55
+ "hit_rate_at_1": 0.7601980300139239,
56
+ "queries": 19391
57
+ },
58
+ "33-128": {
59
+ "hit_rate_at_1": 0.18421052631578946,
60
+ "queries": 38
61
+ }
62
+ },
63
+ "collision_energy": {
64
+ "20-39.999": {
65
+ "hit_rate_at_1": 0.9346868884540117,
66
+ "queries": 4088
67
+ },
68
+ "40-59.999": {
69
+ "hit_rate_at_1": 0.9093809827696235,
70
+ "queries": 1567
71
+ },
72
+ "<20": {
73
+ "hit_rate_at_1": 0.825,
74
+ "queries": 1200
75
+ },
76
+ ">=60": {
77
+ "hit_rate_at_1": 0.9397435897435897,
78
+ "queries": 3120
79
+ },
80
+ "missing": {
81
+ "hit_rate_at_1": 0.5902263592130316,
82
+ "queries": 9454
83
+ }
84
+ },
85
+ "instrument_type": {
86
+ "": {
87
+ "hit_rate_at_1": 0.7407407407407407,
88
+ "queries": 486
89
+ },
90
+ "Orbitrap": {
91
+ "hit_rate_at_1": 0.7518097207859359,
92
+ "queries": 16439
93
+ },
94
+ "QTOF": {
95
+ "hit_rate_at_1": 0.8103035143769968,
96
+ "queries": 2504
97
+ }
98
+ },
99
+ "routing_domain": {
100
+ "fallback": {
101
+ "hit_rate_at_1": 0.5947395564724085,
102
+ "queries": 9695
103
+ },
104
+ "glacier_released": {
105
+ "hit_rate_at_1": 0.9227450174645572,
106
+ "queries": 9734
107
+ }
108
+ }
109
+ },
110
+ "uncertainty": {
111
+ "confidence": 0.95,
112
+ "estimate": 0.7590714910700499,
113
+ "grouping": "supplied target_inchikey connectivity block",
114
+ "lower": 0.7022166703804578,
115
+ "molecule_groups": 3185,
116
+ "queries": 19429,
117
+ "resamples": 2000,
118
+ "seed": 0,
119
+ "statistic": "spectrum-weighted mean; molecules sampled with replacement",
120
+ "upper": 0.806398649397918
121
+ },
122
+ "valid_rankings": 19429
123
+ }
shared/package-files/38/38ed421583fc85afbabc6ad85c91cec0d699f9c2c327af4236588876936bac6c/eval-I3-50k-100k.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ super().__init__(open(name, mode))
2
+ ^^^^^^^^^^^^^^^^
3
+ FileNotFoundError: [Errno 2] No such file or directory: 'runs/I3-ff512x6-real08/step-050000-ema.pt'
4
+ 19429 queries, 1066979 pairs, 24.2 s
shared/package-files/3a/3a37da7fa50fabb4093d8876f075fcf95fb76c68b4cc2c5c3435c592f4016061/fullfold-5Ms-I9b-L.sh ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # Full validation fold with the packaged infer.py and the five members' saved simulations:
3
+ # candidate build and the current release (paired baseline), then the official scorer.
4
+ cd /workspace/research
5
+ export CUDA_MPS_PIPE_DIRECTORY=/tmp/mps
6
+ PY=/opt/glacier/bin/python
7
+ N=5Ms-I9b-L; B=5Ms-I1+I3x2g07-L
8
+ mkdir -p evals/fullfold/$N evals/fullfold/$B
9
+ $PY allcond/checks/measure.py taskset -c 14-21 $PY builds/$N/infer.py validation-input evals/fullfold/$N --load-predictions evals/stack/5M-S2-30k --workers 7 > logs/fullfold-$N-saved-simulations.log 2>&1
10
+ $PY allcond/checks/measure.py taskset -c 14-21 $PY builds/$B/infer.py validation-input evals/fullfold/$B --load-predictions evals/stack/5M-S2-30k --workers 7 > logs/fullfold-$B-saved-simulations-rerun2.log 2>&1
11
+ OMP_NUM_THREADS=4 taskset -c 14-21 $PY /workspace/evaluation/score.py --input-root /workspace/data --rankings evals/fullfold/$B/rankings.jsonl --output logs/fullfold-$B-rerun2.report.json > /dev/null 2>&1
12
+ OMP_NUM_THREADS=4 taskset -c 14-21 $PY /workspace/evaluation/score.py --input-root /workspace/data --rankings evals/fullfold/$N/rankings.jsonl --baseline-rankings evals/fullfold/$B/rankings.jsonl --output logs/fullfold-$N.report.json > logs/fullfold-$N.score.log 2>&1
13
+ date -u +%H:%M:%S > logs/fullfold-$N.done
shared/package-files/3a/3a3b4478d48856473c3dbd30b55c98a2c401226853ab7182d40bd9387deb7f41/report-ind.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "benchmark_score": {
3
+ "at_or_above_upper_reference": false,
4
+ "comparable_to_test_reference": false,
5
+ "direction": "maximize",
6
+ "evaluation_fold": "validation",
7
+ "lower_reference": 0.04004329004329004,
8
+ "metric": "hit_rate_at_1",
9
+ "raw_performance": 0.47275267815556593,
10
+ "saturation": null,
11
+ "score": 0.9533943295305115,
12
+ "unclipped_score": 0.9533943295305115,
13
+ "upper_reference": 0.4939052175894281,
14
+ "version": "massspecgym-retrieval-score/v1"
15
+ },
16
+ "external_reference_context": {
17
+ "comparable_to_this_validation_report": false,
18
+ "fold": "test",
19
+ "hit_rate_at_1": 0.4939052175894281,
20
+ "name": "GLACIER no-CF with strict-clean MIST fallback"
21
+ },
22
+ "global_error": null,
23
+ "metrics": {
24
+ "hit_rate_at_1": 0.47275267815556593,
25
+ "hit_rate_at_20": 0.680018630647415,
26
+ "hit_rate_at_5": 0.592454587796926
27
+ },
28
+ "queries": 2147,
29
+ "schema": "massspecgym-retrieval-development-report/v1",
30
+ "strata": {
31
+ "adduct": {
32
+ "[M+H]+": {
33
+ "hit_rate_at_1": 0.5850422195416164,
34
+ "queries": 1658
35
+ },
36
+ "[M+Na]+": {
37
+ "hit_rate_at_1": 0.09202453987730061,
38
+ "queries": 489
39
+ }
40
+ },
41
+ "candidate_pool_size": {
42
+ "129-256": {
43
+ "hit_rate_at_1": 0.47275267815556593,
44
+ "queries": 2147
45
+ }
46
+ },
47
+ "collision_energy": {
48
+ "20-39.999": {
49
+ "hit_rate_at_1": 0.71484375,
50
+ "queries": 512
51
+ },
52
+ "40-59.999": {
53
+ "hit_rate_at_1": 0.7288135593220338,
54
+ "queries": 177
55
+ },
56
+ "<20": {
57
+ "hit_rate_at_1": 0.5099337748344371,
58
+ "queries": 151
59
+ },
60
+ ">=60": {
61
+ "hit_rate_at_1": 0.7586206896551724,
62
+ "queries": 348
63
+ },
64
+ "missing": {
65
+ "hit_rate_at_1": 0.18665276329509906,
66
+ "queries": 959
67
+ }
68
+ },
69
+ "instrument_type": {
70
+ "": {
71
+ "hit_rate_at_1": 0.5714285714285714,
72
+ "queries": 35
73
+ },
74
+ "Orbitrap": {
75
+ "hit_rate_at_1": 0.4828877005347594,
76
+ "queries": 1870
77
+ },
78
+ "QTOF": {
79
+ "hit_rate_at_1": 0.38016528925619836,
80
+ "queries": 242
81
+ }
82
+ },
83
+ "routing_domain": {
84
+ "fallback": {
85
+ "hit_rate_at_1": 0.18668012108980828,
86
+ "queries": 991
87
+ },
88
+ "glacier_released": {
89
+ "hit_rate_at_1": 0.7179930795847751,
90
+ "queries": 1156
91
+ }
92
+ }
93
+ },
94
+ "uncertainty": {
95
+ "confidence": 0.95,
96
+ "estimate": 0.47275267815556593,
97
+ "grouping": "supplied target_inchikey connectivity block",
98
+ "lower": 0.3972122998952267,
99
+ "molecule_groups": 413,
100
+ "queries": 2147,
101
+ "resamples": 2000,
102
+ "seed": 0,
103
+ "statistic": "spectrum-weighted mean; molecules sampled with replacement",
104
+ "upper": 0.5597915299028767
105
+ },
106
+ "valid_rankings": 2147
107
+ }
shared/package-files/3a/3a94f027d9c7dc0bff6c6f3f083e12827079f68fe14936be6322b7dbae19fed3/swap-I16+I18+I20-w1.0_0.05_0.5.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ released rule from saved branches: Hit@1 0.7710 pool-macro 0.8641 Hit@5 0.8783 Hit@20 0.9334 | twin pools 354 (macro 0.7655), others macro 0.8757
2
+
3
+ k=10 p=3.0 half=0.5 | m=10 tau=0.02 kernel=plain fps=morgan debias=fold control=none shuffle=None: alone Hit@1 0.4485 (domain 0.4195 / other 0.4777) macro 0.3819
4
+ lam 1.00/1.00: Hit@1 0.8044 (+0.0334 [+0.0138, +0.0584]) macro +0.0145 [+0.0083, +0.0207] pools +84 -35 | no top 20 +0.0274 cap30 +0.0267 [+0.0141, +0.0414] | Hit@5 -0.0032 macro@5 +0.0074 cap30@5 +0.0148
5
+ without consensus term: Hit@1 0.7710 (ref 0.7713; -0.0003 [-0.0074, +0.0052]) macro 0.8641 (+0.0050 [+0.0006, +0.0095]) pools +40 -23 cap30 +0.0004 | Hit@5 0.8783 (-0.0014) Hit@20 0.9334 (-0.0042)
6
+ domain 0.9358 (+0.0044) other 0.6056 (-0.0051) | Orbitrap no energy 0.5690 (-0.0074) | above 850 0.4237 (+0.0092) below 850 0.8501 (-0.0025)
7
+ full rule : Hit@1 0.8044 (ref 0.8022; +0.0023 [-0.0035, +0.0082]) macro 0.8786 (+0.0047 [+0.0003, +0.0089]) pools +36 -20 cap30 +0.0026 | Hit@5 0.8751 (+0.0032) Hit@20 0.9168 (+0.0005)
8
+ domain 0.9398 (+0.0034) other 0.6685 (+0.0011) | Orbitrap no energy 0.6376 (-0.0026) | above 850 0.5372 (-0.0075) below 850 0.8653 (+0.0045)
shared/package-files/3c/3c02d5172d26f1587777ace6f7dacea73daab5aaeb92cea81a5670c43cde5cef/netprop-released-sub-0_25-2.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ released rule from saved branches: Hit@1 0.7629 pool-macro 0.8404 Hit@5 0.8640 Hit@20 0.9229 | twin pools 95 (macro 0.7263), others macro 0.8549
2
+
3
+ k=10 p=3.0 half=0.5 | m=10 tau=0.02 kernel=plain debias=fold control=none shuffle=None: alone Hit@1 0.2613 (domain 0.1879 / other 0.3335) macro 0.2175
4
+ lam 1.00/1.00: Hit@1 0.7761 (+0.0132 [-0.0112, +0.0371]) macro +0.0130 [+0.0012, +0.0236] pools +18 -7 | no top 20 +0.0145 cap30 +0.0154 [-0.0112, +0.0414] | Hit@5 +0.0081 macro@5 +0.0059 cap30@5 +0.0080
shared/package-files/3c/3c140bc0e96a66db3c0d7451b4fda9cde77161967797935f6fb8dd197679a5a9/eval-score-I12.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 1 members, 74.8 s
2
+ cos pool rule all=0.5774 dom=0.8090 non=0.3448 h5=0.6880 macro=0.6843
3
+ cen pool rule all=0.5569 dom=0.7836 non=0.3292 h5=0.6811 macro=0.6595
4
+ cos_pm pool rule all=0.5774 dom=0.8090 non=0.3448 h5=0.6880 macro=0.6843
5
+ dot pool rule all=0.5719 dom=0.7781 non=0.3649 h5=0.7033 macro=0.6713
6
+ geo pool rule all=0.5504 dom=0.7874 non=0.3124 h5=0.6751 macro=0.6701
7
+ idf05 pool rule all=0.5816 dom=0.8083 non=0.3540 h5=0.6995 macro=0.6872
8
+ idf1 pool rule all=0.5828 dom=0.8136 non=0.3510 h5=0.7017 macro=0.6896
9
+ lin pool rule all=0.5297 dom=0.7473 non=0.3113 h5=0.6544 macro=0.6344
10
+ log01 pool rule all=0.5749 dom=0.8033 non=0.3455 h5=0.7140 macro=0.6843
11
+ log05 pool rule all=0.5809 dom=0.8184 non=0.3425 h5=0.7138 macro=0.6946
12
+ log20 pool rule all=0.5845 dom=0.8178 non=0.3503 h5=0.7048 macro=0.6952
13
+ nll02 pool rule all=0.4138 dom=0.6669 non=0.1596 h5=0.5605 macro=0.5313
14
+ nll10 pool rule all=0.3690 dom=0.5997 non=0.1375 h5=0.5197 macro=0.4722
15
+ poolcen pool rule all=0.5483 dom=0.7715 non=0.3242 h5=0.6824 macro=0.6542
16
+ pow020 pool rule all=0.5431 dom=0.7596 non=0.3256 h5=0.7052 macro=0.6447
17
+ pow030 pool rule all=0.5636 dom=0.7829 non=0.3435 h5=0.7080 macro=0.6630
18
+ pow040 pool rule all=0.5785 dom=0.8037 non=0.3523 h5=0.7166 macro=0.6807
19
+ pow05 pool rule all=0.5814 dom=0.8161 non=0.3457 h5=0.7090 macro=0.6917
20
+ pow05idf025 pool rule all=0.5841 dom=0.8143 non=0.3530 h5=0.7230 macro=0.6929
21
+ pow05idf05 pool rule all=0.5833 dom=0.8179 non=0.3478 h5=0.7247 macro=0.6937
22
+ pow05m pool rule all=0.5814 dom=0.8161 non=0.3457 h5=0.7090 macro=0.6917
23
+ pow05s pool rule all=0.5750 dom=0.8117 non=0.3374 h5=0.6999 macro=0.6890
24
+ pow060 pool rule all=0.5766 dom=0.8180 non=0.3342 h5=0.7086 macro=0.6934
25
+ pow060idf025 pool rule all=0.5834 dom=0.8213 non=0.3445 h5=0.7148 macro=0.6967
26
+ pow060idf050 pool rule all=0.5867 dom=0.8227 non=0.3498 h5=0.7193 macro=0.6964
27
+ pow060m pool rule all=0.5766 dom=0.8180 non=0.3342 h5=0.7086 macro=0.6934
28
+ pow065 pool rule all=0.5775 dom=0.8199 non=0.3341 h5=0.7042 macro=0.6958
29
+ pow070 pool rule all=0.5770 dom=0.8189 non=0.3342 h5=0.6991 macro=0.6940
30
+ pow070idf025 pool rule all=0.5819 dom=0.8222 non=0.3407 h5=0.7086 macro=0.6964
31
+ pow070idf050 pool rule all=0.5824 dom=0.8221 non=0.3418 h5=0.7140 macro=0.6976
32
+ pow070m pool rule all=0.5770 dom=0.8189 non=0.3342 h5=0.6991 macro=0.6940
33
+ pow075 pool rule all=0.5791 dom=0.8176 non=0.3397 h5=0.6977 macro=0.6943
34
+ pow080 pool rule all=0.5802 dom=0.8142 non=0.3452 h5=0.6956 macro=0.6917
35
+ pow080m pool rule all=0.5802 dom=0.8142 non=0.3452 h5=0.6956 macro=0.6917
36
+ pow085 pool rule all=0.5792 dom=0.8135 non=0.3439 h5=0.6936 macro=0.6893
37
+ pow085idf025 pool rule all=0.5855 dom=0.8179 non=0.3522 h5=0.6997 macro=0.6926
38
+ pow085idf050 pool rule all=0.5866 dom=0.8163 non=0.3561 h5=0.7060 macro=0.6934
39
+ pow090 pool rule all=0.5816 dom=0.8141 non=0.3482 h5=0.6900 macro=0.6872
40
+ pow2 pool rule all=0.5330 dom=0.7677 non=0.2973 h5=0.6614 macro=0.6512
41
+ tani pool rule all=0.5421 dom=0.7674 non=0.3158 h5=0.6806 macro=0.6571
shared/package-files/3c/3ca1aa2268d6a4235b90b917bd4ec950f2712c826e93d39952b9ba05e725c5db/library-reverse-small-znw_8.log ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ base (released rule without consensus): Hit@1 0.7685 domain 0.9301 other 0.6062 macro 0.8650
2
+ znw_8 alone per spectrum 0.2180 (0.2150 / 0.2209) pool mean 0.3129 (0.2614 / 0.3647) macro 0.2227
3
+ other w 0.05 (x0.05/sd): Hit@1 0.7594 (-0.0091 [-0.0291, +0.0043]) macro 0.8647 (-0.0003 [-0.0027, +0.0018]) pools +8 -9 | domain 0.9294 other 0.5888 Hit@5 0.8855
4
+ other w 0.10 (x0.05/sd): Hit@1 0.7567 (-0.0118 [-0.0331, +0.0038]) macro 0.8639 (-0.0012 [-0.0044, +0.0024]) pools +15 -19 | domain 0.9297 other 0.5830 Hit@5 0.8880
5
+ other w 0.20 (x0.05/sd): Hit@1 0.7475 (-0.0209 [-0.0452, -0.0002]) macro 0.8615 (-0.0035 [-0.0077, +0.0012]) pools +25 -37 | domain 0.9286 other 0.5658 Hit@5 0.8959
6
+ other w 0.30 (x0.05/sd): Hit@1 0.7536 (-0.0149 [-0.0412, +0.0091]) macro 0.8606 (-0.0044 [-0.0092, +0.0012]) pools +32 -47 | domain 0.9293 other 0.5772 Hit@5 0.8921
7
+ all w 0.05 (x0.05/sd): Hit@1 0.7610 (-0.0075 [-0.0280, +0.0062]) macro 0.8656 (+0.0006 [-0.0024, +0.0035]) pools +15 -13 | domain 0.9321 other 0.5892 Hit@5 0.8858
8
+ all w 0.10 (x0.05/sd): Hit@1 0.7575 (-0.0110 [-0.0323, +0.0047]) macro 0.8639 (-0.0012 [-0.0050, +0.0030]) pools +22 -26 | domain 0.9317 other 0.5826 Hit@5 0.8890
9
+ all w 0.20 (x0.05/sd): Hit@1 0.7471 (-0.0214 [-0.0455, -0.0005]) macro 0.8606 (-0.0044 [-0.0095, +0.0009]) pools +34 -49 | domain 0.9276 other 0.5659 Hit@5 0.8958
10
+ all w 0.30 (x0.05/sd): Hit@1 0.7533 (-0.0152 [-0.0413, +0.0089]) macro 0.8597 (-0.0053 [-0.0112, +0.0012]) pools +46 -64 | domain 0.9282 other 0.5777 Hit@5 0.8934
shared/package-files/56/5606ac500db6e746b0790eaf33948de692ddb40e206e2a256478df7b39fe0160/validation.jsonl ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {"step": 5000, "elapsed": 122.89247310091741, "val_loss": 0.03769003911150656, "val_cos": 0.6351265645909465, "val_loss_ema": 0.03506690962903995, "val_cos_ema": 0.6560519946524043}
2
+ {"step": 10000, "elapsed": 241.18909381492995, "val_loss": 0.03136553374853031, "val_cos": 0.6962334086409169, "val_loss_ema": 0.030148990683526396, "val_cos_ema": 0.7082829277835323, "ind_cos": 0.45205620464254465, "agg_cos": 0.4891656801688198, "ind_nll": 0.3135519069432292, "agg_nll": 0.2895156724484019, "retrieval_seconds": 11.547416899935342}
3
+ {"step": 15000, "elapsed": 354.96333089098334, "val_loss": 0.029885894965448113, "val_cos": 0.7181207809122317, "val_loss_ema": 0.02857240176899102, "val_cos_ema": 0.7286305014009342}
4
+ {"step": 20000, "elapsed": 455.4616900680121, "val_loss": 0.02877548722760898, "val_cos": 0.730062623848371, "val_loss_ema": 0.02760639295715932, "val_cos_ema": 0.741453834112345, "ind_cos": 0.5139224870039631, "agg_cos": 0.5566421328941273, "ind_nll": 0.4076895362602296, "agg_nll": 0.38900612486489267, "retrieval_seconds": 11.382992881932296}
5
+ {"step": 25000, "elapsed": 566.4225651749875, "val_loss": 0.028477733112766957, "val_cos": 0.7375770131793967, "val_loss_ema": 0.027340128698009822, "val_cos_ema": 0.7472576683453215}
6
+ {"step": 30000, "elapsed": 665.1743251919979, "val_loss": 0.02754206203127111, "val_cos": 0.7456102060507167, "val_loss_ema": 0.026788575189456845, "val_cos_ema": 0.7542761992583573, "ind_cos": 0.5410983581244532, "agg_cos": 0.5792372227083226, "ind_nll": 0.4386226774409388, "agg_nll": 0.4280199701477173, "retrieval_seconds": 11.324746005935594}
7
+ {"step": 35000, "elapsed": 774.31287277292, "val_loss": 0.028099377737330473, "val_cos": 0.7470852280860762, "val_loss_ema": 0.027340684733969275, "val_cos_ema": 0.7524030052128905}
8
+ {"step": 40000, "elapsed": 868.8698491069954, "val_loss": 0.027492598784406123, "val_cos": 0.752357755823286, "val_loss_ema": 0.027221920008126545, "val_cos_ema": 0.7546660527191955, "ind_cos": 0.536774924082557, "agg_cos": 0.5725976632868393, "ind_nll": 0.44433578671058727, "agg_nll": 0.4258582531267693, "retrieval_seconds": 11.217270076973364}
9
+ {"step": 45000, "elapsed": 987.770604688907, "val_loss": 0.02711030717700807, "val_cos": 0.756390153732822, "val_loss_ema": 0.02730759745293561, "val_cos_ema": 0.7553835590081193}
10
+ {"step": 50000, "elapsed": 1090.247666540905, "val_loss": 0.02738600794755705, "val_cos": 0.754768325638382, "val_loss_ema": 0.027265339505378558, "val_cos_ema": 0.7557645033278219, "ind_cos": 0.5379072520459107, "agg_cos": 0.5739873385145916, "ind_nll": 0.4516444490195069, "agg_nll": 0.43522569355087753, "retrieval_seconds": 11.337262545945123}
shared/package-files/56/56382d924fa83294fbf0719351015cf3cd417b347da8c0754849c9ed485ed27d/stack_ceiling.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Ceiling diagnostic for a learned re-ranker over the released branches (lab, E65)
2
+
3
+ stack_ceiling.py --pred evals/fullfold/NAME/pred --rankings evals/fullfold/NAME/rankings.jsonl
4
+
5
+ DIAGNOSTIC ONLY - nothing fitted here may enter a package. A LightGBM ranker is fitted on the
6
+ branch scores of VALIDATION pools in a pool-grouped 5-fold cross-validation, to measure how much a
7
+ learned fusion could add at most over the hand-set released rule if its training features had the
8
+ validation distribution. It decides whether a train-fold stacker (cross-fitted forward members,
9
+ 10+ GPU hours) is worth building. The fitted models are discarded.
10
+ """
11
+ import argparse
12
+ import json
13
+ import pickle
14
+ import sys
15
+ from pathlib import Path
16
+
17
+ import lightgbm as lgb
18
+ import numpy as np
19
+
20
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
21
+ import scorelab
22
+ sys.path.insert(0, str(Path(__file__).resolve().parent.parent / 'pkg'))
23
+ import infer
24
+
25
+ ROOT = Path('/workspace/research')
26
+ BRANCHES = ('distances', 'inverse', 'library', 'ion')
27
+
28
+
29
+ def zscore(x):
30
+ out = np.full(len(x), np.nan)
31
+ ok = np.isfinite(x)
32
+ if ok.sum() > 1:
33
+ s = x[ok].std()
34
+ out[ok] = (x[ok] - x[ok].mean()) / (s if s > 0 else 1.0)
35
+ return out
36
+
37
+
38
+ def pct(x):
39
+ out = np.full(len(x), np.nan)
40
+ ok = np.isfinite(x)
41
+ if ok.sum() > 1:
42
+ out[ok] = np.argsort(np.argsort(x[ok])) / (ok.sum() - 1)
43
+ return out
44
+
45
+
46
+ def main():
47
+ p = argparse.ArgumentParser(description=__doc__)
48
+ p.add_argument('--pred', required=True)
49
+ p.add_argument('--rankings', required=True)
50
+ p.add_argument('--folds', type=int, default=5)
51
+ p.add_argument('--seed', type=int, default=0)
52
+ args = p.parse_args()
53
+ root = ROOT / 'validation-input'
54
+ queries = list(infer.read_jsonl(root / 'queries.jsonl'))
55
+ pools = {r['candidate_pool_id']: r['candidates'] for r in infer.read_jsonl(root / 'candidate_pools.jsonl')}
56
+ ev = scorelab.Evaluator(queries, pools)
57
+ branch = {name: pickle.load(open(ROOT / args.pred / f'{name}.pkl', 'rb')) for name in BRANCHES}
58
+ index = {q['query_id']: i for i, q in enumerate(queries)}
59
+ released = {}
60
+ for line in open(ROOT / args.rankings):
61
+ row = json.loads(line)
62
+ released[index[row['query_id']]] = row['ranking']
63
+ domain = ev.domain
64
+ X, y, group, pool_of, sizes, base_hit = [], [], [], [], [], []
65
+ names = None
66
+ for pid, members in ev.by_pool.items():
67
+ members = np.array(members)
68
+ n = len(pools[pid])
69
+ ids = ev.ids[pid]
70
+ feats, cols = [], []
71
+ for part, sel in (('dom', members[domain[members]]), ('oth', members[~domain[members]]), ('all', members)):
72
+ for name in BRANCHES:
73
+ vals = [branch[name].get(queries[i]['query_id']) for i in sel]
74
+ vals = [np.where(np.isfinite(v), v, np.nan) for v in vals if v is not None]
75
+ mean = np.nanmean(vals, axis=0) if vals else np.full(n, np.nan)
76
+ feats += [mean, zscore(mean), pct(mean)]
77
+ cols += [f'{part}_{name}', f'{part}_{name}_z', f'{part}_{name}_pct']
78
+ order = {cid: k for k, cid in enumerate(released[int(members[0])])}
79
+ rel = np.array([order[c] for c in ids], dtype=float)
80
+ feats += [rel, rel / max(n - 1, 1), np.full(n, len(members)), np.full(n, domain[members].mean()), np.full(n, n),
81
+ np.full(n, np.median([float(queries[i]['precursor_mz']) for i in members])),
82
+ np.full(n, np.mean([queries[i]['adduct'] == '[M+Na]+' for i in members]))]
83
+ cols += ['release_rank', 'release_pct', 'n_spectra', 'domain_share', 'n_cands', 'mz', 'na_share']
84
+ names = cols
85
+ X.append(np.stack(feats, axis=1))
86
+ label = np.zeros(n, dtype=int)
87
+ label[ev.target[members[0]]] = 1
88
+ y.append(label)
89
+ group.append(n)
90
+ sizes.append(len(members))
91
+ base_hit.append(float(rel[ev.target[members[0]]] == 0))
92
+ sizes = np.array(sizes, dtype=float)
93
+ base_hit = np.array(base_hit)
94
+ n_pools = len(X)
95
+ print(f'pools {n_pools}, rows {sum(group)}, features {len(names)}')
96
+ print(f'release: per query {np.sum(base_hit * sizes) / sizes.sum():.4f} pool-macro {base_hit.mean():.4f}')
97
+ rng = np.random.default_rng(args.seed)
98
+ fold = rng.permutation(n_pools) % args.folds
99
+ variants = {
100
+ 'branches only (no release rank)': [k for k, c in enumerate(names) if not c.startswith('release')],
101
+ 'branches + release rank': list(range(len(names))),
102
+ 'all-part means only (4 raw + z)': [k for k, c in enumerate(names) if c.startswith('all_') and not c.endswith('pct')] + [names.index(c) for c in ('n_spectra', 'domain_share', 'mz', 'na_share')],
103
+ }
104
+ boot = np.random.default_rng(1).integers(0, n_pools, size=(1000, n_pools))
105
+ for label, cols in variants.items():
106
+ hit = np.zeros(n_pools)
107
+ top5 = np.zeros(n_pools)
108
+ for f in range(args.folds):
109
+ tr = np.where(fold != f)[0]
110
+ te = np.where(fold == f)[0]
111
+ model = lgb.LGBMRanker(objective='lambdarank', n_estimators=300, learning_rate=0.05, num_leaves=15, min_child_samples=40,
112
+ subsample=0.8, subsample_freq=1, colsample_bytree=0.8, lambdarank_truncation_level=10, verbose=-1, n_jobs=6,
113
+ random_state=args.seed)
114
+ model.fit(np.concatenate([X[k][:, cols] for k in tr]), np.concatenate([y[k] for k in tr]), group=[group[k] for k in tr])
115
+ for k in te:
116
+ score = model.predict(X[k][:, cols])
117
+ order = np.argsort(-score, kind='stable')
118
+ pos = int(np.where(order == int(np.argmax(y[k])))[0][0])
119
+ hit[k] = pos == 0
120
+ top5[k] = pos < 5
121
+ d = hit - base_hit
122
+ per_query = np.sum(d[boot] * sizes[boot], axis=1) / np.sum(sizes[boot], axis=1)
123
+ macro = d[boot].mean(axis=1)
124
+ print(f'{label}: per query {np.sum(hit * sizes) / sizes.sum():.4f} ({np.sum(d * sizes) / sizes.sum():+.4f} [{np.percentile(per_query, 2.5):+.4f}, {np.percentile(per_query, 97.5):+.4f}])'
125
+ f' pool-macro {hit.mean():.4f} ({d.mean():+.4f} [{np.percentile(macro, 2.5):+.4f}, {np.percentile(macro, 97.5):+.4f}])'
126
+ f' top-5 per query {np.sum(top5 * sizes) / sizes.sum():.4f} pools +{int((d > 0).sum())} -{int((d < 0).sum())}')
127
+
128
+
129
+ if __name__ == '__main__':
130
+ main()
shared/package-files/56/565d1106b459cf127e3695d8351dab7b2c257a2e849af804956e688275602ac4/report-np075-val-sub-replay.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "benchmark_score": {
3
+ "at_or_above_upper_reference": true,
4
+ "comparable_to_test_reference": false,
5
+ "direction": "maximize",
6
+ "evaluation_fold": "validation",
7
+ "lower_reference": 0.04004329004329004,
8
+ "metric": "hit_rate_at_1",
9
+ "raw_performance": 0.5770843036795529,
10
+ "saturation": "upper",
11
+ "score": 1.0,
12
+ "unclipped_score": 1.1832695827558022,
13
+ "upper_reference": 0.4939052175894281,
14
+ "version": "massspecgym-retrieval-score/v1"
15
+ },
16
+ "external_reference_context": {
17
+ "comparable_to_this_validation_report": false,
18
+ "fold": "test",
19
+ "hit_rate_at_1": 0.4939052175894281,
20
+ "name": "GLACIER no-CF with strict-clean MIST fallback"
21
+ },
22
+ "global_error": null,
23
+ "metrics": {
24
+ "hit_rate_at_1": 0.5770843036795529,
25
+ "hit_rate_at_20": 0.8164881229622729,
26
+ "hit_rate_at_5": 0.7177456916627852
27
+ },
28
+ "queries": 2147,
29
+ "schema": "massspecgym-retrieval-development-report/v1",
30
+ "strata": {
31
+ "adduct": {
32
+ "[M+H]+": {
33
+ "hit_rate_at_1": 0.6628468033775633,
34
+ "queries": 1658
35
+ },
36
+ "[M+Na]+": {
37
+ "hit_rate_at_1": 0.28629856850715746,
38
+ "queries": 489
39
+ }
40
+ },
41
+ "candidate_pool_size": {
42
+ "129-256": {
43
+ "hit_rate_at_1": 0.5770843036795529,
44
+ "queries": 2147
45
+ }
46
+ },
47
+ "collision_energy": {
48
+ "20-39.999": {
49
+ "hit_rate_at_1": 0.796875,
50
+ "queries": 512
51
+ },
52
+ "40-59.999": {
53
+ "hit_rate_at_1": 0.7401129943502824,
54
+ "queries": 177
55
+ },
56
+ "<20": {
57
+ "hit_rate_at_1": 0.6291390728476821,
58
+ "queries": 151
59
+ },
60
+ ">=60": {
61
+ "hit_rate_at_1": 0.8132183908045977,
62
+ "queries": 348
63
+ },
64
+ "missing": {
65
+ "hit_rate_at_1": 0.3357664233576642,
66
+ "queries": 959
67
+ }
68
+ },
69
+ "instrument_type": {
70
+ "": {
71
+ "hit_rate_at_1": 0.6,
72
+ "queries": 35
73
+ },
74
+ "Orbitrap": {
75
+ "hit_rate_at_1": 0.5941176470588235,
76
+ "queries": 1870
77
+ },
78
+ "QTOF": {
79
+ "hit_rate_at_1": 0.44214876033057854,
80
+ "queries": 242
81
+ }
82
+ },
83
+ "routing_domain": {
84
+ "fallback": {
85
+ "hit_rate_at_1": 0.32896064581231077,
86
+ "queries": 991
87
+ },
88
+ "glacier_released": {
89
+ "hit_rate_at_1": 0.7897923875432526,
90
+ "queries": 1156
91
+ }
92
+ }
93
+ },
94
+ "uncertainty": {
95
+ "confidence": 0.95,
96
+ "estimate": 0.5770843036795529,
97
+ "grouping": "supplied target_inchikey connectivity block",
98
+ "lower": 0.476581171734675,
99
+ "molecule_groups": 413,
100
+ "queries": 2147,
101
+ "resamples": 2000,
102
+ "seed": 0,
103
+ "statistic": "spectrum-weighted mean; molecules sampled with replacement",
104
+ "upper": 0.6835037618988236
105
+ },
106
+ "valid_rankings": 2147
107
+ }
shared/package-files/56/56d8f62123d1d469f6fba33cb1d2528e433075d82052bd9ce09172670c665e98/matched-B3-30k.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {"event": "canonicalized", "unique_candidates": 828117, "unsupported": 7007, "seconds": 73.11523634800687}
2
+ {"event": "planned", "queries": 19429, "groups": 10176, "molecules": 944948, "simulations": 2540748, "seconds": 76.45933417799824}
3
+ ind meanD all=0.4820 dom=0.7160 non=0.2470 h5=0.6341 h20=0.7118 | diff=+0.0000 [+0.0000,+0.0000]
4
+ agg meanD all=0.6241 dom=0.7888 non=0.4588 h5=0.8174 h20=0.9036 | diff=+0.0000 [+0.0000,+0.0000]
5
+ ind Dmean all=0.4820 dom=0.7160 non=0.2470 h5=0.6341 h20=0.7118 | diff=+0.0000 [+0.0000,+0.0000]
6
+ agg Dmean all=0.6241 dom=0.7888 non=0.4588 h5=0.8174 h20=0.9036 | diff=+0.0000 [+0.0000,+0.0000]
7
+ ind gmean all=0.4820 dom=0.7160 non=0.2470 h5=0.6341 h20=0.7118 | diff=+0.0000 [+0.0000,+0.0000]
8
+ agg gmean all=0.6241 dom=0.7888 non=0.4588 h5=0.8174 h20=0.9036 | diff=+0.0000 [+0.0000,+0.0000]
shared/package-files/75/751915c72e387c626ac9a076693ea766f19ade5559e92addc37cec166f8c5f9d/train.py ADDED
@@ -0,0 +1,371 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Train native GLACIER from random initialization on all acquisition conditions.
3
+
4
+ Plain PyTorch loop around the pinned ``JointModel`` (same losses, optimizer,
5
+ LR and MAGMa schedules as the native recipe). Differences from the starter
6
+ wrapper: native batch size 32 without gradient accumulation, several loader
7
+ workers, every adduct/instrument/collision-energy condition, optional EMA
8
+ weights and optional warm start from one of OUR OWN train-fold checkpoints.
9
+ No test input is read. Validation rows are used for a monitoring loss only.
10
+ """
11
+
12
+ import argparse
13
+ import functools
14
+ import hashlib
15
+ import json
16
+ import math
17
+ import os
18
+ from pathlib import Path
19
+ import random
20
+ import shutil
21
+ import sys
22
+ import time
23
+
24
+
25
+ def sha256(path):
26
+ h = hashlib.sha256()
27
+ with Path(path).open('rb') as f:
28
+ for block in iter(lambda: f.read(1024 * 1024), b''):
29
+ h.update(block)
30
+ return h.hexdigest()
31
+
32
+
33
+ def seed_worker(base, worker):
34
+ import numpy as np
35
+ random.seed(base + worker)
36
+ np.random.seed((base + worker) % 2**32)
37
+
38
+
39
+ def main():
40
+ p = argparse.ArgumentParser(description=__doc__)
41
+ p.add_argument('--data-root', type=Path, required=True)
42
+ p.add_argument('--source-root', type=Path, default=Path('/workspace/research/ms-pred'))
43
+ p.add_argument('--config', type=Path, default=Path(__file__).with_name('config.json'))
44
+ p.add_argument('--work-dir', type=Path, required=True)
45
+ p.add_argument('--subset', choices=['all', 'released', 'other'], default='all')
46
+ p.add_argument('--batch-size', type=int, default=32)
47
+ p.add_argument('--workers', type=int, default=5)
48
+ p.add_argument('--max-steps', type=int, required=True)
49
+ p.add_argument('--max-seconds', type=float, default=1e9)
50
+ p.add_argument('--checkpoint-every', type=int, default=5000)
51
+ p.add_argument('--resume-every', type=int, default=1000)
52
+ p.add_argument('--validate-every', type=int, default=5000)
53
+ p.add_argument('--validation-rows', type=int, default=2048)
54
+ p.add_argument('--ema-decay', type=float, default=0.999)
55
+ p.add_argument('--seed', type=int, default=1)
56
+ p.add_argument('--init', type=Path, help='warm start from our own exported train-fold checkpoint')
57
+ p.add_argument('--override', action='append', default=[], help='config override key=value (JSON value)')
58
+ p.add_argument('--limit-train', type=int)
59
+ p.add_argument('--fork', type=Path, help='resume state (last.ckpt copy) of another of our runs to branch from')
60
+ p.add_argument('--loss-precursor-window', type=float,
61
+ help='fragment-only intensity loss: ignore predicted and target peaks at or above precursor - window')
62
+ p.add_argument('--decoys', type=Path, help='decoy table (prepare_decoys.py): adds the contrastive ranking term')
63
+ p.add_argument('--num-decoys', type=int, default=3)
64
+ p.add_argument('--mutation-rate', type=float, default=0.5, help='share of decoys taken from mutations')
65
+ p.add_argument('--contr-weight', type=float, default=1.0)
66
+ p.add_argument('--contr-loss', choices=['sinkhorn', 'softmax'], default='sinkhorn')
67
+ p.add_argument('--contr-tau', type=float, default=0.05)
68
+ p.add_argument('--contr-threshold', type=float, default=0.5)
69
+ p.add_argument('--contr-window', type=float, default=0.5, help='fragment-only ranked distance; negative = full')
70
+ p.add_argument('--contr-power', type=float, default=0.75)
71
+ p.add_argument('--contr-topk', type=int, default=100)
72
+ p.add_argument('--synth', type=Path,
73
+ help='synthetic-augmentation store (prepare_synth.py): adds an end-to-end intensity loss on it')
74
+ p.add_argument('--synth-per-item', type=int, default=1, help='synthetic spectra per real spectrum per step')
75
+ p.add_argument('--synth-weight', type=float, default=1.0)
76
+ p.add_argument('--synth-schedule', choices=['e2e', 'constant'], default='e2e',
77
+ help="e2e: weight x (1 - magma_weight), the stock end-to-end intensity weight")
78
+ p.add_argument('--synth-instrument', choices=['flag', 'label'], default='flag',
79
+ help='flag: synthetic spectra use their own instrument index; label: their labelled instrument')
80
+ args = p.parse_args()
81
+ args.work_dir = args.work_dir.resolve()
82
+ args.work_dir.mkdir(parents=True, exist_ok=True)
83
+ source = args.source_root.resolve() / 'src'
84
+ sys.path.insert(0, str(source))
85
+ sys.path.insert(0, str(Path(__file__).resolve().parent))
86
+ sys.path.insert(0, str(Path(__file__).resolve().parent / 'pkg'))
87
+ import numpy as np
88
+ import torch
89
+ from torch.utils.data import DataLoader, Sampler
90
+ from ms_pred.glacier.joint_model import JointModel
91
+ from dataset_all import AllConditionDataset
92
+
93
+ config = json.loads(args.config.read_text())
94
+ for item in args.override:
95
+ key, value = item.split('=', 1)
96
+ config[key] = json.loads(value)
97
+ manifest_path = args.data_root / 'manifest.json'
98
+ source_digest = hashlib.sha256()
99
+ for path in sorted((source / 'ms_pred').rglob('*.py')):
100
+ source_digest.update(str(path.relative_to(source)).encode())
101
+ source_digest.update(path.read_bytes())
102
+ identity = {
103
+ 'initialization': 'random' if not args.init else f'own-checkpoint:{sha256(args.init)}',
104
+ 'data_manifest_sha256': sha256(manifest_path), 'subset': args.subset,
105
+ 'config': config, 'source_sha256': source_digest.hexdigest(), 'seed': args.seed,
106
+ 'batch_size': args.batch_size, 'ema_decay': args.ema_decay, 'limit_train': args.limit_train,
107
+ 'trainer': 'allcond-plain-loop-v1',
108
+ }
109
+ if args.loss_precursor_window is not None:
110
+ identity['loss_precursor_window'] = args.loss_precursor_window
111
+ if args.decoys:
112
+ identity['contrastive'] = {
113
+ 'decoys_sha256': sha256(args.decoys), 'num_decoys': args.num_decoys, 'mutation_rate': args.mutation_rate,
114
+ 'weight': args.contr_weight, 'loss': args.contr_loss, 'tau': args.contr_tau,
115
+ 'threshold': args.contr_threshold, 'window': args.contr_window, 'power': args.contr_power,
116
+ 'topk': args.contr_topk,
117
+ }
118
+ if args.synth:
119
+ if args.decoys:
120
+ raise ValueError('--synth and --decoys are exclusive')
121
+ identity['synthetic'] = {
122
+ 'manifest_sha256': sha256(args.synth / 'manifest.json'), 'per_item': args.synth_per_item,
123
+ 'weight': args.synth_weight, 'schedule': args.synth_schedule, 'instrument': args.synth_instrument,
124
+ }
125
+ if args.fork:
126
+ if args.init:
127
+ raise ValueError('--fork and --init are exclusive')
128
+ identity['initialization'] = f'own-checkpoint:{sha256(args.fork)}'
129
+ identity['fork'] = str(args.fork)
130
+ recipe = args.work_dir / 'recipe.json'
131
+ last_path = args.work_dir / 'last.ckpt'
132
+ if recipe.exists() and json.loads(recipe.read_text()) != identity:
133
+ raise ValueError('recipe changed; use a new work directory')
134
+ recipe.write_text(json.dumps(identity, indent=2) + '\n')
135
+ if not (args.work_dir / 'training-source').exists():
136
+ shutil.copytree(Path(__file__).parent, args.work_dir / 'training-source',
137
+ ignore=shutil.ignore_patterns('__pycache__', '*.pyc'))
138
+
139
+ random.seed(args.seed)
140
+ np.random.seed(args.seed)
141
+ torch.manual_seed(args.seed)
142
+ torch.cuda.manual_seed_all(args.seed)
143
+
144
+ train = AllConditionDataset(args.data_root, 'train', config, subset=args.subset, limit=args.limit_train)
145
+ validation = AllConditionDataset(args.data_root, 'validation', config, subset=args.subset,
146
+ limit=args.validation_rows)
147
+ model = JointModel(**dict(config, node_feats=train.processor.get_node_feats(),
148
+ edge_feats=train.processor.get_edge_feats()))
149
+ logged = {}
150
+ model.log = lambda name, value, **kwargs: logged.__setitem__(name, float(value))
151
+ if args.init:
152
+ saved = torch.load(args.init, map_location='cpu', weights_only=False)
153
+ if saved.get('task_training', {}).get('pretrained_assets', []) != []:
154
+ raise ValueError('warm start must come from our own random-initialized train-fold run')
155
+ model.load_state_dict(saved['state_dict'])
156
+ if args.loss_precursor_window is not None:
157
+ full_loss, precursor_window = model.inten_loss_fn, float(args.loss_precursor_window)
158
+
159
+ def fragment_loss(pred, targ, parent_mass=None, use_hun=False):
160
+ # compare fragment regions only; spectra without any fragment target keep the full comparison
161
+ limit = parent_mass.view(-1, 1).to(pred.dtype) - precursor_window
162
+ targ_low = targ[:, :, 0] < limit
163
+ full = ~((targ_low & (targ[:, :, 1] > 0)).any(dim=1, keepdim=True))
164
+ targ_inten = torch.where(targ_low | full, targ[:, :, 1], torch.zeros_like(targ[:, :, 1]))
165
+ pred_inten = torch.where((pred[:, :, 0] < limit) | full, pred[:, :, 1], torch.zeros_like(pred[:, :, 1]))
166
+ return full_loss(torch.stack([pred[:, :, 0], pred_inten], dim=-1),
167
+ torch.stack([targ[:, :, 0], targ_inten], dim=-1), parent_mass=parent_mass, use_hun=use_hun)
168
+
169
+ model.inten_loss_fn = fragment_loss
170
+ contrast, train_items, train_collate = None, train, train.collate_fn
171
+ if args.decoys:
172
+ from contrastive import ContrastiveDataset, ContrastiveStep, collate_contrastive
173
+ from fastfeat import expand_graphormer
174
+ train_items = ContrastiveDataset(train, args.data_root, args.decoys, num_decoys=args.num_decoys,
175
+ mutation_rate=args.mutation_rate, seed=args.seed)
176
+ train_collate = collate_contrastive
177
+ contrast = ContrastiveStep(model, weight=args.contr_weight, kind=args.contr_loss, tau=args.contr_tau,
178
+ threshold=args.contr_threshold,
179
+ window=args.contr_window if args.contr_window >= 0 else None,
180
+ power=args.contr_power, topk=args.contr_topk)
181
+ synth_step = None
182
+ if args.synth:
183
+ from synth import SynthMixDataset, SynthStep, collate_synth
184
+ from fastfeat import expand_graphormer
185
+ train_items = SynthMixDataset(train, args.synth, per_item=args.synth_per_item, seed=args.seed,
186
+ instrument=args.synth_instrument)
187
+ train_collate = collate_synth
188
+ synth_step = SynthStep(model, weight=args.synth_weight, schedule=args.synth_schedule)
189
+ device = torch.device('cuda')
190
+ model.to(device)
191
+ optim_config = model.configure_optimizers()
192
+ optimizer = optim_config['optimizer']
193
+ scheduler = optim_config['lr_scheduler']['scheduler']
194
+ ema = {k: v.detach().clone() for k, v in model.state_dict().items() if v.dtype.is_floating_point}
195
+
196
+ state = {'step': 0, 'epoch': 0, 'batch_in_epoch': 0, 'elapsed': 0.0, 'skipped': 0}
197
+ resume_path = last_path if last_path.exists() else args.fork
198
+ if resume_path is not None and resume_path.exists():
199
+ saved = torch.load(resume_path, map_location='cpu', weights_only=False)
200
+ if resume_path != last_path and saved['identity'].get('initialization') != 'random' \
201
+ and not str(saved['identity'].get('initialization')).startswith('own-checkpoint:'):
202
+ raise ValueError('fork source must be one of our own train-fold runs')
203
+ model.load_state_dict(saved['state_dict'])
204
+ optimizer.load_state_dict(saved['optimizer'])
205
+ scheduler.load_state_dict(saved['scheduler'])
206
+ ema = {k: v.to(device) for k, v in saved['ema_state_dict'].items()}
207
+ state = saved['loop_state']
208
+ model.step = saved['magma_step']
209
+ print(json.dumps({'event': 'resumed', 'source': str(resume_path), **state}), flush=True)
210
+
211
+ class EpochSampler(Sampler):
212
+ """Deterministic per-epoch permutation that can restart mid-epoch."""
213
+
214
+ def __init__(self, size, epoch, start):
215
+ self.size, self.epoch, self.start = size, epoch, start
216
+
217
+ def __iter__(self):
218
+ generator = torch.Generator().manual_seed(args.seed * 1000003 + self.epoch)
219
+ order = torch.randperm(self.size, generator=generator).tolist()
220
+ usable = (self.size // args.batch_size) * args.batch_size
221
+ return iter(order[self.start * args.batch_size:usable])
222
+
223
+ def __len__(self):
224
+ return (self.size // args.batch_size) * args.batch_size - self.start * args.batch_size
225
+
226
+ def loader_for(epoch, start):
227
+ if contrast is not None or synth_step is not None:
228
+ train_items.epoch = epoch
229
+ return DataLoader(train_items, batch_size=args.batch_size, sampler=EpochSampler(len(train), epoch, start),
230
+ num_workers=args.workers, collate_fn=train_collate, drop_last=True,
231
+ multiprocessing_context='spawn' if args.workers else None,
232
+ prefetch_factor=4 if args.workers else None,
233
+ worker_init_fn=functools.partial(seed_worker, args.seed * 7919 + epoch * 131))
234
+
235
+ def to_device(batch):
236
+ out = {}
237
+ for key, value in batch.items():
238
+ if torch.is_tensor(value):
239
+ out[key] = value.to(device, non_blocking=True)
240
+ elif isinstance(value, dict):
241
+ out[key] = to_device(value)
242
+ else:
243
+ out[key] = value
244
+ return out
245
+
246
+ validation_batches = None
247
+
248
+ def validate(weights=None):
249
+ nonlocal validation_batches
250
+ if validation_batches is None:
251
+ validation_batches = [validation.collate_fn([validation[i] for i in range(s, min(s + 64, len(validation)))])
252
+ for s in range(0, len(validation), 64)]
253
+ backup = None
254
+ if weights is not None:
255
+ backup = {k: v.detach().clone() for k, v in model.state_dict().items() if k in weights}
256
+ model.load_state_dict(weights, strict=False)
257
+ model.eval()
258
+ total, count = 0.0, 0
259
+ with torch.no_grad():
260
+ for batch in validation_batches:
261
+ size = len(batch['names'])
262
+ loss = model.validation_step(to_device(batch), 0)['loss']
263
+ total += float(loss) * size
264
+ count += size
265
+ model.train()
266
+ if backup is not None:
267
+ model.load_state_dict(backup, strict=False)
268
+ return total / max(count, 1)
269
+
270
+ def export(path, weights, kind):
271
+ state_dict = dict(model.state_dict())
272
+ state_dict.update(weights)
273
+ payload = {
274
+ 'state_dict': {k: v.detach().cpu() for k, v in state_dict.items()},
275
+ 'hyper_parameters': dict(model.hparams), 'global_step': state['step'], 'weights': kind,
276
+ 'task_training': dict(identity, elapsed_seconds=state['elapsed'], pretrained_assets=[],
277
+ data='MassSpecGym v1.5 train fold'),
278
+ }
279
+ temporary = Path(str(path) + '.tmp')
280
+ torch.save(payload, temporary)
281
+ os.replace(temporary, path)
282
+
283
+ def save_resume():
284
+ payload = {
285
+ 'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict(),
286
+ 'scheduler': scheduler.state_dict(), 'ema_state_dict': ema, 'loop_state': state,
287
+ 'magma_step': model.step, 'identity': identity,
288
+ }
289
+ temporary = Path(str(last_path) + '.tmp')
290
+ torch.save(payload, temporary)
291
+ os.replace(temporary, last_path)
292
+
293
+ checkpoints = args.work_dir / 'checkpoints'
294
+ checkpoints.mkdir(exist_ok=True)
295
+ print(json.dumps({'parameters': sum(p.numel() for p in model.parameters()), 'train_rows': len(train),
296
+ 'validation_rows': len(validation), 'steps_per_epoch': len(train) // args.batch_size,
297
+ 'identity': identity['initialization'], 'gpu': torch.cuda.get_device_name()}), flush=True)
298
+ model.train()
299
+ started = time.monotonic()
300
+ base_elapsed = state['elapsed']
301
+ window = []
302
+ done = state['step'] >= args.max_steps
303
+ while not done:
304
+ loader = loader_for(state['epoch'], state['batch_in_epoch'])
305
+ for batch in loader:
306
+ batch = to_device(batch)
307
+ if contrast is not None:
308
+ if 'decoy' in batch:
309
+ batch['decoy']['graphormer_input'] = expand_graphormer(batch['decoy'].pop('graphormer_compact'))
310
+ loss = contrast(batch)
311
+ elif synth_step is not None:
312
+ if 'synth' in batch:
313
+ batch['synth']['graphormer_input'] = expand_graphormer(batch['synth'].pop('graphormer_compact'))
314
+ loss = synth_step(batch, state['step'])
315
+ else:
316
+ loss = model.training_step(batch, state['step'])['loss']
317
+ if not torch.isfinite(loss):
318
+ state['skipped'] += 1
319
+ optimizer.zero_grad(set_to_none=True)
320
+ else:
321
+ optimizer.zero_grad(set_to_none=True)
322
+ loss.backward()
323
+ torch.nn.utils.clip_grad_value_(model.parameters(), 5)
324
+ optimizer.step()
325
+ window.append(float(loss))
326
+ scheduler.step()
327
+ state['step'] += 1
328
+ state['batch_in_epoch'] += 1
329
+ with torch.no_grad():
330
+ decay = min(args.ema_decay, (1 + state['step']) / (10 + state['step']))
331
+ for key, value in model.state_dict().items():
332
+ if key in ema:
333
+ ema[key].mul_(decay).add_(value.detach(), alpha=1 - decay)
334
+ state['elapsed'] = base_elapsed + time.monotonic() - started
335
+ step = state['step']
336
+ if step % 100 == 0:
337
+ record = {'step': step, 'epoch': state['epoch'], 'elapsed_seconds': state['elapsed'],
338
+ 'train_loss': float(np.mean(window)) if window else None,
339
+ 'lr': scheduler.get_last_lr()[0], 'magma_weight': model.magma_weight_scheduler(),
340
+ 'skipped': state['skipped'], 'metrics': dict(logged)}
341
+ window = []
342
+ with (args.work_dir / 'metrics.jsonl').open('a') as f:
343
+ f.write(json.dumps(record) + '\n')
344
+ if step % args.validate_every == 0 or step == args.max_steps:
345
+ record = {'step': step, 'elapsed_seconds': state['elapsed'], 'validation_loss': validate(),
346
+ 'validation_loss_ema': validate(ema)}
347
+ print(json.dumps(record), flush=True)
348
+ with (args.work_dir / 'validation.jsonl').open('a') as f:
349
+ f.write(json.dumps(record) + '\n')
350
+ if step % args.checkpoint_every == 0 or step == args.max_steps:
351
+ export(checkpoints / f'step-{step:07d}.ckpt', {}, 'raw')
352
+ export(checkpoints / f'step-{step:07d}-ema.ckpt', ema, 'ema')
353
+ if step % args.resume_every == 0 or step == args.max_steps:
354
+ save_resume()
355
+ if step >= args.max_steps or state['elapsed'] >= args.max_seconds:
356
+ done = True
357
+ break
358
+ else:
359
+ state['epoch'] += 1
360
+ state['batch_in_epoch'] = 0
361
+ del loader
362
+ save_resume()
363
+ (args.work_dir / 'result.json').write_text(json.dumps({
364
+ 'global_step': state['step'], 'elapsed_seconds': state['elapsed'], 'skipped_batches': state['skipped'],
365
+ 'hardware': torch.cuda.get_device_name(),
366
+ 'peak_cuda_allocated_bytes': torch.cuda.max_memory_allocated(),
367
+ }, indent=2) + '\n')
368
+
369
+
370
+ if __name__ == '__main__':
371
+ main()
shared/package-files/75/75a4f93f95b85ff5388ef21c1eeaa788e8ece72ef666c843530028f7c82eabad/e73-store-real-validation.log ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "allcond-inverse-real/v1",
3
+ "fold": "validation",
4
+ "spectra": 19429,
5
+ "structures": 3386,
6
+ "skipped": {
7
+ "structure": 0,
8
+ "adduct": 0
9
+ },
10
+ "ppm": 20.0,
11
+ "floor": 0.004,
12
+ "max_peaks": 48,
13
+ "mean_selected_peaks": 25.164084615780535,
14
+ "mean_annotated_peaks": 18.12532811776211,
15
+ "no_annotated_peak": 20,
16
+ "mean_explained_intensity": 0.8726378016248633,
17
+ "source_sha256": "26788b56a3cb9d19c265c925504a4ce4e77e4c6b65b2a11068be6f24d15bf70f",
18
+ "sha256": "4c0dec880ce9abc3419fabbb4b8d501fa4858c972c748fe6c03ade121918efc9",
19
+ "elapsed_seconds": 1.9333453039871529
20
+ }
shared/package-files/7c/7cf300e8d128316f3531a381d6e0cf81481efe9217b1c01a28a078b8628e2611/eval-score-N5-recur-mop-10.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 5 members, 151.7 s
2
+ cos pool rule all=0.6373 dom=0.8503 non=0.4235 h5=0.7486 macro=0.7348
3
+ cen pool rule all=0.6204 dom=0.8281 non=0.4119 h5=0.7458 macro=0.7103
4
+ cos_pm pool rule all=0.6387 dom=0.8478 non=0.4288 h5=0.7492 macro=0.7333
5
+ dot pool rule all=0.6437 dom=0.8113 non=0.4754 h5=0.7768 macro=0.7194
6
+ geo pool rule all=0.6299 dom=0.8336 non=0.4255 h5=0.7402 macro=0.7212
7
+ idf05 pool rule all=0.6400 dom=0.8470 non=0.4321 h5=0.7617 macro=0.7327
8
+ idf1 pool rule all=0.6367 dom=0.8481 non=0.4245 h5=0.7638 macro=0.7327
9
+ lin pool rule all=0.6058 dom=0.8006 non=0.4103 h5=0.7239 macro=0.6929
10
+ log01 pool rule all=0.6548 dom=0.8419 non=0.4670 h5=0.7800 macro=0.7277
11
+ log05 pool rule all=0.6716 dom=0.8547 non=0.4878 h5=0.7903 macro=0.7383
12
+ log20 pool rule all=0.6398 dom=0.8566 non=0.4222 h5=0.7573 macro=0.7404
13
+ nll02 pool rule all=0.5049 dom=0.7335 non=0.2754 h5=0.6406 macro=0.6063
14
+ nll10 pool rule all=0.4476 dom=0.6822 non=0.2121 h5=0.6114 macro=0.5540
15
+ poolcen pool rule all=0.6179 dom=0.8197 non=0.4153 h5=0.7570 macro=0.7123
16
+ pow020 pool rule all=0.6036 dom=0.8127 non=0.3936 h5=0.7745 macro=0.7050
17
+ pow030 pool rule all=0.6170 dom=0.8254 non=0.4077 h5=0.7702 macro=0.7126
18
+ pow040 pool rule all=0.6586 dom=0.8395 non=0.4768 h5=0.7800 macro=0.7274
19
+ pow05 pool rule all=0.6717 dom=0.8505 non=0.4921 h5=0.7924 macro=0.7342
20
+ pow05idf025 pool rule all=0.6709 dom=0.8521 non=0.4889 h5=0.7959 macro=0.7372
21
+ pow05idf05 pool rule all=0.6698 dom=0.8523 non=0.4866 h5=0.7993 macro=0.7374
22
+ pow05m pool rule all=0.6706 dom=0.8517 non=0.4888 h5=0.7902 macro=0.7333
23
+ pow05s pool rule all=0.6715 dom=0.8497 non=0.4925 h5=0.7859 macro=0.7333
24
+ pow060 pool rule all=0.6434 dom=0.8542 non=0.4318 h5=0.7625 macro=0.7395
25
+ pow060idf025 pool rule all=0.6405 dom=0.8519 non=0.4284 h5=0.7699 macro=0.7357
26
+ pow060idf050 pool rule all=0.6726 dom=0.8543 non=0.4900 h5=0.7954 macro=0.7398
27
+ pow060m pool rule all=0.6382 dom=0.8547 non=0.4207 h5=0.7584 macro=0.7374
28
+ pow065 pool rule all=0.6440 dom=0.8551 non=0.4320 h5=0.7613 macro=0.7389
29
+ pow070 pool rule all=0.6434 dom=0.8543 non=0.4316 h5=0.7578 macro=0.7389
30
+ pow070idf025 pool rule all=0.6395 dom=0.8500 non=0.4282 h5=0.7589 macro=0.7377
31
+ pow070idf050 pool rule all=0.6440 dom=0.8511 non=0.4361 h5=0.7659 macro=0.7380
32
+ pow070m pool rule all=0.6415 dom=0.8559 non=0.4262 h5=0.7554 macro=0.7389
33
+ pow075 pool rule all=0.6389 dom=0.8488 non=0.4282 h5=0.7542 macro=0.7369
34
+ pow080 pool rule all=0.6361 dom=0.8476 non=0.4237 h5=0.7515 macro=0.7345
35
+ pow080m pool rule all=0.6381 dom=0.8517 non=0.4236 h5=0.7499 macro=0.7369
36
+ pow085 pool rule all=0.6362 dom=0.8470 non=0.4244 h5=0.7491 macro=0.7342
37
+ pow085idf025 pool rule all=0.6385 dom=0.8492 non=0.4270 h5=0.7569 macro=0.7360
38
+ pow085idf050 pool rule all=0.6405 dom=0.8520 non=0.4282 h5=0.7626 macro=0.7363
39
+ pow090 pool rule all=0.6373 dom=0.8469 non=0.4268 h5=0.7490 macro=0.7342
40
+ pow2 pool rule all=0.6021 dom=0.8090 non=0.3944 h5=0.7278 macro=0.7047
41
+ tani pool rule all=0.6097 dom=0.8257 non=0.3928 h5=0.7489 macro=0.7147
shared/package-files/93/93355d7b360441f5ecd9b3cb24b5c389ef9a69d901c4f0f60c5cd3ca863ace0d/inv-pool-strata.log ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ pools | all (3386) | domain only (2153) | mixed (254) | other only (979) | other only < 850 (831) | other only >= 850 (148)
2
+ I16 | 0.7289 | 0.8291 | 0.7008 | 0.5158 | 0.5572 | 0.2838
3
+ I18 | 0.7295 | 0.8291 | 0.7283 | 0.5107 | 0.5535 | 0.2703
4
+ I19 | 0.7224 | 0.8258 | 0.7087 | 0.4985 | 0.5355 | 0.2905
5
+ I20 | 0.7215 | 0.8268 | 0.7087 | 0.4934 | 0.5355 | 0.2568
6
+ I21 | 0.7212 | 0.8240 | 0.6890 | 0.5036 | 0.5487 | 0.2500
7
+ I22 | 0.7197 | 0.8263 | 0.6732 | 0.4974 | 0.5367 | 0.2770
8
+ I25 | 0.7044 | 0.8012 | 0.6575 | 0.5036 | 0.5439 | 0.2770
9
+ I23 | 0.7259 | 0.8281 | 0.7362 | 0.4985 | 0.5391 | 0.2703
10
+ I24 | 0.7097 | 0.8170 | 0.6732 | 0.4831 | 0.5211 | 0.2703
11
+ N5 | 0.7357 | 0.8360 | 0.7087 | 0.5220 | 0.5704 | 0.2500
12
+ I14 | 0.7132 | 0.8216 | 0.6732 | 0.4852 | 0.5295 | 0.2365
shared/package-files/93/93356a643cbdd9fba86186125a89b55bda768140821198c6378c5aef321ac800/eval-score-I16+I17.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 2 members, 68.8 s
2
+ cos pool rule all=0.6003 dom=0.8409 non=0.3588 h5=0.7008 macro=0.7194
3
+ cen pool rule all=0.5749 dom=0.8124 non=0.3365 h5=0.6960 macro=0.6961
4
+ cos_pm pool rule all=0.5995 dom=0.8399 non=0.3581 h5=0.7047 macro=0.7180
5
+ dot pool rule all=0.5787 dom=0.8027 non=0.3539 h5=0.7140 macro=0.7050
6
+ geo pool rule all=0.5722 dom=0.8251 non=0.3182 h5=0.6903 macro=0.7050
7
+ idf05 pool rule all=0.6020 dom=0.8360 non=0.3671 h5=0.7125 macro=0.7203
8
+ idf1 pool rule all=0.6050 dom=0.8354 non=0.3736 h5=0.7205 macro=0.7215
9
+ lin pool rule all=0.5575 dom=0.7901 non=0.3240 h5=0.6812 macro=0.6766
10
+ log01 pool rule all=0.5913 dom=0.8268 non=0.3548 h5=0.7273 macro=0.7141
11
+ log05 pool rule all=0.6056 dom=0.8421 non=0.3682 h5=0.7197 macro=0.7245
12
+ log20 pool rule all=0.6019 dom=0.8445 non=0.3584 h5=0.7135 macro=0.7221
13
+ nll02 pool rule all=0.4479 dom=0.7168 non=0.1780 h5=0.5869 macro=0.5856
14
+ nll10 pool rule all=0.4169 dom=0.6652 non=0.1676 h5=0.5418 macro=0.5384
15
+ poolcen pool rule all=0.5776 dom=0.8078 non=0.3465 h5=0.7073 macro=0.6967
16
+ pow020 pool rule all=0.5669 dom=0.7998 non=0.3332 h5=0.7168 macro=0.6911
17
+ pow030 pool rule all=0.5834 dom=0.8167 non=0.3490 h5=0.7230 macro=0.7011
18
+ pow040 pool rule all=0.5923 dom=0.8246 non=0.3589 h5=0.7282 macro=0.7118
19
+ pow05 pool rule all=0.6004 dom=0.8320 non=0.3678 h5=0.7244 macro=0.7188
20
+ pow05idf025 pool rule all=0.6019 dom=0.8385 non=0.3643 h5=0.7246 macro=0.7224
21
+ pow05idf05 pool rule all=0.6011 dom=0.8388 non=0.3625 h5=0.7329 macro=0.7236
22
+ pow05m pool rule all=0.6008 dom=0.8340 non=0.3667 h5=0.7240 macro=0.7194
23
+ pow05s pool rule all=0.6031 dom=0.8411 non=0.3641 h5=0.7154 macro=0.7191
24
+ pow060 pool rule all=0.6045 dom=0.8413 non=0.3668 h5=0.7181 macro=0.7242
25
+ pow060idf025 pool rule all=0.6087 dom=0.8508 non=0.3657 h5=0.7215 macro=0.7271
26
+ pow060idf050 pool rule all=0.6075 dom=0.8496 non=0.3645 h5=0.7295 macro=0.7283
27
+ pow060m pool rule all=0.6038 dom=0.8421 non=0.3645 h5=0.7173 macro=0.7253
28
+ pow065 pool rule all=0.6053 dom=0.8426 non=0.3670 h5=0.7158 macro=0.7224
29
+ pow070 pool rule all=0.6028 dom=0.8422 non=0.3625 h5=0.7127 macro=0.7209
30
+ pow070idf025 pool rule all=0.6071 dom=0.8489 non=0.3644 h5=0.7176 macro=0.7245
31
+ pow070idf050 pool rule all=0.6043 dom=0.8472 non=0.3604 h5=0.7234 macro=0.7239
32
+ pow070m pool rule all=0.6026 dom=0.8424 non=0.3617 h5=0.7142 macro=0.7215
33
+ pow075 pool rule all=0.6036 dom=0.8432 non=0.3630 h5=0.7147 macro=0.7230
34
+ pow080 pool rule all=0.6024 dom=0.8414 non=0.3625 h5=0.7109 macro=0.7218
35
+ pow080m pool rule all=0.5993 dom=0.8410 non=0.3566 h5=0.7102 macro=0.7206
36
+ pow085 pool rule all=0.6009 dom=0.8427 non=0.3581 h5=0.7078 macro=0.7206
37
+ pow085idf025 pool rule all=0.6049 dom=0.8421 non=0.3667 h5=0.7109 macro=0.7209
38
+ pow085idf050 pool rule all=0.6028 dom=0.8394 non=0.3652 h5=0.7188 macro=0.7212
39
+ pow090 pool rule all=0.6004 dom=0.8418 non=0.3580 h5=0.7066 macro=0.7191
40
+ pow2 pool rule all=0.5689 dom=0.8066 non=0.3304 h5=0.6896 macro=0.6920
41
+ tani pool rule all=0.5794 dom=0.8150 non=0.3430 h5=0.7057 macro=0.7017
shared/package-files/93/934b804bda1e0387c4be73816a4bfcef9dda4e1f50a3af300746b49a84e4c801/dataset_all.py ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """All-condition MAGMa supervision for native GLACIER featurization.
2
+
3
+ Rows come from ``prepare_all.py``. Unlike the starter dataset, a missing
4
+ instrument maps to ms-pred's native ``Unknown`` class and a missing collision
5
+ energy is passed as NaN (ms-pred's native missing-energy embedding).
6
+ """
7
+
8
+ import json
9
+ import math
10
+ from pathlib import Path
11
+ import sqlite3
12
+
13
+ from torch.utils.data import Dataset
14
+ from ms_pred import common
15
+ from ms_pred.glacier.dataset import IntenDataset, TreeProcessor
16
+
17
+
18
+ def is_released_domain(row):
19
+ energy = row.get('collision_energy')
20
+ return (row['adduct'] == '[M+H]+' and isinstance(energy, (int, float))
21
+ and math.isfinite(energy) and row['instrument_type'] in {'Orbitrap', 'QTOF'})
22
+
23
+
24
+ def instrument_index(name):
25
+ return common.instrument2onehot_pos.get(name or 'Unknown', common.instrument2onehot_pos['Unknown'])
26
+
27
+
28
+ class AllConditionDataset(Dataset):
29
+ def __init__(self, root, split, model_config, subset='all', limit=None):
30
+ if split not in {'train', 'validation'}:
31
+ raise ValueError('only train and validation are development inputs')
32
+ root = Path(root).resolve()
33
+ rows = []
34
+ with (root / f'{split}.jsonl').open() as stream:
35
+ for line in stream:
36
+ row = json.loads(line)
37
+ released = is_released_domain(row)
38
+ if subset == 'released' and not released:
39
+ continue
40
+ if subset == 'other' and released:
41
+ continue
42
+ # keep only light metadata: spectra live in the target payload
43
+ rows.append({'query_id': row['query_id'], 'adduct': row['adduct'],
44
+ 'instrument_type': row['instrument_type'],
45
+ 'precursor_mz': row['precursor_mz'], 'released': released})
46
+ self.rows = rows[:limit] if limit else rows
47
+ self.database = root / 'targets.sqlite'
48
+ self.connection = None
49
+ self.processor = TreeProcessor(
50
+ pe_embed_k=model_config['pe_embed_k'], root_encode='graphormer',
51
+ embed_elem_group=model_config['embed_elem_group'],
52
+ multi_hop_max_dist=model_config['multi_hop_max_dist'],
53
+ )
54
+
55
+ def __getstate__(self):
56
+ return dict(self.__dict__, connection=None)
57
+
58
+ def __len__(self):
59
+ return len(self.rows)
60
+
61
+ def __getitem__(self, index):
62
+ if self.connection is None:
63
+ self.connection = sqlite3.connect(self.database.as_uri() + '?mode=ro', uri=True)
64
+ row = self.rows[index]
65
+ result = self.connection.execute('SELECT payload FROM targets WHERE query_id=?', (row['query_id'],)).fetchone()
66
+ if result is None or result[0] is None:
67
+ raise ValueError('missing prepared supervision')
68
+ tree = json.loads(result[0])
69
+ if tree.get('collision_energy') in (None, 'nan'):
70
+ tree['collision_energy'] = float('nan')
71
+ item = self.processor.featurize_tree(tree)
72
+ item.update(name=row['query_id'], adduct=common.ion2onehot_pos[row['adduct']],
73
+ instrument=instrument_index(row['instrument_type']),
74
+ precursor=row['precursor_mz'])
75
+ return item
76
+
77
+ collate_fn = staticmethod(IntenDataset.collate_fn)
shared/package-files/93/9352c4e7d7af5238ccfb70c06597c1e117aa2489a28a318dccc5e459fe864cfc/replay-B20k+B30k-I1-50k-val-sub.log ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ reloc
2
+ reloc2
3
+ reloc3
4
+ reloc4
5
+ /opt/glacier/lib/python3.11/site-packages/torch/nn/modules/transformer.py:508: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active develo
6
+ output = torch._nested_tensor_from_mask(
7
+ {"event": "simulating", "finished": 64000, "total": 295422, "seconds": 402.3761534569785, "per_second": 159.0551513804925}
8
+ {"event": "simulating", "finished": 128000, "total": 295422, "seconds": 880.0155879409285, "per_second": 145.45196898102225}
9
+ {"event": "simulating", "finished": 192000, "total": 295422, "seconds": 1195.0014332679566, "per_second": 160.66926336225373}
10
+ {"event": "simulating", "finished": 256000, "total": 295422, "seconds": 1676.3089701608988, "per_second": 152.71647682910634}
11
+ {"event": "simulated", "finished": 295422, "total": 295422, "failed": 0, "seconds": 2013.4490474959603}
12
+ {"event": "finished", "queries": 2147, "unscored": 0, "directly_scored": 2147, "scoring_seconds": 25.9191494329134, "total_seconds": 2076.9211527400184}
13
+
14
+ real 34m38.615s
15
+ user 52m47.035s
16
+ sys 0m33.758s
17
+
18
+ [exited with code 0]
shared/package-files/93/93e3bd183a2c73f60c7d014cc0ba3f6c0e33353d079873b4416ef58a826a3aee/ce-jitter-fused-B30k.log ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 718 pools; 2436 queries; jittered members: B30k=evals/tta-B30k-x scales ['0.8', '1.25']
2
+ forward B20k alone, recorded energy 0.8444 (pool macro 0.8524)
3
+ forward B20k alone, jitter mean 0.8444 (pool macro 0.8524) | +0.0000 [+0.0000,+0.0000] pools gained 0 lost 0
4
+ forward B30k alone, recorded energy 0.8247 (pool macro 0.8454)
5
+ forward B30k alone, jitter mean 0.8333 (pool macro 0.8524) | +0.0086 [+0.0024,+0.0161] pools gained 10 lost 5
6
+ forward pair, recorded energy 0.8461 (pool macro 0.8579)
7
+ forward pair, jitter mean 0.8571 (pool macro 0.8649) | +0.0111 [+0.0013,+0.0243] pools gained 7 lost 2
8
+ fused (released rule), recorded energy 0.9183 (pool macro 0.9234)
9
+ fused (released rule), jitter mean 0.9158 (pool macro 0.9192) | -0.0025 [-0.0058,+0.0000] pools gained 0 lost 3
10
+ fused, domain weight 0.5, recorded energy 0.9122 (pool macro 0.9164) | -0.0062 [-0.0139,+0.0016] pools gained 3 lost 8
11
+ fused, domain weight 0.5, jitter mean 0.9109 (pool macro 0.9150) | -0.0074 [-0.0154,+0.0004] pools gained 3 lost 9
12
+ fused, domain weight 1.0, recorded energy 0.9130 (pool macro 0.9164) | -0.0053 [-0.0106,-0.0012] pools gained 0 lost 5
13
+ fused, domain weight 1.0, jitter mean 0.9126 (pool macro 0.9164) | -0.0057 [-0.0115,-0.0012] pools gained 1 lost 6
shared/package-files/a5/a535d77b8ebc5ee1cd2b998a0eae5cd64daea829ff52d893e59d012178cc519d/publish.py ADDED
@@ -0,0 +1,226 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Build, release and document the GLACIER all-condition retrieval package.
3
+
4
+ ``build`` assembles one complete, standalone package directory (used for the
5
+ relocated validation replay).
6
+ ``release`` copies a validated build byte-for-byte into a new immutable
7
+ generation of the submission and switches the manifest atomically.
8
+ ``records`` copies EXPERIMENTS.md and research records (file-by-file, each
9
+ file replaced atomically) into the submission root.
10
+ """
11
+
12
+ import argparse
13
+ import hashlib
14
+ import json
15
+ import os
16
+ from pathlib import Path
17
+ import shutil
18
+ import sys
19
+
20
+ HERE = Path(__file__).resolve().parent
21
+ RESEARCH = HERE.parent
22
+
23
+
24
+ def sha256(path):
25
+ h = hashlib.sha256()
26
+ with Path(path).open('rb') as f:
27
+ for block in iter(lambda: f.read(1024 * 1024), b''):
28
+ h.update(block)
29
+ return h.hexdigest()
30
+
31
+
32
+ INFER_SH = '''#!/usr/bin/env bash
33
+ set -euo pipefail
34
+ package_root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
35
+ scratch="$(mktemp -d)"
36
+ trap 'rm -rf "$scratch"' EXIT
37
+ export PYTHONDONTWRITEBYTECODE=1 DGLBACKEND=pytorch MPLCONFIGDIR="$scratch/mpl" NUMBA_CACHE_DIR="$scratch/numba"
38
+ if [ ! -w "${HOME:-/nonexistent}" ]; then export HOME="$scratch"; fi
39
+ "${MASSSPECGYM_GLACIER_PYTHON:-/opt/glacier/bin/python}" "$package_root/infer.py" "$@"
40
+ '''
41
+
42
+
43
+ ROOT_INFER_SH = '''#!/usr/bin/env bash
44
+ set -euo pipefail
45
+ root="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
46
+ for candidate in python3 "${MASSSPECGYM_GLACIER_PYTHON:-/opt/glacier/bin/python}" /opt/conda/bin/python3; do
47
+ if command -v "$candidate" >/dev/null 2>&1; then exec "$candidate" "$root/dispatch.py" "$@"; fi
48
+ done
49
+ echo "infer.sh: no Python interpreter found" >&2
50
+ exit 1
51
+ '''
52
+
53
+
54
+ def build(args):
55
+ import torch
56
+ out = args.output.resolve()
57
+ if out.exists():
58
+ raise FileExistsError(out)
59
+ out.mkdir(parents=True)
60
+ members, provenance = [], []
61
+ for path in args.checkpoint:
62
+ saved = torch.load(path, map_location='cpu', weights_only=False)
63
+ training = saved['task_training']
64
+ if training.get('pretrained_assets') != [] or not str(training.get('initialization', '')).startswith(('random', 'own-checkpoint:')):
65
+ raise ValueError(f'{path}: not a train-fold, scratch-initialized checkpoint')
66
+ digest = sha256(path)
67
+ members.append({'state_dict': saved['state_dict'], 'hyper_parameters': saved['hyper_parameters'],
68
+ 'global_step': saved['global_step'], 'weights': saved['weights'],
69
+ 'task_training': training, 'source_checkpoint_sha256': digest})
70
+ provenance.append({'source': str(path), 'sha256': digest, 'global_step': saved['global_step'],
71
+ 'weights': saved['weights'], 'subset': training.get('subset'), 'seed': training.get('seed'),
72
+ 'elapsed_seconds': training.get('elapsed_seconds'),
73
+ 'data_manifest_sha256': training.get('data_manifest_sha256')})
74
+ bundle = {'format': 'glacier-members/v1', 'members': members}
75
+ inverse_provenance = []
76
+ for path in args.inverse or []:
77
+ saved = torch.load(path, map_location='cpu', weights_only=True)
78
+ digest = sha256(path)
79
+ recipe = json.loads((path.parent / 'recipe.json').read_text())
80
+ bundle.setdefault('inverse', []).append({'ema': saved['ema'], 'config': saved['config'],
81
+ 'class': saved.get('class', 'FormulaFormer'), 'step': saved['step'],
82
+ 'source_checkpoint_sha256': digest})
83
+ inverse_provenance.append({'source': str(path), 'sha256': digest, 'step': saved['step'],
84
+ 'class': saved.get('class', 'FormulaFormer'), 'weights': 'ema',
85
+ 'initialization': 'random', 'seed': recipe['args']['seed'],
86
+ 'training_data': 'MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus',
87
+ 'stores': recipe.get('stores'), 'recipe': recipe['args']})
88
+ torch.save(bundle, out / 'model.pt')
89
+ config = json.loads((HERE / 'pkg' / 'model_config.json').read_text())
90
+ for item in args.set:
91
+ key, value = item.split('=', 1)
92
+ config[key] = json.loads(value)
93
+ if inverse_provenance:
94
+ if not isinstance(config.get('inverse'), dict):
95
+ raise ValueError('--inverse needs --set inverse={...} (fusion options)')
96
+ config['inverse_members'] = inverse_provenance
97
+ elif config.get('inverse'):
98
+ raise ValueError('fusion options without --inverse checkpoints')
99
+ config['checkpoint'] = 'model.pt'
100
+ config['method'] = 'native-glacier-all-conditions-forward-simulation' + ('+formula-transformer-inverse-fusion' if inverse_provenance else '')
101
+ config['inference_runtime'] = '/opt/glacier/bin/python'
102
+ config['members'] = provenance
103
+ (out / 'model_config.json').write_text(json.dumps(config, indent=2) + '\n')
104
+ manifest = {'schema': 'massspecgym-retrieval-model/v1',
105
+ 'checkpoint': {'path': 'model.pt', 'sha256': sha256(out / 'model.pt'), 'format': 'pytorch'},
106
+ 'model_config': 'model_config.json',
107
+ 'training_provenance': {'data': 'MassSpecGym v1.5 train fold', 'pretrained_assets': []}}
108
+ (out / 'model_manifest.json').write_text(json.dumps(manifest, indent=2) + '\n')
109
+ for name in ['infer.py', 'fastfeat.py', 'inverse.py', 'invrun.py']:
110
+ shutil.copyfile(HERE / 'pkg' / name, out / name)
111
+ (out / 'infer.sh').write_text(INFER_SH)
112
+ (out / 'infer.sh').chmod(0o755)
113
+ source_root = args.source_root.resolve()
114
+ shutil.copytree(source_root / 'src' / 'ms_pred', out / 'vendor' / 'ms_pred',
115
+ ignore=shutil.ignore_patterns('__pycache__', '*.pyc'))
116
+ shutil.copyfile(source_root / 'LICENSE', out / 'LICENSE.ms-pred')
117
+ shutil.copytree(HERE, out / 'training', ignore=shutil.ignore_patterns('__pycache__', '*.pyc'))
118
+ readme = args.readme.read_text() if args.readme else '# GLACIER all-condition retrieval package\n'
119
+ (out / 'README.md').write_text(readme)
120
+ for path in [out, *out.rglob('*')]:
121
+ path.chmod(0o755 if path.is_dir() or path.stat().st_mode & 0o111 else 0o644)
122
+ print(json.dumps({'built': str(out), 'checkpoint_sha256': manifest['checkpoint']['sha256'],
123
+ 'members': provenance, 'config': {k: v for k, v in config.items() if k != 'members'}}, indent=2))
124
+
125
+
126
+ def release(args):
127
+ sys.path.insert(0, str(RESEARCH / 'starter'))
128
+ from publication import publish_directory
129
+ source = args.source.resolve()
130
+ manifest = json.loads((source / 'model_manifest.json').read_text())
131
+ if sha256(source / manifest['checkpoint']['path']) != manifest['checkpoint']['sha256']:
132
+ raise ValueError('staged checkpoint hash mismatch')
133
+ if args.readme:
134
+ readme = args.readme.read_text()
135
+
136
+ def copy(stage):
137
+ shutil.copytree(source, stage, dirs_exist_ok=True)
138
+ if args.readme:
139
+ (stage / 'README.md').write_text(readme)
140
+
141
+ destination = args.output_dir.resolve()
142
+ publish_directory(destination, copy)
143
+ # The stock root launcher needs `python3` on PATH; fall back to the documented runtimes.
144
+ temporary = destination / '.infer.sh.tmp'
145
+ temporary.write_text(ROOT_INFER_SH)
146
+ temporary.chmod(0o755)
147
+ os.replace(temporary, destination / 'infer.sh')
148
+ current = json.loads((destination / 'model_manifest.json').read_text())
149
+ selected = Path(current['model_config']).parts[1]
150
+ # Keep the selected generation and its predecessor; older ones are unreferenced.
151
+ generations = sorted((p for p in (destination / '.versions').iterdir() if p.is_dir()),
152
+ key=lambda p: p.stat().st_mtime, reverse=True)
153
+ keep = {selected}
154
+ for path in generations:
155
+ if path.name != selected and len(keep) < 2:
156
+ keep.add(path.name)
157
+ for path in generations:
158
+ if path.name not in keep:
159
+ shutil.rmtree(path, ignore_errors=True)
160
+ print(json.dumps({'released': selected, 'checkpoint_sha256': current['checkpoint']['sha256'], 'kept': sorted(keep)}))
161
+
162
+
163
+ def records(args):
164
+ destination = args.output_dir.resolve()
165
+
166
+ def put(source, relative):
167
+ target = destination / relative
168
+ target.parent.mkdir(parents=True, exist_ok=True)
169
+ target.parent.chmod(0o755)
170
+ if target.exists() and target.stat().st_size == source.stat().st_size and target.stat().st_mtime >= source.stat().st_mtime:
171
+ return
172
+ temporary = target.with_name(target.name + '.tmp')
173
+ shutil.copyfile(source, temporary)
174
+ temporary.chmod(0o644)
175
+ os.replace(temporary, target)
176
+
177
+ put(RESEARCH / 'EXPERIMENTS.md', 'EXPERIMENTS.md')
178
+ skip = {'__pycache__'}
179
+ for path in sorted(HERE.rglob('*')):
180
+ if path.is_file() and not (set(path.parts) & skip) and path.suffix != '.pyc':
181
+ put(path, Path('research') / 'source' / path.relative_to(HERE))
182
+ for run in sorted((RESEARCH / 'runs').iterdir()):
183
+ for name in ['recipe.json', 'metrics.jsonl', 'validation.jsonl', 'result.json']:
184
+ if (run / name).is_file():
185
+ put(run / name, Path('research') / 'runs' / run.name / name)
186
+ if (run / 'training-source').is_dir():
187
+ for path in sorted((run / 'training-source').rglob('*')):
188
+ if path.is_file() and not (set(path.parts) & skip):
189
+ put(path, Path('research') / 'runs' / run.name / 'training-source' / path.relative_to(run / 'training-source'))
190
+ for path in sorted((RESEARCH / 'logs').glob('*')):
191
+ if path.is_file() and path.stat().st_size < 50 * 1024 * 1024:
192
+ put(path, Path('research') / 'logs' / path.name)
193
+ for path in sorted((RESEARCH / 'evals').glob('*/report*.json')) + sorted((RESEARCH / 'evals').glob('*/*.log')) \
194
+ + sorted((RESEARCH / 'evals').glob('*/summary*.json')) + sorted((RESEARCH / 'evals').glob('*/summary*.txt')):
195
+ put(path, Path('research') / 'evals' / path.parent.name / path.name)
196
+ for name in ['manifest.json', 'preparation.json']:
197
+ if (RESEARCH / 'prep_all' / name).is_file():
198
+ put(RESEARCH / 'prep_all' / name, Path('research') / 'prep_all' / name)
199
+ for path in [destination / 'research', *(destination / 'research').rglob('*')]:
200
+ if path.is_dir():
201
+ path.chmod(0o755)
202
+ print(json.dumps({'records': str(destination / 'research')}))
203
+
204
+
205
+ def main():
206
+ parser = argparse.ArgumentParser(description=__doc__)
207
+ sub = parser.add_subparsers(dest='command', required=True)
208
+ b = sub.add_parser('build')
209
+ b.add_argument('--checkpoint', type=Path, action='append', required=True)
210
+ b.add_argument('--set', action='append', default=[])
211
+ b.add_argument('--inverse', type=Path, action='append', help='inverse-branch EMA checkpoint (train_inv.py / train_set.py)')
212
+ b.add_argument('--output', type=Path, required=True)
213
+ b.add_argument('--source-root', type=Path, default=RESEARCH / 'ms-pred')
214
+ b.add_argument('--readme', type=Path)
215
+ r = sub.add_parser('release')
216
+ r.add_argument('--source', type=Path, required=True)
217
+ r.add_argument('--output-dir', type=Path, default=Path('/workspace/submission'))
218
+ r.add_argument('--readme', type=Path)
219
+ c = sub.add_parser('records')
220
+ c.add_argument('--output-dir', type=Path, default=Path('/workspace/submission'))
221
+ args = parser.parse_args()
222
+ {'build': build, 'release': release, 'records': records}[args.command](args)
223
+
224
+
225
+ if __name__ == '__main__':
226
+ main()
shared/package-files/b6/b63740294886c81680e5a0189d2abc271cd9d7726978ea72eb962bad61eb5d32/model_config.json ADDED
@@ -0,0 +1,610 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "sparse_k": 100,
3
+ "inference_batch_size": 128,
4
+ "workers": 14,
5
+ "num_bins": 150000,
6
+ "upper_limit": 1500,
7
+ "max_observed_peaks": 50,
8
+ "observed_intensity_threshold": 0.001,
9
+ "ensemble": "prediction",
10
+ "pool_aggregation": 1.0,
11
+ "autocast": false,
12
+ "model": {
13
+ "pe_embed_k": 14,
14
+ "embed_elem_group": true,
15
+ "multi_hop_max_dist": 10
16
+ },
17
+ "checkpoint": "model.pt",
18
+ "views": [
19
+ {
20
+ "precursor_window": 0.5,
21
+ "pred_power": 0.75,
22
+ "match_ppm": 20,
23
+ "match_floor_mda": 4,
24
+ "match_by_instrument": {
25
+ "Orbitrap": [
26
+ 10,
27
+ 1
28
+ ]
29
+ }
30
+ }
31
+ ],
32
+ "inverse": {
33
+ "pooled": true,
34
+ "score": {
35
+ "members": "probability",
36
+ "power": 0.7
37
+ },
38
+ "weight_domain": 0.75,
39
+ "weight_other": 0.05,
40
+ "weight_other_single": 0.5,
41
+ "batch": 2048
42
+ },
43
+ "library": {
44
+ "kind": "sum",
45
+ "k": 20,
46
+ "top": 32,
47
+ "power": 6,
48
+ "debias": true,
49
+ "weight_domain": 0.0,
50
+ "weight_other": 0.4,
51
+ "weight_other_single": 0.0
52
+ },
53
+ "ion": {
54
+ "weight_domain": 0.05,
55
+ "weight_other": 0.01
56
+ },
57
+ "inverse_members": [
58
+ {
59
+ "source": "runs/I9b-ff512x6-real08-corpus2+train3/step-0050000-ema.pt",
60
+ "sha256": "f2741a9676a4743fa586d3effe8753d6eade9356fb2ceac8055ff50baf34195f",
61
+ "step": 50000,
62
+ "class": "FormulaFormer",
63
+ "weights": "ema",
64
+ "initialization": "random",
65
+ "seed": 1,
66
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
67
+ "self_simulated_store": {
68
+ "path": "inv/selfsim/corpus2+train3.npz",
69
+ "rows": 2031301,
70
+ "manifest": {
71
+ "schema": "allcond-inverse-selfsim/v1",
72
+ "merged": [
73
+ "inv/selfsim/corpus2.npz",
74
+ "inv/selfsim/train3.npz"
75
+ ],
76
+ "parts": [
77
+ {
78
+ "schema": "allcond-inverse-selfsim/v1",
79
+ "merged": [
80
+ "inv/selfsim/sub75.npz",
81
+ "inv/selfsim/sub25.npz"
82
+ ],
83
+ "parts": [
84
+ {
85
+ "name": "sub75",
86
+ "source": "synth",
87
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
88
+ "bundle": "builds/5Ms-I1+I3x2-L",
89
+ "members": "1,3",
90
+ "structures": 435076,
91
+ "tasks": 651127,
92
+ "failed": 0,
93
+ "rows": 651127,
94
+ "seconds": 635.7592854518443,
95
+ "schema": "allcond-inverse-selfsim/v1",
96
+ "floor": 0.02,
97
+ "spectra": 651083,
98
+ "without_peak": 44,
99
+ "mean_annotated_peaks": 15.61580167198345,
100
+ "mean_candidates": 20.62178653319552,
101
+ "ppm": 20.0,
102
+ "floor_da": 0.004,
103
+ "max_peaks": 24,
104
+ "annotate_seconds": 25.62758301384747
105
+ },
106
+ {
107
+ "name": "sub25",
108
+ "source": "synth",
109
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
110
+ "bundle": "builds/5Ms-I1+I3x2-L",
111
+ "members": "1,3",
112
+ "structures": 435076,
113
+ "tasks": 217012,
114
+ "failed": 0,
115
+ "rows": 217012,
116
+ "seconds": 227.74617605702952,
117
+ "schema": "allcond-inverse-selfsim/v1",
118
+ "floor": 0.02,
119
+ "spectra": 217000,
120
+ "without_peak": 12,
121
+ "mean_annotated_peaks": 15.62026728110599,
122
+ "mean_candidates": 20.61567563084069,
123
+ "ppm": 20.0,
124
+ "floor_da": 0.004,
125
+ "max_peaks": 24,
126
+ "annotate_seconds": 10.8742212112993
127
+ }
128
+ ],
129
+ "spectra": 868083,
130
+ "structures": 434944
131
+ },
132
+ {
133
+ "name": "train3",
134
+ "source": "synth",
135
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
136
+ "bundle": "builds/5Ms-I1+I3x2-L",
137
+ "members": "1,3",
138
+ "structures": 435076,
139
+ "tasks": 1163260,
140
+ "failed": 0,
141
+ "rows": 1163260,
142
+ "seconds": 1262.926542488858,
143
+ "schema": "allcond-inverse-selfsim/v1",
144
+ "floor": 0.02,
145
+ "spectra": 1163218,
146
+ "without_peak": 42,
147
+ "mean_annotated_peaks": 17.868643710809152,
148
+ "mean_candidates": 25.376552103570997,
149
+ "ppm": 20.0,
150
+ "floor_da": 0.004,
151
+ "max_peaks": 24,
152
+ "annotate_seconds": 43.40271363919601
153
+ }
154
+ ],
155
+ "spectra": 2031301,
156
+ "structures": 434950
157
+ }
158
+ },
159
+ "stores": {
160
+ "gen-x1.manifest.json": null,
161
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
162
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
163
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
164
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
165
+ },
166
+ "recipe": {
167
+ "store": "/workspace/research/inv",
168
+ "out": "runs/I9b-ff512x6-real08-corpus2+train3",
169
+ "pools": "pools-validation-input.npz",
170
+ "input": "/workspace/research/validation-input",
171
+ "steps": 50000,
172
+ "batch": 512,
173
+ "real_frac": 0.08,
174
+ "real_balance": 0.5,
175
+ "no_synth": false,
176
+ "target": "",
177
+ "heavy_boost": 1.0,
178
+ "synth_structures": 1.0,
179
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
180
+ "extra_frac": 0.5,
181
+ "heavy_lo": 450.0,
182
+ "heavy_hi": 850.0,
183
+ "lr": 0.0006,
184
+ "weight_decay": 0.0001,
185
+ "warmup": 2000,
186
+ "ema": 0.999,
187
+ "dim": 512,
188
+ "layers": 6,
189
+ "heads": 8,
190
+ "dropout": 0.1,
191
+ "peak_drop": 0.3,
192
+ "truncate": 0.3,
193
+ "min_keep": 4,
194
+ "jitter": 0.15,
195
+ "synth_energy_drop": 0.3,
196
+ "synth_instrument_drop": 0.3,
197
+ "validate_every": 5000,
198
+ "retrieve_every": 10000,
199
+ "retrieve_pools": 0,
200
+ "seed": 1,
201
+ "contrast_weight": 0.0,
202
+ "contrast_k": 31,
203
+ "contrast_tau": 0.05,
204
+ "init": null,
205
+ "max_steps_this_run": 0
206
+ }
207
+ },
208
+ {
209
+ "source": "runs/I13-ff512x6-real08-corpus2+train3-seed2/step-0050000-ema.pt",
210
+ "sha256": "bd0f9c965cf7e0101877c780d9f3592e6fc07b4b20cdfc35adb9f60e3aa295a7",
211
+ "step": 50000,
212
+ "class": "FormulaFormer",
213
+ "weights": "ema",
214
+ "initialization": "random",
215
+ "seed": 2,
216
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus + spectra of the corpus structures simulated by our own train-fold-trained GLACIER forward members (self-simulation, E60)",
217
+ "self_simulated_store": {
218
+ "path": "inv/selfsim/corpus2+train3.npz",
219
+ "rows": 2031301,
220
+ "manifest": {
221
+ "schema": "allcond-inverse-selfsim/v1",
222
+ "merged": [
223
+ "inv/selfsim/corpus2.npz",
224
+ "inv/selfsim/train3.npz"
225
+ ],
226
+ "parts": [
227
+ {
228
+ "schema": "allcond-inverse-selfsim/v1",
229
+ "merged": [
230
+ "inv/selfsim/sub75.npz",
231
+ "inv/selfsim/sub25.npz"
232
+ ],
233
+ "parts": [
234
+ {
235
+ "name": "sub75",
236
+ "source": "synth",
237
+ "conditions": "from-rows 2 hash-range 0.25,1 (534 conditions)",
238
+ "bundle": "builds/5Ms-I1+I3x2-L",
239
+ "members": "1,3",
240
+ "structures": 435076,
241
+ "tasks": 651127,
242
+ "failed": 0,
243
+ "rows": 651127,
244
+ "seconds": 635.7592854518443,
245
+ "schema": "allcond-inverse-selfsim/v1",
246
+ "floor": 0.02,
247
+ "spectra": 651083,
248
+ "without_peak": 44,
249
+ "mean_annotated_peaks": 15.61580167198345,
250
+ "mean_candidates": 20.62178653319552,
251
+ "ppm": 20.0,
252
+ "floor_da": 0.004,
253
+ "max_peaks": 24,
254
+ "annotate_seconds": 25.62758301384747
255
+ },
256
+ {
257
+ "name": "sub25",
258
+ "source": "synth",
259
+ "conditions": "from-rows 2 hash-range 0,0.25 (485 conditions)",
260
+ "bundle": "builds/5Ms-I1+I3x2-L",
261
+ "members": "1,3",
262
+ "structures": 435076,
263
+ "tasks": 217012,
264
+ "failed": 0,
265
+ "rows": 217012,
266
+ "seconds": 227.74617605702952,
267
+ "schema": "allcond-inverse-selfsim/v1",
268
+ "floor": 0.02,
269
+ "spectra": 217000,
270
+ "without_peak": 12,
271
+ "mean_annotated_peaks": 15.62026728110599,
272
+ "mean_candidates": 20.61567563084069,
273
+ "ppm": 20.0,
274
+ "floor_da": 0.004,
275
+ "max_peaks": 24,
276
+ "annotate_seconds": 10.8742212112993
277
+ }
278
+ ],
279
+ "spectra": 868083,
280
+ "structures": 434944
281
+ },
282
+ {
283
+ "name": "train3",
284
+ "source": "synth",
285
+ "conditions": "from-train 3 window 2000 seed 0 hash-range 0,1 (301 conditions)",
286
+ "bundle": "builds/5Ms-I1+I3x2-L",
287
+ "members": "1,3",
288
+ "structures": 435076,
289
+ "tasks": 1163260,
290
+ "failed": 0,
291
+ "rows": 1163260,
292
+ "seconds": 1262.926542488858,
293
+ "schema": "allcond-inverse-selfsim/v1",
294
+ "floor": 0.02,
295
+ "spectra": 1163218,
296
+ "without_peak": 42,
297
+ "mean_annotated_peaks": 17.868643710809152,
298
+ "mean_candidates": 25.376552103570997,
299
+ "ppm": 20.0,
300
+ "floor_da": 0.004,
301
+ "max_peaks": 24,
302
+ "annotate_seconds": 43.40271363919601
303
+ }
304
+ ],
305
+ "spectra": 2031301,
306
+ "structures": 434950
307
+ }
308
+ },
309
+ "stores": {
310
+ "gen-x1.manifest.json": null,
311
+ "pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758",
312
+ "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
313
+ "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597",
314
+ "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"
315
+ },
316
+ "recipe": {
317
+ "store": "/workspace/research/inv",
318
+ "out": "runs/I13-ff512x6-real08-corpus2+train3-seed2",
319
+ "pools": "pools-validation-input.npz",
320
+ "input": "/workspace/research/validation-input",
321
+ "steps": 50000,
322
+ "batch": 512,
323
+ "real_frac": 0.08,
324
+ "real_balance": 0.5,
325
+ "no_synth": false,
326
+ "target": "",
327
+ "heavy_boost": 1.0,
328
+ "synth_structures": 1.0,
329
+ "extra_store": "inv/selfsim/corpus2+train3.npz",
330
+ "extra_frac": 0.5,
331
+ "heavy_lo": 450.0,
332
+ "heavy_hi": 850.0,
333
+ "lr": 0.0006,
334
+ "weight_decay": 0.0001,
335
+ "warmup": 2000,
336
+ "ema": 0.999,
337
+ "dim": 512,
338
+ "layers": 6,
339
+ "heads": 8,
340
+ "dropout": 0.1,
341
+ "peak_drop": 0.3,
342
+ "truncate": 0.3,
343
+ "min_keep": 4,
344
+ "jitter": 0.15,
345
+ "synth_energy_drop": 0.3,
346
+ "synth_instrument_drop": 0.3,
347
+ "validate_every": 5000,
348
+ "retrieve_every": 10000,
349
+ "retrieve_pools": 0,
350
+ "seed": 2,
351
+ "contrast_weight": 0.0,
352
+ "contrast_k": 31,
353
+ "contrast_tau": 0.05,
354
+ "init": null,
355
+ "max_steps_this_run": 0
356
+ }
357
+ }
358
+ ],
359
+ "library_asset": {
360
+ "schema": "train-library/v1",
361
+ "fold": "train",
362
+ "spectra": 194119,
363
+ "molecules": 25046,
364
+ "bin": 0.01,
365
+ "prior_sample": 1024,
366
+ "prior_seed": 0,
367
+ "fragment_entries": 6765621,
368
+ "loss_entries": 6763777,
369
+ "source": "/workspace/data/train.jsonl",
370
+ "source_sha256": "de7f46e91f7b67eb5ca67468381abbdb20a8f739a0ae4b27070276798ecad0fc",
371
+ "structure_store": "/run-state/workspace/research/inv/real-train.npz",
372
+ "structure_store_sha256": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6",
373
+ "sha256": "22d1ddb216eef78d5040209ff6526371e5fcd8aac88e68a36823b6d5e73b3712",
374
+ "elapsed_seconds": 13.499361125999712,
375
+ "training_data": "MassSpecGym v1.5 train fold"
376
+ },
377
+ "ion_asset": {
378
+ "source": "inv/ion_error_table.json",
379
+ "sha256": "39aca3a714fd25a20cbaba4d46942764b52d730828a4b8205bf46ac5fc2a6225",
380
+ "fit": "MassSpecGym v1.5 train fold (194,119 spectra with [M+H]+ / [M+Na]+), analysis/ion_fit.py",
381
+ "training_data": "MassSpecGym v1.5 train fold",
382
+ "window_da": 0.02,
383
+ "classes": {
384
+ "domain|Orbitrap": {
385
+ "median_ppm": 0.358,
386
+ "sigma_ppm": 1.51,
387
+ "background": 0.0204
388
+ },
389
+ "domain|QTOF": {
390
+ "median_ppm": 0.027,
391
+ "sigma_ppm": 4.348,
392
+ "background": 0.0869
393
+ },
394
+ "domain|other": {
395
+ "median_ppm": 0.027,
396
+ "sigma_ppm": 4.348,
397
+ "background": 0.0869
398
+ },
399
+ "H-no-energy|Orbitrap": {
400
+ "median_ppm": -0.943,
401
+ "sigma_ppm": 2.937,
402
+ "background": 0.1196
403
+ },
404
+ "H-no-energy|QTOF": {
405
+ "median_ppm": 0.665,
406
+ "sigma_ppm": 4.58,
407
+ "background": 0.08
408
+ },
409
+ "H-no-energy|other": {
410
+ "median_ppm": 0.851,
411
+ "sigma_ppm": 4.269,
412
+ "background": 0.1021
413
+ },
414
+ "Na|Orbitrap": {
415
+ "median_ppm": -0.872,
416
+ "sigma_ppm": 2.935,
417
+ "background": 0.1834
418
+ },
419
+ "Na|QTOF": {
420
+ "median_ppm": 1.119,
421
+ "sigma_ppm": 4.103,
422
+ "background": 0.0798
423
+ },
424
+ "Na|other": {
425
+ "median_ppm": 3.003,
426
+ "sigma_ppm": 2.969,
427
+ "background": 0.0319
428
+ }
429
+ },
430
+ "fallback": {
431
+ "domain|other": "domain|QTOF"
432
+ },
433
+ "train_fold_statistics": {
434
+ "domain|Orbitrap": {
435
+ "spectra": 64738,
436
+ "present": 0.5068,
437
+ "median_ppm": 0.176,
438
+ "sigma_ppm": 1.093,
439
+ "beyond_3_sigma": 0.0763,
440
+ "ml_median_ppm": 0.358,
441
+ "ml_sigma_ppm": 1.51,
442
+ "ml_background": 0.0204,
443
+ "ml_mean_log_density": 6.2307
444
+ },
445
+ "domain|QTOF": {
446
+ "spectra": 34603,
447
+ "present": 0.5111,
448
+ "median_ppm": 0.045,
449
+ "sigma_ppm": 3.847,
450
+ "beyond_3_sigma": 0.0914,
451
+ "ml_median_ppm": 0.027,
452
+ "ml_sigma_ppm": 4.348,
453
+ "ml_background": 0.0869,
454
+ "ml_mean_log_density": 4.9233
455
+ },
456
+ "domain|other": {
457
+ "spectra": 36,
458
+ "present": 0.8611,
459
+ "median_ppm": 2.185,
460
+ "sigma_ppm": 4.203,
461
+ "beyond_3_sigma": 0.0,
462
+ "ml_median_ppm": 2.789,
463
+ "ml_sigma_ppm": 3.688,
464
+ "ml_background": 0.0,
465
+ "ml_mean_log_density": 5.6271
466
+ },
467
+ "H-no-energy|Orbitrap": {
468
+ "spectra": 51704,
469
+ "present": 0.5479,
470
+ "median_ppm": -0.989,
471
+ "sigma_ppm": 2.647,
472
+ "beyond_3_sigma": 0.1156,
473
+ "ml_median_ppm": -0.943,
474
+ "ml_sigma_ppm": 2.937,
475
+ "ml_background": 0.1196,
476
+ "ml_mean_log_density": 5.0128
477
+ },
478
+ "H-no-energy|QTOF": {
479
+ "spectra": 11931,
480
+ "present": 0.626,
481
+ "median_ppm": 0.801,
482
+ "sigma_ppm": 4.239,
483
+ "beyond_3_sigma": 0.0754,
484
+ "ml_median_ppm": 0.665,
485
+ "ml_sigma_ppm": 4.58,
486
+ "ml_background": 0.08,
487
+ "ml_mean_log_density": 4.7319
488
+ },
489
+ "H-no-energy|other": {
490
+ "spectra": 3467,
491
+ "present": 0.67,
492
+ "median_ppm": 0.61,
493
+ "sigma_ppm": 4.033,
494
+ "beyond_3_sigma": 0.0895,
495
+ "ml_median_ppm": 0.851,
496
+ "ml_sigma_ppm": 4.269,
497
+ "ml_background": 0.1021,
498
+ "ml_mean_log_density": 4.7426
499
+ },
500
+ "Na|Orbitrap": {
501
+ "spectra": 24454,
502
+ "present": 0.4954,
503
+ "median_ppm": -0.875,
504
+ "sigma_ppm": 2.797,
505
+ "beyond_3_sigma": 0.1574,
506
+ "ml_median_ppm": -0.872,
507
+ "ml_sigma_ppm": 2.935,
508
+ "ml_background": 0.1834,
509
+ "ml_mean_log_density": 4.589
510
+ },
511
+ "Na|QTOF": {
512
+ "spectra": 2361,
513
+ "present": 0.8602,
514
+ "median_ppm": 1.126,
515
+ "sigma_ppm": 4.211,
516
+ "beyond_3_sigma": 0.0551,
517
+ "ml_median_ppm": 1.119,
518
+ "ml_sigma_ppm": 4.103,
519
+ "ml_background": 0.0798,
520
+ "ml_mean_log_density": 4.5016
521
+ },
522
+ "Na|other": {
523
+ "spectra": 825,
524
+ "present": 0.9309,
525
+ "median_ppm": 3.305,
526
+ "sigma_ppm": 3.261,
527
+ "beyond_3_sigma": 0.026,
528
+ "ml_median_ppm": 3.003,
529
+ "ml_sigma_ppm": 2.969,
530
+ "ml_background": 0.0319,
531
+ "ml_mean_log_density": 4.9698
532
+ }
533
+ }
534
+ },
535
+ "method": "native-glacier-all-conditions-forward-simulation+formula-transformer-inverse-fusion+train-library-neighbour-fusion+measured-intact-ion-term",
536
+ "inference_runtime": "/opt/glacier/bin/python",
537
+ "members": [
538
+ {
539
+ "source": "runs/B-all-b32/checkpoints/step-0020000-ema.ckpt",
540
+ "sha256": "80ec4715663e505007412e3d89dbe6020cd145713840c15affaf82604fc5765c",
541
+ "global_step": 20000,
542
+ "weights": "ema",
543
+ "subset": "all",
544
+ "seed": 1,
545
+ "elapsed_seconds": 5136.2172065590275,
546
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
547
+ "initialization": "random",
548
+ "training_data": "MassSpecGym v1.5 train fold",
549
+ "synthetic": null
550
+ },
551
+ {
552
+ "source": "runs/B-all-b32/checkpoints/step-0030000-ema.ckpt",
553
+ "sha256": "0228e4fa0589333f02db8223e49d6ecd23cd100d1a6182ad268d9db4f8468798",
554
+ "global_step": 30000,
555
+ "weights": "ema",
556
+ "subset": "all",
557
+ "seed": 1,
558
+ "elapsed_seconds": 8314.79353913304,
559
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
560
+ "initialization": "random",
561
+ "training_data": "MassSpecGym v1.5 train fold",
562
+ "synthetic": null
563
+ },
564
+ {
565
+ "source": "runs/B2-all-b32-seed2/checkpoints/step-0020000-ema.ckpt",
566
+ "sha256": "ac3c85cdabfb5f68ca3ca506732f1f8e79fa6e9f3d5bf5fa574b6e1f3249312c",
567
+ "global_step": 20000,
568
+ "weights": "ema",
569
+ "subset": "all",
570
+ "seed": 2,
571
+ "elapsed_seconds": 6945.18486677701,
572
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
573
+ "initialization": "random",
574
+ "training_data": "MassSpecGym v1.5 train fold",
575
+ "synthetic": null
576
+ },
577
+ {
578
+ "source": "runs/B2-all-b32-seed2/checkpoints/step-0030000-ema.ckpt",
579
+ "sha256": "f11667b8ee56a15c82d26ab9052816ecc63f2081eaa6fd4ea82d38d1e2a36ce1",
580
+ "global_step": 30000,
581
+ "weights": "ema",
582
+ "subset": "all",
583
+ "seed": 2,
584
+ "elapsed_seconds": 9298.063542745003,
585
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
586
+ "initialization": "random",
587
+ "training_data": "MassSpecGym v1.5 train fold",
588
+ "synthetic": null
589
+ },
590
+ {
591
+ "source": "runs/S2-B20k-synth/checkpoints/step-0030000-ema.ckpt",
592
+ "sha256": "44742fa31f337e59b543d903018780560569c2e8e50a008ba008859c8440e3a2",
593
+ "global_step": 30000,
594
+ "weights": "ema",
595
+ "subset": "all",
596
+ "seed": 1,
597
+ "elapsed_seconds": 8906.84206694903,
598
+ "data_manifest_sha256": "695645989510ffe27c4ef0d2f8f8f14f8c5514a5b23b5bfda55a7d035503ed4e",
599
+ "initialization": "own-checkpoint:6c5a43fada335e5f4e9f8265ec6cfd54a8f317eb8d2100d96ea3731809005be2",
600
+ "training_data": "MassSpecGym v1.5 train fold + supplied synthetic_augmentation corpus (auxiliary intensity term)",
601
+ "synthetic": {
602
+ "manifest_sha256": "431454e702f8da3194684fc9a565135d29a38a407aaa31e9399d871360c6e962",
603
+ "per_item": 1,
604
+ "weight": 1.0,
605
+ "schedule": "e2e",
606
+ "instrument": "flag"
607
+ }
608
+ }
609
+ ]
610
+ }
shared/package-files/b6/b665f6a2fef711b478e6d456fccce87c974a50619a8d892d36877cff76f550fb/anchor-B2-30k-v1.log ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"event": "canonicalized", "unique_candidates": 828117, "unsupported": 7007, "seconds": 97.26345496000431}
2
+ {"event": "planned", "queries": 19429, "groups": 10176, "molecules": 944948, "simulations": 2540748, "seconds": 101.55645595600072}
3
+ ind noprec_pp0.75 all=0.4708 dom=0.7065 non=0.2342 h5=0.6277 h20=0.7027 | diff=+0.0000 [+0.0000,+0.0000]
4
+ agg noprec_pp0.75 all=0.6118 dom=0.7970 non=0.4259 h5=0.7901 h20=0.8848 | diff=+0.0000 [+0.0000,+0.0000]
5
+ ind a5 all=0.4807 dom=0.7139 non=0.2466 h5=0.6380 h20=0.7151 | diff=+0.0099 [+0.0063,+0.0137]
6
+ agg a5 all=0.6308 dom=0.8033 non=0.4576 h5=0.7779 h20=0.8902 | diff=+0.0189 [+0.0046,+0.0343]
7
+ ind a10 all=0.4778 dom=0.7033 non=0.2515 h5=0.6410 h20=0.7231 | diff=+0.0070 [+0.0024,+0.0115]
8
+ agg a10 all=0.6282 dom=0.8031 non=0.4527 h5=0.7753 h20=0.8969 | diff=+0.0164 [-0.0010,+0.0353]
9
+ ind p10f1 all=0.4786 dom=0.7161 non=0.2400 h5=0.6326 h20=0.7086 | diff=+0.0077 [+0.0034,+0.0122]
10
+ agg p10f1 all=0.6340 dom=0.8107 non=0.4566 h5=0.7848 h20=0.8930 | diff=+0.0222 [+0.0059,+0.0390]
11
+ ind p10f2 all=0.4803 dom=0.7191 non=0.2405 h5=0.6343 h20=0.7105 | diff=+0.0095 [+0.0055,+0.0138]
12
+ agg p10f2 all=0.6290 dom=0.8101 non=0.4470 h5=0.7792 h20=0.8933 | diff=+0.0171 [+0.0029,+0.0319]
13
+ ind p20f2 all=0.4804 dom=0.7137 non=0.2461 h5=0.6392 h20=0.7178 | diff=+0.0095 [+0.0058,+0.0133]
14
+ agg p20f2 all=0.6282 dom=0.8080 non=0.4477 h5=0.7778 h20=0.8939 | diff=+0.0164 [-0.0012,+0.0330]
15
+ ind p20f4 all=0.4818 dom=0.7145 non=0.2482 h5=0.6397 h20=0.7193 | diff=+0.0110 [+0.0074,+0.0148]
16
+ agg p20f4 all=0.6335 dom=0.8097 non=0.4566 h5=0.7739 h20=0.8931 | diff=+0.0217 [+0.0068,+0.0360]
shared/package-files/b6/b672973a2274bbae420e6ec878372a79c2cb4916874eb9a6436bc7dd9292db77/eval-A10k-sub.log ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"event": "loaded", "queries": 2147, "pools": 413, "members": 1, "workers": 10}
2
+ {"event": "canonicalized", "unique_candidates": 105046, "unsupported": 928, "seconds": 6.584986881003715}
3
+ {"event": "planned", "queries": 2147, "groups": 1168, "molecules": 116802, "simulations": 295422, "seconds": 6.909893014002591}
4
+ /opt/glacier/lib/python3.11/site-packages/torch/nn/modules/transformer.py:508: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)
5
+ output = torch._nested_tensor_from_mask(
6
+ {"event": "simulating", "finished": 64000, "total": 295422, "seconds": 25.547479233937338, "per_second": 2505.1395252719194}
7
+ {"event": "simulating", "finished": 128000, "total": 295422, "seconds": 53.30762945394963, "per_second": 2401.1572322977554}
8
+ {"event": "simulating", "finished": 192000, "total": 295422, "seconds": 83.09802224196028, "per_second": 2310.5243039472693}
9
+ {"event": "simulating", "finished": 256000, "total": 295422, "seconds": 118.18431927694473, "per_second": 2166.1080045662216}
10
+ {"event": "simulated", "finished": 295422, "total": 295422, "failed": 0, "seconds": 157.38810374191962}
11
+ {"event": "finished", "queries": 2147, "unscored": 0, "directly_scored": 2147, "scoring_seconds": 1.8268624959746376, "total_seconds": 178.23354486201424}
shared/package-files/b6/b6da4868833e77578579a777da141035ab34a9f6aa4814e9233c49b7697c721d/netprop-oracle-rerank.log ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ released rule from saved branches: Hit@1 0.7713 pool-macro 0.8591 Hit@5 0.8797 Hit@20 0.9377 | twin pools 354 (macro 0.7429), others macro 0.8727
2
+
3
+ k=10 p=3.0 half=0.5 | m=10 tau=0.02 kernel=plain debias=fold control=none shuffle=None: alone Hit@1 0.6827 (domain 0.7351 / other 0.6302) macro 0.6917
4
+ lam 1.00/1.00: Hit@1 0.8167 (+0.0453 [+0.0172, +0.0750]) macro +0.0269 [+0.0195, +0.0337] pools +120 -29 | no top 20 +0.0386 cap30 +0.0410 [+0.0233, +0.0597] | Hit@5 +0.0072 macro@5 +0.0171 cap30@5 +0.0252 | twin +31 -5 other +89 -24 | dom +0.0078 oth +0.0830
5
+ lam 2.00/2.00: Hit@1 0.8186 (+0.0473 [+0.0188, +0.0776]) macro +0.0298 [+0.0222, +0.0375] pools +143 -42 | no top 20 +0.0408 cap30 +0.0427 [+0.0239, +0.0617] | Hit@5 +0.0097 macro@5 +0.0177 cap30@5 +0.0279 | twin +35 -7 other +108 -35 | dom +0.0079 oth +0.0868
shared/package-files/b9/b9052f01862eff09dd56dcd554fe689280543be8842bdb1fbbfc31cd4651f67d/consensus-subset-5Ms-N5u48-L-ion-net2.log ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ == val-tiny
2
+ consensus {"scored": 168, "pools": 40, "answers": 400, "search_seconds": 0.02126318495720625, "seconds": 0.5167010120349005}
3
+ subset: 180 labelled queries, 40 pools
4
+ base (no term) Hit@1 0.9056 Hit@5 0.9111 Hit@20 0.9889 macro 0.9000
5
+ lab, with term Hit@1 0.9056 Hit@5 0.9111 Hit@20 0.9889 macro 0.9000 pools +0 -0
6
+ replay Hit@1 0.9056 Hit@5 0.9111 Hit@20 0.9889 macro 0.9000
7
+ label rank: replay == lab for 180 / 180 queries; top-1 status equal for 180; largest rank difference 0
8
+ == val-sub
9
+ consensus {"scored": 2147, "pools": 413, "answers": 4130, "search_seconds": 0.3022799019818194, "seconds": 1.2762607790064067}
10
+ subset: 2147 labelled queries, 413 pools
11
+ base (no term) Hit@1 0.7946 Hit@5 0.8836 Hit@20 0.8975 macro 0.9080
12
+ lab, with term Hit@1 0.7960 Hit@5 0.8896 Hit@20 0.8985 macro 0.9128 pools +2 -0
13
+ replay Hit@1 0.7960 Hit@5 0.8896 Hit@20 0.8985 macro 0.9128
14
+ label rank: replay == lab for 2147 / 2147 queries; top-1 status equal for 2147; largest rank difference 0
shared/package-files/b9/b90cf50c87342676bbe09c8ae51cac8f1e1bdf71c9feefd8e3b49ed81e3d8435/eval-B3-20k30k-full.log ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"event": "loaded", "queries": 19429, "pools": 3386, "members": 2, "inverse_members": 0, "workers": 13}
2
+ {"event": "canonicalized", "unique_candidates": 828117, "unsupported": 7007, "seconds": 46.74947982298909}
3
+ {"event": "planned", "queries": 19429, "groups": 10176, "molecules": 944948, "simulations": 2540748, "seconds": 51.1962361879996}
4
+ /opt/glacier/lib/python3.11/site-packages/torch/nn/modules/transformer.py:508: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)
5
+ output = torch._nested_tensor_from_mask(
6
+ {"event": "simulating", "finished": 64000, "total": 2540748, "seconds": 44.795899763994385, "per_second": 1428.7021878605349}
7
+ {"event": "simulating", "finished": 128000, "total": 2540748, "seconds": 95.81436122699233, "per_second": 1335.9166450711618}
8
+ {"event": "simulating", "finished": 192000, "total": 2540748, "seconds": 156.3520112769911, "per_second": 1227.9982740986645}
9
+ {"event": "simulating", "finished": 256000, "total": 2540748, "seconds": 224.53731132199755, "per_second": 1140.122318614938}
10
+ {"event": "simulating", "finished": 320000, "total": 2540748, "seconds": 293.4880397050001, "per_second": 1090.3340399208378}
11
+ {"event": "simulating", "finished": 384000, "total": 2540748, "seconds": 360.1647454939957, "per_second": 1066.1787551508191}
12
+ {"event": "simulating", "finished": 448000, "total": 2540748, "seconds": 418.1166265310021, "per_second": 1071.4713827979815}
13
+ {"event": "simulating", "finished": 512000, "total": 2540748, "seconds": 470.3000572279998, "per_second": 1088.6666759468078}
14
+ {"event": "simulating", "finished": 576000, "total": 2540748, "seconds": 522.1164347479935, "per_second": 1103.2022010148257}
15
+ {"event": "simulating", "finished": 640000, "total": 2540748, "seconds": 575.3273913500016, "per_second": 1112.4100983585097}
16
+ {"event": "simulating", "finished": 704000, "total": 2540748, "seconds": 628.9084987160022, "per_second": 1119.3997241845304}
17
+ {"event": "simulating", "finished": 768000, "total": 2540748, "seconds": 681.5814538519917, "per_second": 1126.7912230586815}
18
+ {"event": "simulating", "finished": 832000, "total": 2540748, "seconds": 733.6929246200016, "per_second": 1133.9894008531078}
19
+ {"event": "simulating", "finished": 896000, "total": 2540748, "seconds": 786.778832786993, "per_second": 1138.820673182214}
20
+ {"event": "simulating", "finished": 960000, "total": 2540748, "seconds": 841.7193612550036, "per_second": 1140.5226542117791}
21
+ {"event": "simulating", "finished": 1024000, "total": 2540748, "seconds": 896.5344541950035, "per_second": 1142.1758474630487}
22
+ {"event": "simulating", "finished": 1088000, "total": 2540748, "seconds": 951.7571967349941, "per_second": 1143.148697727096}
23
+ {"event": "simulating", "finished": 1152000, "total": 2540748, "seconds": 1007.1731549319957, "per_second": 1143.7953785392374}
24
+ {"event": "simulating", "finished": 1216000, "total": 2540748, "seconds": 1064.3267858329928, "per_second": 1142.5062454369224}
25
+ {"event": "simulating", "finished": 1280000, "total": 2540748, "seconds": 1121.7517844069953, "per_second": 1141.072399253335}
26
+ {"event": "simulating", "finished": 1344000, "total": 2540748, "seconds": 1178.0490213000012, "per_second": 1140.8693320052748}
27
+ {"event": "simulating", "finished": 1408000, "total": 2540748, "seconds": 1234.53782451, "per_second": 1140.507785218204}
28
+ {"event": "simulating", "finished": 1472000, "total": 2540748, "seconds": 1294.210176624998, "per_second": 1137.3732231333838}
29
+ {"event": "simulating", "finished": 1536000, "total": 2540748, "seconds": 1354.8910518829944, "per_second": 1133.6704880184313}
30
+ {"event": "simulating", "finished": 1600000, "total": 2540748, "seconds": 1415.960358005992, "per_second": 1129.9751373358922}
31
+ {"event": "simulating", "finished": 1664000, "total": 2540748, "seconds": 1473.6734530209942, "per_second": 1129.1510996475108}
32
+ {"event": "simulating", "finished": 1728000, "total": 2540748, "seconds": 1530.6084766400018, "per_second": 1128.9627794256785}
33
+ {"event": "simulating", "finished": 1792000, "total": 2540748, "seconds": 1592.978536744995, "per_second": 1124.9366885141308}
34
+ {"event": "simulating", "finished": 1856000, "total": 2540748, "seconds": 1658.1823645760014, "per_second": 1119.2978767896743}
35
+ {"event": "simulating", "finished": 1920000, "total": 2540748, "seconds": 1723.811496430004, "per_second": 1113.8108801201874}
36
+ {"event": "simulating", "finished": 1984000, "total": 2540748, "seconds": 1791.5775208889972, "per_second": 1107.403936959156}
37
+ {"event": "simulating", "finished": 2048000, "total": 2540748, "seconds": 1859.568839575004, "per_second": 1101.3305645990827}
38
+ {"event": "simulating", "finished": 2112000, "total": 2540748, "seconds": 1930.9721073730034, "per_second": 1093.749615510126}
39
+ {"event": "simulating", "finished": 2176000, "total": 2540748, "seconds": 2005.1599653080048, "per_second": 1085.200202301941}
40
+ {"event": "simulating", "finished": 2240000, "total": 2540748, "seconds": 2078.349586412005, "per_second": 1077.7782595598187}
41
+ {"event": "simulating", "finished": 2304000, "total": 2540748, "seconds": 2164.397603610996, "per_second": 1064.4994229138383}
42
+ {"event": "simulating", "finished": 2368000, "total": 2540748, "seconds": 2263.5635348249925, "per_second": 1046.1380754585632}
43
+ {"event": "simulating", "finished": 2432000, "total": 2540748, "seconds": 2378.516571646993, "per_second": 1022.4860440286833}
44
+ {"event": "simulating", "finished": 2496000, "total": 2540748, "seconds": 2509.6223233649944, "per_second": 994.5719627857275}
45
+ {"event": "simulated", "finished": 2540748, "total": 2540748, "failed": 0, "seconds": 2611.8383371969976}
46
+ {"event": "finished", "queries": 19429, "unscored": 0, "directly_scored": 19429, "scoring_seconds": 26.95632052900328, "total_seconds": 2709.5768285280064}
shared/package-files/b9/b9694748013b9646c614eb704e331f78559c87cce1b9ed50225c4135bcc38eb9/e74-term-union-hyb-both.log ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ released rule from saved branches: Hit@1 0.7783 pool-macro 0.8706 Hit@5 0.8754 Hit@20 0.9314 | twin pools 354 (macro 0.7655), others macro 0.8829
2
+
3
+ k=10 p=3.0 half=0.5 | m=10 tau=0.02 kernel=plain fps=morgan debias=fold control=none shuffle=None: alone Hit@1 0.4557 (domain 0.4201 / other 0.4915) macro 0.3783
4
+ lam 1.00/1.00: Hit@1 0.8029 (+0.0246 [+0.0086, +0.0460]) macro +0.0115 [+0.0053, +0.0177] pools +81 -42 | no top 20 +0.0186 cap30 +0.0186 [+0.0076, +0.0313] | Hit@5 +0.0239 macro@5 +0.0065 cap30@5 +0.0150
5
+ without consensus term: Hit@1 0.7783 (ref 0.7783; +0.0000 [+0.0000, +0.0000]) macro 0.8706 (+0.0000 [+0.0000, +0.0000]) pools +0 -0 cap30 +0.0000 | Hit@5 0.8754 (+0.0000) Hit@20 0.9314 (+0.0000)
6
+ macro Hit@5 0.9333 (+0.0000 [+0.0000, +0.0000]) pools +0 -0
7
+ macro Hit@20 0.9652 (+0.0000 [+0.0000, +0.0000]) pools +0 -0
8
+ domain 0.9374 (+0.0000) other 0.6185 (+0.0000) | Orbitrap no energy 0.5840 (+0.0000) | above 850 0.4653 (+0.0000) below 850 0.8495 (+0.0000)
9
+ full rule : Hit@1 0.8029 (ref 0.8173; -0.0145 [-0.0273, -0.0036]) macro 0.8822 (-0.0044 [-0.0074, -0.0015]) pools +5 -20 cap30 -0.0118 | Hit@5 0.8993 (+0.0255) Hit@20 0.9288 (+0.0243)
10
+ macro Hit@5 0.9398 (-0.0006 [-0.0030, +0.0015]) pools +7 -9
11
+ macro Hit@20 0.9628 (-0.0032 [-0.0059, -0.0009]) pools +4 -15
12
+ domain 0.9431 (-0.0022) other 0.6621 (-0.0268) | Orbitrap no energy 0.6314 (-0.0304) | above 850 0.5144 (-0.0411) below 850 0.8686 (-0.0084)
shared/package-files/b9/b98c577aafa80ccc1c381a527016ca758c6239da8627aba8bfb5a6bde98a3876/na_ion_types.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Which ion type explains the fragments of [M+Na]+ spectra? (train fold, labels of the TRAIN fold only)
2
+
3
+ For sampled train spectra: share of fragment intensity (m/z < precursor - 0.5; sqrt intensities
4
+ as in the inverse branch) that has a sub-formula of the label formula when the ion is taken as
5
+ sodiated (m/z - Na + e), as protonated (m/z - H + e), either, or neither. [M+H]+ spectra are a
6
+ reference (their 'sodiated' share measures chance matches).
7
+ """
8
+ import json
9
+ import sys
10
+ from pathlib import Path
11
+
12
+ import numpy as np
13
+
14
+ sys.path.insert(0, '/workspace/research/allcond/pkg')
15
+ import inverse
16
+
17
+ rng = np.random.default_rng(0)
18
+ rows = {'[M+Na]+': [], '[M+H]+ no energy': [], '[M+H]+ energy': []}
19
+ with open('/workspace/data/train.jsonl') as stream:
20
+ for line in stream:
21
+ row = json.loads(line)
22
+ key = '[M+Na]+' if row['adduct'] == '[M+Na]+' else ('[M+H]+ energy' if row['collision_energy'] is not None else '[M+H]+ no energy')
23
+ rows[key].append(row)
24
+ for key, items in rows.items():
25
+ print(key, len(items), 'train spectra', len({r['target_inchikey'] for r in items}), 'molecules')
26
+ pick = rng.choice(len(items), size=min(4000, len(items)), replace=False)
27
+ shares = {'sodiated': [], 'protonated': [], 'either': [], 'both': [], 'peaks': [], 'top': {'sodiated': 0, 'protonated': 0, 'none': 0}}
28
+ for i in pick:
29
+ row = items[i]
30
+ root = inverse.parse_formula(row['target_formula'])
31
+ if root is None:
32
+ continue
33
+ table = inverse.SubformulaTable(root)
34
+ mz, inten = inverse.select_peaks(row['mzs'], row['intensities'])
35
+ keep = mz < row['precursor_mz'] - 0.5
36
+ mz, inten = mz[keep], inten[keep]
37
+ if len(mz) == 0:
38
+ continue
39
+ tol = inverse.tolerance_of(mz, 20, 0.004)
40
+ na, _, _ = table.assign(mz - inverse.ION_SHIFT['[M+Na]+'], tol)
41
+ h, _, _ = table.assign(mz - inverse.ION_SHIFT['[M+H]+'], tol)
42
+ total = inten.sum()
43
+ shares['sodiated'].append(inten[na].sum() / total)
44
+ shares['protonated'].append(inten[h].sum() / total)
45
+ shares['either'].append(inten[na | h].sum() / total)
46
+ shares['both'].append(inten[na & h].sum() / total)
47
+ shares['peaks'].append(len(mz))
48
+ shares['top']['sodiated' if na[0] else ('protonated' if h[0] else 'none')] += 1
49
+ n = len(shares['peaks'])
50
+ print(f' sampled {n}; fragment peaks per spectrum (top 50) mean {np.mean(shares["peaks"]):.1f}; '
51
+ f'intensity share sodiated {np.mean(shares["sodiated"]):.3f} protonated {np.mean(shares["protonated"]):.3f} '
52
+ f'either {np.mean(shares["either"]):.3f} both {np.mean(shares["both"]):.3f}; strongest fragment: {shares["top"]}', flush=True)
53
+ for lo, hi in ((1, 5), (5, 10), (10, 25), (25, 51)):
54
+ m = (np.array(shares['peaks']) >= lo) & (np.array(shares['peaks']) < hi)
55
+ if m.sum():
56
+ print(f' {lo}-{hi - 1} fragment peaks: n={int(m.sum())} sodiated {np.array(shares["sodiated"])[m].mean():.3f} protonated {np.array(shares["protonated"])[m].mean():.3f} either {np.array(shares["either"])[m].mean():.3f}')
shared/package-files/b9/b9f06e0872fa67848191be9925014bb305879bf9c389875d28f68a13f145a45a/recur-mof-5.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ DGL backend not selected or invalid. Assuming PyTorch for now.
2
+ Setting the default backend to "pytorch". You can change it in the ~/.dgl/config.json file or export the DGLBACKEND environment variable. Valid options are: pytorch, mxnet, tensorflow (all lowercase)
3
+ merged tol 5 mDa: 7962 of 19429 spectra in groups of >= 3; peaks 139302 -> 1015127; 0 s
4
+ wrote /workspace/research/inv/recur/mof-5.npz 169 s
shared/package-files/cb/cb1d041f6c7ed56984943cbd6808c4525c639a0326a92a31c24b585fe260ee53/eval-B10k-sub-hce10.log ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {"event": "loaded", "queries": 502, "pools": 103, "members": 1, "workers": 8}
2
+ {"event": "canonicalized", "unique_candidates": 26064, "unsupported": 328, "seconds": 3.0041393630672246}
3
+ {"event": "planned", "queries": 502, "groups": 129, "molecules": 25736, "simulations": 32261, "seconds": 3.046831817016937}
4
+ /opt/glacier/lib/python3.11/site-packages/torch/nn/modules/transformer.py:508: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)
5
+ output = torch._nested_tensor_from_mask(
6
+ {"event": "simulated", "finished": 32261, "total": 32261, "failed": 0, "seconds": 34.809406278072856}
7
+ {"event": "finished", "queries": 502, "unscored": 0, "directly_scored": 502, "scoring_seconds": 0.7901481019798666, "total_seconds": 50.95482438802719}
shared/package-files/e6/e614f3e10a43efafcb939209fd868bbdc2c1c38b8b8cb0ec8082fa0750438a87/eval-score-I9b+I13.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 2 members, 47.0 s
2
+ cos pool rule all=0.5852 dom=0.8266 non=0.3428 h5=0.6985 macro=0.6976
3
+ cen pool rule all=0.5716 dom=0.7976 non=0.3446 h5=0.6960 macro=0.6713
4
+ cos_pm pool rule all=0.5858 dom=0.8257 non=0.3449 h5=0.6989 macro=0.6973
5
+ dot pool rule all=0.5725 dom=0.7855 non=0.3587 h5=0.7065 macro=0.6825
6
+ geo pool rule all=0.5630 dom=0.8004 non=0.3247 h5=0.6818 macro=0.6804
7
+ idf05 pool rule all=0.5992 dom=0.8297 non=0.3677 h5=0.7050 macro=0.7026
8
+ idf1 pool rule all=0.5975 dom=0.8238 non=0.3703 h5=0.7113 macro=0.6994
9
+ lin pool rule all=0.5391 dom=0.7615 non=0.3158 h5=0.6561 macro=0.6435
10
+ log01 pool rule all=0.5847 dom=0.8154 non=0.3531 h5=0.7102 macro=0.6926
11
+ log05 pool rule all=0.5991 dom=0.8323 non=0.3648 h5=0.7168 macro=0.7082
12
+ log20 pool rule all=0.5998 dom=0.8300 non=0.3687 h5=0.7084 macro=0.7050
13
+ nll02 pool rule all=0.4247 dom=0.6731 non=0.1752 h5=0.5737 macro=0.5425
14
+ nll10 pool rule all=0.3830 dom=0.6158 non=0.1493 h5=0.5220 macro=0.4879
15
+ poolcen pool rule all=0.5665 dom=0.7832 non=0.3488 h5=0.6996 macro=0.6651
16
+ pow020 pool rule all=0.5475 dom=0.7712 non=0.3229 h5=0.7090 macro=0.6583
17
+ pow030 pool rule all=0.5691 dom=0.7880 non=0.3495 h5=0.7065 macro=0.6740
18
+ pow040 pool rule all=0.5828 dom=0.8059 non=0.3588 h5=0.7124 macro=0.6878
19
+ pow05 pool rule all=0.5956 dom=0.8244 non=0.3659 h5=0.7206 macro=0.6999
20
+ pow05idf025 pool rule all=0.5979 dom=0.8260 non=0.3688 h5=0.7229 macro=0.7044
21
+ pow05idf05 pool rule all=0.6011 dom=0.8242 non=0.3770 h5=0.7236 macro=0.7050
22
+ pow05m pool rule all=0.5958 dom=0.8254 non=0.3652 h5=0.7185 macro=0.7008
23
+ pow05s pool rule all=0.5897 dom=0.8237 non=0.3548 h5=0.7123 macro=0.6982
24
+ pow060 pool rule all=0.5929 dom=0.8295 non=0.3554 h5=0.7133 macro=0.7035
25
+ pow060idf025 pool rule all=0.5990 dom=0.8328 non=0.3643 h5=0.7181 macro=0.7094
26
+ pow060idf050 pool rule all=0.5957 dom=0.8323 non=0.3580 h5=0.7203 macro=0.7088
27
+ pow060m pool rule all=0.5910 dom=0.8300 non=0.3510 h5=0.7121 macro=0.7029
28
+ pow065 pool rule all=0.5953 dom=0.8314 non=0.3583 h5=0.7121 macro=0.7058
29
+ pow070 pool rule all=0.5931 dom=0.8314 non=0.3538 h5=0.7110 macro=0.7050
30
+ pow070idf025 pool rule all=0.5969 dom=0.8304 non=0.3625 h5=0.7147 macro=0.7053
31
+ pow070idf050 pool rule all=0.5984 dom=0.8319 non=0.3639 h5=0.7203 macro=0.7064
32
+ pow070m pool rule all=0.5949 dom=0.8324 non=0.3564 h5=0.7104 macro=0.7056
33
+ pow075 pool rule all=0.5939 dom=0.8299 non=0.3570 h5=0.7108 macro=0.7023
34
+ pow080 pool rule all=0.5914 dom=0.8292 non=0.3528 h5=0.7060 macro=0.7020
35
+ pow080m pool rule all=0.5918 dom=0.8311 non=0.3516 h5=0.7053 macro=0.7044
36
+ pow085 pool rule all=0.5930 dom=0.8296 non=0.3555 h5=0.7051 macro=0.7014
37
+ pow085idf025 pool rule all=0.5949 dom=0.8307 non=0.3581 h5=0.7071 macro=0.7050
38
+ pow085idf050 pool rule all=0.5996 dom=0.8302 non=0.3681 h5=0.7111 macro=0.7035
39
+ pow090 pool rule all=0.5916 dom=0.8296 non=0.3528 h5=0.7018 macro=0.7014
40
+ pow2 pool rule all=0.5439 dom=0.7692 non=0.3177 h5=0.6686 macro=0.6568
41
+ tani pool rule all=0.5531 dom=0.7859 non=0.3193 h5=0.6950 macro=0.6707
shared/package-files/e6/e648ff9e2e19f94d3b1ab306363508326ffb239560531b291ee3c5e1d6e65b62/train-I1.log ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {"args": {"store": "/workspace/research/inv", "out": "/workspace/research/runs/I1-ff512x6", "pools": "pools-validation-input.npz", "input": "/workspace/research/validation-input", "steps": 100000, "batch": 512, "real_frac": 0.2, "real_balance": 0.5, "no_synth": false, "lr": 0.0006, "weight_decay": 0.0001, "warmup": 2000, "ema": 0.999, "dim": 512, "layers": 6, "heads": 8, "dropout": 0.1, "peak_drop": 0.3, "truncate": 0.3, "min_keep": 4, "jitter": 0.15, "synth_energy_drop": 0.3, "synth_instrument_drop": 0.3, "validate_every": 5000, "retrieve_every": 10000, "retrieve_pools": 0, "seed": 1, "max_steps_this_run": 0}, "parameters": 24997832, "model": {"dim": 512, "layers": 6, "heads": 8, "dropout": 0.1, "pair_dim": 64, "fp_bits": 4096, "max_energy": 200.0}, "real_rows": 194119, "synth_rows": 2333477, "stores": {"pools-validation-input.manifest.json": "13490ff1249eca04dc3a23c94838008071da73002c497d5a614d236258301758", "real-train.manifest.json": "59912f8d472101e1f789f5f673f8ef2bdcb521e7973fcd9f4db96bc46d8e46c6", "real-validation.manifest.json": "ff9c1d5b46543ad3da990c1a3aeaa2e4f68939761224689af930a39671fdb597", "synth.manifest.json": "28d43b29c6334a14a568b8f8f412b260ad87934f687903ad927ac57a0f8e20cc"}, "gpu": "NVIDIA H100 80GB HBM3", "load_seconds": 10.601237821974792}
2
+ {"step": 5000, "elapsed": 186.13890091900248, "val_loss": 0.040747812007923885, "val_cos": 0.607664645954501, "val_loss_ema": 0.038519610207106245, "val_cos_ema": 0.6252824411358343}
3
+ {"step": 10000, "elapsed": 362.25227257201914, "val_loss": 0.03945653315487239, "val_cos": 0.6359164749295512, "val_loss_ema": 0.0381329625920117, "val_cos_ema": 0.6478792650282761, "ind_cos": 0.34844819599567656, "agg_cos": 0.37366822790673737, "ind_nll": 0.2355242163775799, "agg_nll": 0.22008338051366513, "retrieval_seconds": 37.391662998939864}
4
+ {"step": 15000, "elapsed": 580.5557692999719, "val_loss": 0.042058228550608145, "val_cos": 0.6346390561366868, "val_loss_ema": 0.039651410800196185, "val_cos_ema": 0.6541293982255906}
5
+ {"step": 20000, "elapsed": 768.8411250649951, "val_loss": 0.043526858921199264, "val_cos": 0.6386919928208156, "val_loss_ema": 0.04045961220871004, "val_cos_ema": 0.6593355409238445, "ind_cos": 0.3928148643779917, "agg_cos": 0.4162334654382624, "ind_nll": 0.2944567399248546, "agg_nll": 0.27572185907663804, "retrieval_seconds": 40.76780899090227}
6
+ {"step": 25000, "elapsed": 1026.3959472359857, "val_loss": 0.04438528773334159, "val_cos": 0.6429097172014933, "val_loss_ema": 0.04203060502970149, "val_cos_ema": 0.6585786283350598}
7
+ {"step": 30000, "elapsed": 1215.0871545899427, "val_loss": 0.045366839635107716, "val_cos": 0.6461526382919862, "val_loss_ema": 0.04303870764853159, "val_cos_ema": 0.660825981176656, "ind_cos": 0.40310875495393483, "agg_cos": 0.43069638169746255, "ind_nll": 0.31144165937516083, "agg_nll": 0.289309794636883, "retrieval_seconds": 49.146196437999606}
8
+ {"step": 35000, "elapsed": 1459.3268311870052, "val_loss": 0.04336966292182835, "val_cos": 0.6546922511230795, "val_loss_ema": 0.042460188359348866, "val_cos_ema": 0.6636769754518858}
9
+ {"step": 40000, "elapsed": 1622.9072967729298, "val_loss": 0.04611901335245276, "val_cos": 0.6464374920584243, "val_loss_ema": 0.043956673539815645, "val_cos_ema": 0.661682640961964, "ind_cos": 0.39749858459004583, "agg_cos": 0.4301816871686654, "ind_nll": 0.31205929280971745, "agg_nll": 0.2888980390138453, "retrieval_seconds": 18.959456807002425}
10
+ {"step": 45000, "elapsed": 1787.2611129719298, "val_loss": 0.04581547190853556, "val_cos": 0.6533226999990349, "val_loss_ema": 0.04388413401527175, "val_cos_ema": 0.6655436792697128}
11
+ {"step": 50000, "elapsed": 1942.9571517329896, "val_loss": 0.046504271270329425, "val_cos": 0.6506629946596422, "val_loss_ema": 0.04458398223567521, "val_cos_ema": 0.6630306731802331, "ind_cos": 0.40151320191466366, "agg_cos": 0.43676977713726906, "ind_nll": 0.31123578156364196, "agg_nll": 0.2834422770085954, "retrieval_seconds": 25.88446972204838}
12
+ {"step": 55000, "elapsed": 2100.881621839013, "val_loss": 0.04540461709516662, "val_cos": 0.6555062814848294, "val_loss_ema": 0.04555788825031806, "val_cos_ema": 0.6603534322887179}
13
+ {"step": 60000, "elapsed": 2259.801618231926, "val_loss": 0.04693030476588536, "val_cos": 0.6539168919628037, "val_loss_ema": 0.046336081780336415, "val_cos_ema": 0.658901367941447, "ind_cos": 0.3936383756240671, "agg_cos": 0.42210098306655003, "ind_nll": 0.30778732822070104, "agg_nll": 0.2840084409902723, "retrieval_seconds": 49.03888284205459}
14
+ {"step": 65000, "elapsed": 2478.1023216280155, "val_loss": 0.04754811765555034, "val_cos": 0.6530461082358362, "val_loss_ema": 0.04653202319784536, "val_cos_ema": 0.6593040133707988}
15
+ {"step": 70000, "elapsed": 2667.324969599955, "val_loss": 0.04719317626324836, "val_cos": 0.6555910125690012, "val_loss_ema": 0.04687043076664956, "val_cos_ema": 0.6569648611913087, "ind_cos": 0.38818261361881723, "agg_cos": 0.4099027227340573, "ind_nll": 0.3042359359720006, "agg_nll": 0.28241288795100106, "retrieval_seconds": 51.90488003694918}
16
+ {"step": 75000, "elapsed": 2874.183002720936, "val_loss": 0.04826307073415746, "val_cos": 0.6527934354432324, "val_loss_ema": 0.04762201781893569, "val_cos_ema": 0.6551878330938127}
17
+ {"step": 80000, "elapsed": 2990.818481713999, "val_loss": 0.04715615471365694, "val_cos": 0.6553639488449933, "val_loss_ema": 0.04705576155727967, "val_cos_ema": 0.656813556581415, "ind_cos": 0.3877193885428998, "agg_cos": 0.4161819959853827, "ind_nll": 0.30372124144320345, "agg_nll": 0.28040557928869214, "retrieval_seconds": 20.976532778004184}
shared/package-files/e6/e6a6e6659a649411fa75007439d647bdf23583d2c21903147bb73d17300d2669/model_manifest.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "massspecgym-retrieval-model/v1",
3
+ "checkpoint": {
4
+ "path": "model.pt",
5
+ "sha256": "5ad8c63711cebaebf946a95ac33f3edd7c57fd06af9184a7a6bc3324863a2e6f",
6
+ "format": "pytorch"
7
+ },
8
+ "model_config": "model_config.json",
9
+ "training_provenance": {
10
+ "data": "MassSpecGym v1.5 train fold",
11
+ "pretrained_assets": []
12
+ }
13
+ }
shared/package-files/ef/ef251257843f98236987449f06d2bcc25c03935762f8f88980d317dc4dfb85af/swap-recompare-N5-recur-hyb-s10-f5-p48-w0.75_0.1_0.5.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ without consensus term: Hit@1 0.7772 (ref 0.7685; +0.0087 [+0.0001, +0.0191]) macro 0.8698 (+0.0047 [+0.0015, +0.0083]) pools +27 -11 cap30 +0.0095 | Hit@5 0.8832 (+0.0039) Hit@20 0.9416 (+0.0058)
2
+ macro Hit@5 0.9365 (+0.0015 [-0.0012, +0.0041]) pools +12 -7
3
+ macro Hit@20 0.9675 (-0.0012 [-0.0035, +0.0015]) pools +7 -11
4
+ domain 0.9329 (+0.0028) other 0.6209 (+0.0147) | Orbitrap no energy 0.5867 (+0.0170) | above 850 0.4514 (+0.0297) below 850 0.8514 (+0.0040)
5
+ full rule : Hit@1 0.8093 (ref 0.8063; +0.0030 [-0.0084, +0.0129]) macro 0.8842 (+0.0030 [+0.0003, +0.0056]) pools +18 -8 cap30 +0.0046 | Hit@5 0.8821 (+0.0063) Hit@20 0.9086 (-0.0082)
6
+ macro Hit@5 0.9424 (+0.0012 [-0.0012, +0.0035]) pools +11 -7
7
+ macro Hit@20 0.9675 (-0.0021 [-0.0041, +0.0000]) pools +3 -10
8
+ domain 0.9442 (+0.0033) other 0.6737 (+0.0027) | Orbitrap no energy 0.6444 (+0.0020) | above 850 0.5286 (-0.0153) below 850 0.8732 (+0.0071)
shared/package-files/ef/efbf894e84d19de3b41b6f12aa3ad68b9e2229115175b9403d0f89821c7b034d/eval_blocks.py ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Pooled inverse distances per fingerprint block for a multi-family target (lab, E56; labels: evaluation only).
3
+
4
+ eval_blocks.py CHECKPOINT POOLS.npz CAND_FPS.npy OUT.pkl [--block 2048]
5
+
6
+ For every pool the predicted bit probabilities of its spectra are averaged per candidate formula
7
+ (as in the released `pooled` inverse distance). Saved per query (candidate order):
8
+ all cosine distance over all bits
9
+ b0..bK cosine distance inside block k
10
+ mean mean of the block cosine distances
11
+ zmean mean of the block distances after standardizing each block over the pool's supported candidates
12
+ Prints the pool-rule Hit@1 of every variant alone.
13
+ """
14
+ import argparse
15
+ import pickle
16
+ import sys
17
+ import time
18
+ from pathlib import Path
19
+
20
+ import numpy as np
21
+ import torch
22
+
23
+ HERE = Path(__file__).resolve().parent
24
+ sys.path.insert(0, str(HERE.parent / 'pkg'))
25
+ sys.path.insert(0, str(HERE.parent / 'analysis'))
26
+ import inverse
27
+ import invrun
28
+
29
+
30
+ def cosine(prob, bits):
31
+ dot = (prob * bits).sum(-1)
32
+ return 1 - dot / (prob.norm(dim=-1).clamp(min=1e-6) * bits.norm(dim=-1).clamp(min=1e-6))
33
+
34
+
35
+ def main():
36
+ p = argparse.ArgumentParser(description=__doc__)
37
+ p.add_argument('checkpoint', type=Path)
38
+ p.add_argument('pools', type=Path)
39
+ p.add_argument('cand_fps', type=Path)
40
+ p.add_argument('out', type=Path)
41
+ p.add_argument('--block', type=int, default=2048)
42
+ p.add_argument('--batch', type=int, default=2048)
43
+ args = p.parse_args()
44
+ device = torch.device('cuda')
45
+ model = invrun.load_model(args.checkpoint, device)
46
+ store = invrun.PoolStore(args.pools, device)
47
+ store.cand_fps = torch.from_numpy(np.load(args.cand_fps)).to(device)
48
+ bits_total = store.cand_fps.shape[1] * 8
49
+ blocks = [slice(start, start + args.block) for start in range(0, bits_total, args.block)]
50
+ names = ['all'] + [f'b{k}' for k in range(len(blocks))] + ['mean', 'zmean']
51
+ out = {name: [None] * store.queries for name in names}
52
+ tick = time.monotonic()
53
+ with torch.no_grad():
54
+ for group in store.groups(None, 6000):
55
+ low, high = int(store.pair_start[group[0]]), int(store.pair_start[group[-1] + 1])
56
+ logits = store.logits(model, low, high, batch=args.batch)
57
+ for s in group:
58
+ nq, nf = len(store.members[s]), int(store.formula_start[s + 1] - store.formula_start[s])
59
+ if nq == 0 or nf == 0:
60
+ continue
61
+ part = logits[int(store.pair_start[s]) - low:int(store.pair_start[s + 1]) - low].view(nq, nf, -1)
62
+ prob = torch.sigmoid(part).mean(0) # [F, bits]
63
+ cs = slice(int(store.cand_start[s]), int(store.cand_start[s + 1]))
64
+ which = store.cand_formula[cs]
65
+ supported = which >= 0
66
+ cand = inverse.unpack_bits(store.cand_fps[cs])
67
+ pred = prob[which.clamp(min=0)] # [C, bits]
68
+ result = {'all': cosine(pred, cand)}
69
+ per = []
70
+ for k, block in enumerate(blocks):
71
+ d = cosine(pred[:, block], cand[:, block])
72
+ result[f'b{k}'] = d
73
+ per.append(d)
74
+ per = torch.stack(per) # [K, C]
75
+ result['mean'] = per.mean(0)
76
+ if supported.sum() > 1:
77
+ sub = per[:, supported]
78
+ z = (per - sub.mean(1, keepdim=True)) / sub.std(1, keepdim=True).clamp(min=1e-6)
79
+ else:
80
+ z = per
81
+ result['zmean'] = z.mean(0)
82
+ for name, d in result.items():
83
+ worst = d[supported].max() + 1 if supported.any() else torch.tensor(1.0, device=device)
84
+ d = torch.where(supported, d, worst.expand_as(d)).cpu().numpy()
85
+ for qi in store.members[s]:
86
+ out[name][qi] = d
87
+ print(f'{store.queries} queries, {time.monotonic() - tick:.1f} s')
88
+ with args.out.open('wb') as stream:
89
+ pickle.dump(out, stream)
90
+ import scorelab
91
+ import infer
92
+ root = Path('/workspace/research/validation-input')
93
+ queries = list(infer.read_jsonl(root / 'queries.jsonl'))
94
+ pools = {r['candidate_pool_id']: r['candidates'] for r in infer.read_jsonl(root / 'candidate_pools.jsonl')}
95
+ ev = scorelab.Evaluator(queries, pools)
96
+ for name in names:
97
+ rank = ev.ranks(out[name], 1.0)
98
+ hit = rank == 0
99
+ first = np.array([m[0] for m in ev.by_pool.values()])
100
+ print(f'{name:6s} pool rule all={hit.mean():.4f} dom={hit[ev.domain].mean():.4f} non={hit[~ev.domain].mean():.4f} '
101
+ f'h5={np.mean(rank < 5):.4f} macro={hit[first].mean():.4f}', flush=True)
102
+
103
+
104
+ if __name__ == '__main__':
105
+ main()
shared/package-files/ef/efed6ca2099a4feb065029ab8f1e0c6676887395b891b8cfc8858597e71ca195/eval-score-I16.log ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 19429 queries, 1 members, 52.3 s
2
+ cos pool rule all=0.6007 dom=0.8360 non=0.3644 h5=0.7108 macro=0.7185
3
+ cen pool rule all=0.5839 dom=0.8146 non=0.3522 h5=0.7007 macro=0.6949
4
+ cos_pm pool rule all=0.6007 dom=0.8360 non=0.3644 h5=0.7108 macro=0.7185
5
+ dot pool rule all=0.5949 dom=0.8070 non=0.3819 h5=0.7083 macro=0.7070
6
+ geo pool rule all=0.5895 dom=0.8284 non=0.3497 h5=0.7005 macro=0.7123
7
+ idf05 pool rule all=0.6092 dom=0.8411 non=0.3765 h5=0.7216 macro=0.7242
8
+ idf1 pool rule all=0.6106 dom=0.8398 non=0.3804 h5=0.7285 macro=0.7209
9
+ lin pool rule all=0.5586 dom=0.7924 non=0.3240 h5=0.6765 macro=0.6810
10
+ log01 pool rule all=0.6051 dom=0.8307 non=0.3786 h5=0.7344 macro=0.7144
11
+ log05 pool rule all=0.6179 dom=0.8467 non=0.3882 h5=0.7306 macro=0.7277
12
+ log20 pool rule all=0.6132 dom=0.8446 non=0.3809 h5=0.7235 macro=0.7259
13
+ nll02 pool rule all=0.4443 dom=0.7104 non=0.1771 h5=0.5842 macro=0.5836
14
+ nll10 pool rule all=0.4087 dom=0.6546 non=0.1618 h5=0.5515 macro=0.5340
15
+ poolcen pool rule all=0.5770 dom=0.7945 non=0.3585 h5=0.7126 macro=0.6902
16
+ pow020 pool rule all=0.5642 dom=0.7923 non=0.3351 h5=0.7152 macro=0.6828
17
+ pow030 pool rule all=0.5885 dom=0.8126 non=0.3634 h5=0.7295 macro=0.6991
18
+ pow040 pool rule all=0.6005 dom=0.8261 non=0.3740 h5=0.7326 macro=0.7120
19
+ pow05 pool rule all=0.6047 dom=0.8346 non=0.3738 h5=0.7288 macro=0.7212
20
+ pow05idf025 pool rule all=0.6049 dom=0.8362 non=0.3727 h5=0.7351 macro=0.7215
21
+ pow05idf05 pool rule all=0.6063 dom=0.8406 non=0.3711 h5=0.7366 macro=0.7245
22
+ pow05m pool rule all=0.6047 dom=0.8346 non=0.3738 h5=0.7288 macro=0.7212
23
+ pow05s pool rule all=0.6046 dom=0.8407 non=0.3675 h5=0.7241 macro=0.7215
24
+ pow060 pool rule all=0.6123 dom=0.8425 non=0.3812 h5=0.7254 macro=0.7253
25
+ pow060idf025 pool rule all=0.6118 dom=0.8466 non=0.3761 h5=0.7277 macro=0.7283
26
+ pow060idf050 pool rule all=0.6162 dom=0.8475 non=0.3840 h5=0.7312 macro=0.7274
27
+ pow060m pool rule all=0.6123 dom=0.8425 non=0.3812 h5=0.7254 macro=0.7253
28
+ pow065 pool rule all=0.6121 dom=0.8444 non=0.3789 h5=0.7224 macro=0.7277
29
+ pow070 pool rule all=0.6140 dom=0.8454 non=0.3817 h5=0.7203 macro=0.7289
30
+ pow070idf025 pool rule all=0.6149 dom=0.8492 non=0.3796 h5=0.7236 macro=0.7307
31
+ pow070idf050 pool rule all=0.6124 dom=0.8486 non=0.3752 h5=0.7289 macro=0.7283
32
+ pow070m pool rule all=0.6140 dom=0.8454 non=0.3817 h5=0.7203 macro=0.7289
33
+ pow075 pool rule all=0.6132 dom=0.8448 non=0.3806 h5=0.7204 macro=0.7271
34
+ pow080 pool rule all=0.6121 dom=0.8446 non=0.3786 h5=0.7180 macro=0.7259
35
+ pow080m pool rule all=0.6121 dom=0.8446 non=0.3786 h5=0.7180 macro=0.7259
36
+ pow085 pool rule all=0.6095 dom=0.8414 non=0.3767 h5=0.7152 macro=0.7242
37
+ pow085idf025 pool rule all=0.6109 dom=0.8437 non=0.3771 h5=0.7241 macro=0.7262
38
+ pow085idf050 pool rule all=0.6111 dom=0.8437 non=0.3775 h5=0.7242 macro=0.7259
39
+ pow090 pool rule all=0.6091 dom=0.8404 non=0.3770 h5=0.7145 macro=0.7233
40
+ pow2 pool rule all=0.5739 dom=0.8078 non=0.3391 h5=0.6888 macro=0.6920
41
+ tani pool rule all=0.5812 dom=0.8111 non=0.3504 h5=0.7130 macro=0.7014