Upload SCALE checkpoints for T01–T09

#1
Files changed (36) hide show
  1. .gitattributes +19 -0
  2. README.md +21 -0
  3. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/context/model.yaml +198 -0
  4. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/.metadata +3 -0
  5. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_0.distcp +3 -0
  6. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_1.distcp +3 -0
  7. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/common.pt +3 -0
  8. checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/metadata.json +1 -0
  9. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/context/model.yaml +200 -0
  10. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/.metadata +3 -0
  11. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_0.distcp +3 -0
  12. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_1.distcp +3 -0
  13. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_0.distcp +3 -0
  14. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_1.distcp +3 -0
  15. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/common.pt +3 -0
  16. checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/metadata.json +1 -0
  17. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/context/model.yaml +200 -0
  18. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/.metadata +3 -0
  19. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_0.distcp +3 -0
  20. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_1.distcp +3 -0
  21. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_0.distcp +3 -0
  22. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_1.distcp +3 -0
  23. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/common.pt +3 -0
  24. checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/metadata.json +1 -0
  25. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/context/model.yaml +188 -0
  26. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/.metadata +3 -0
  27. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_0.distcp +3 -0
  28. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_1.distcp +3 -0
  29. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/common.pt +3 -0
  30. checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/metadata.json +1 -0
  31. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/context/model.yaml +216 -0
  32. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/.metadata +3 -0
  33. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_0.distcp +3 -0
  34. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_1.distcp +3 -0
  35. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/common.pt +3 -0
  36. checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/metadata.json +1 -0
.gitattributes CHANGED
@@ -33,3 +33,22 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/.metadata filter=lfs diff=lfs merge=lfs -text
37
+ checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
38
+ checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_1.distcp filter=lfs diff=lfs merge=lfs -text
39
+ checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/.metadata filter=lfs diff=lfs merge=lfs -text
40
+ checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
41
+ checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_1.distcp filter=lfs diff=lfs merge=lfs -text
42
+ checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/.metadata filter=lfs diff=lfs merge=lfs -text
43
+ checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
44
+ checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_1.distcp filter=lfs diff=lfs merge=lfs -text
45
+ checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/.metadata filter=lfs diff=lfs merge=lfs -text
46
+ checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
47
+ checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_1.distcp filter=lfs diff=lfs merge=lfs -text
48
+ checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_0.distcp filter=lfs diff=lfs merge=lfs -text
49
+ checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_1.distcp filter=lfs diff=lfs merge=lfs -text
50
+ checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/.metadata filter=lfs diff=lfs merge=lfs -text
51
+ checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
52
+ checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_1.distcp filter=lfs diff=lfs merge=lfs -text
53
+ checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_0.distcp filter=lfs diff=lfs merge=lfs -text
54
+ checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_1.distcp filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: nemo
3
+ tags:
4
+ - single-cell
5
+ - perturbation
6
+ ---
7
+
8
+ # SCALE checkpoints
9
+
10
+ These checkpoints accompany [SCALE: Scalable Conditional Atlas-Level Endpoint Transport for virtual cell perturbation prediction](https://arxiv.org/abs/2603.17380v3).
11
+
12
+ This upload covers T01–T09. T02 includes all five PBMC folds. T09 includes
13
+ `patient_1015_val_patient_101` for single-fold inference and evaluation.
14
+ The paper's Kang aggregation requires all 56 folds.
15
+
16
+ Use the YAML and commands supplied with the SCALE source repository. Keep
17
+ the `weights/` directories intact, including the distributed shards,
18
+ `common.pt`, and `.metadata`. The tensor shards are copied from the selected
19
+ runs. Server paths saved by checkpoint callbacks were replaced in this copy.
20
+ The experiment YAMLs provide the data paths. Prepared datasets and evaluator
21
+ inputs are separate.
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/context/model.yaml ADDED
@@ -0,0 +1,198 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ checkpoint_path: checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last
2
+ _target_: models.JIT_FP.nemo_model.JIT_FP_Nemo_Model
3
+ conf:
4
+ Interpolant:
5
+ prior_distribution: gaussian
6
+ prior_kwargs:
7
+ logits: 0.1
8
+ total_count: 1
9
+ zi_logits: 0
10
+ probability_path:
11
+ constant_noise: 0.0
12
+ cell_sentence_len: 512
13
+ cellmap_width: 512
14
+ dataset:
15
+ batch_size: 1
16
+ cell_sentence_len: 512
17
+ control_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/control_train
18
+ control_pert: '[(''control'', 0.0, ''uM'')]'
19
+ fingerprint_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/drug_fingerprints.pkl
20
+ global_keys_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/global_keys.pkl
21
+ num_workers: 4
22
+ random_seed: 42
23
+ test_control_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/control_test
24
+ test_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/test
25
+ train_control_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/control_train
26
+ train_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/train
27
+ val_control_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/control_val
28
+ val_lmdb_path: resources/nemo_cellflow/analyzing/task04_combo_drug/data_processed/lmdb_by_split.backup_before_unseencombo_fix_20260609_1920/split_seen_single_unseen_combo/val
29
+ dataset_name: tahoe
30
+ experiment:
31
+ accumulate_grad_batches: 1
32
+ check_val_every_n_epoch: 1
33
+ ckpt_freq: 50
34
+ group: aivc-combosciplex-jit-fp-split_seen_single_unseen_combo-dit12-1gpu-bs1
35
+ monitor: val_loss
36
+ num_epoch: 500
37
+ num_gpus: 1
38
+ num_nodes: 1
39
+ num_sanity_val_steps: 2
40
+ pipeline_model_parallel_size: 1
41
+ precision: 32
42
+ seed: 42
43
+ strategy: ddp
44
+ tensor_model_parallel_size: 1
45
+ time_step: 201
46
+ use_wandb: true
47
+ layers_after_pool:
48
+ act_fn: SiLU
49
+ act_last_layer: true
50
+ dims:
51
+ - 512
52
+ - 512
53
+ dropout_rate: 0.0
54
+ input_dim: -1
55
+ layer_type: mlp
56
+ layers_before_pool:
57
+ batch_onehot:
58
+ dims:
59
+ - 256
60
+ - 256
61
+ dropout_rate: 0.0
62
+ input_dim: 256
63
+ layer_type: embedding
64
+ num_embeddings: 1
65
+ cell_type_onehot:
66
+ dims:
67
+ - 256
68
+ - 256
69
+ dropout_rate: 0.0
70
+ input_dim: 256
71
+ layer_type: embedding
72
+ num_embeddings: 1
73
+ pert_fingerprint:
74
+ dims:
75
+ - 256
76
+ - 256
77
+ dropout_rate: 0.0
78
+ fp_dim: 2048
79
+ fp_mode: concat
80
+ input_dim: 256
81
+ layer_type: film_continuous
82
+ layers_for_aggregation:
83
+ act_fn: SiLU
84
+ act_last_layer: true
85
+ dims:
86
+ - 1024
87
+ - 512
88
+ dropout_rate: 0.0
89
+ input_dim: 1024
90
+ layer_type: mlp
91
+ model_name: JIT_FP
92
+ prepare_model:
93
+ batch_class: 1
94
+ cell_class: 1
95
+ cell_dim: 2000
96
+ cell_sentence_len: 512
97
+ cellmap_decoder_kwargs:
98
+ act_fn: SiLU
99
+ dims:
100
+ - 512
101
+ - 1024
102
+ - 2000
103
+ dropout_rate: 0.0
104
+ gene_decoder_type: mlp
105
+ cellmap_encoder_kwargs:
106
+ act_fn: SiLU
107
+ dims:
108
+ - 2000
109
+ - 1024
110
+ - 512
111
+ dropout_rate: 0.0
112
+ gene_encoder_type: mlp
113
+ cellmap_width: 512
114
+ cond_output_dropout: 0.9
115
+ condition_embedding_dim: 256
116
+ condition_encoder_kwargs: {}
117
+ condition_mode: deterministic
118
+ conditioning: concatenation
119
+ dit_depth: 12
120
+ dit_kwargs:
121
+ mlp_ratio: 4.0
122
+ num_heads: 16
123
+ hidden_channels: 32
124
+ hidden_dims:
125
+ - 2048
126
+ - 2048
127
+ - 2048
128
+ hidden_dropout: 0.0
129
+ layer_norm_before_concatenation: false
130
+ layers_after_pool:
131
+ act_fn: SiLU
132
+ act_last_layer: true
133
+ dims:
134
+ - 512
135
+ - 512
136
+ dropout_rate: 0.0
137
+ input_dim: -1
138
+ layer_type: mlp
139
+ layers_before_pool:
140
+ batch_onehot:
141
+ dims:
142
+ - 256
143
+ - 256
144
+ dropout_rate: 0.0
145
+ input_dim: 256
146
+ layer_type: embedding
147
+ num_embeddings: 1
148
+ cell_type_onehot:
149
+ dims:
150
+ - 256
151
+ - 256
152
+ dropout_rate: 0.0
153
+ input_dim: 256
154
+ layer_type: embedding
155
+ num_embeddings: 1
156
+ pert_fingerprint:
157
+ dims:
158
+ - 256
159
+ - 256
160
+ dropout_rate: 0.0
161
+ fp_dim: 2048
162
+ fp_mode: concat
163
+ input_dim: 256
164
+ layer_type: film_continuous
165
+ layers_for_aggregation:
166
+ act_fn: SiLU
167
+ act_last_layer: true
168
+ dims:
169
+ - 1024
170
+ - 512
171
+ dropout_rate: 0.0
172
+ input_dim: 1024
173
+ layer_type: mlp
174
+ learning_rate: 0.001
175
+ linear_projection_before_concatenation: false
176
+ loss_type: x
177
+ match_fn: match_linear
178
+ megatron_ckpt: null
179
+ pool_sample_covariates: true
180
+ pooling: mean
181
+ pooling_kwargs:
182
+ num_heads: 8
183
+ pool_method: mean
184
+ pred_type: x
185
+ regularization: 0.0
186
+ seed: 0
187
+ solver_kwargs: {}
188
+ src_loss_weight: 1.0
189
+ time_encoder_dims:
190
+ - 1024
191
+ - 512
192
+ time_encoder_dropout: 0.0
193
+ time_freqs: 512
194
+ time_max_period: 10000
195
+ transformer_backbone_kwargs:
196
+ hidden_size: 512
197
+ configure_init_model_parallel: false
198
+ model_transform: null
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/.metadata ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05b20cc67039570f61fea632a92c6487587c480454749532c5f867994478c1ae
3
+ size 232979
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad1ae82d8493f0bdd80311c47a88dd3bb41686d0886c6a378a2ff5ccc2ce65b0
3
+ size 437781932
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/__0_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae20b2f098614ee3d88c9163f5104e071bdd4d4070ed4716a5aa4e174a62b770
3
+ size 437784292
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/common.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bebea1a78be638f256742fa7f0db0a7f6510389faf00c7ec4841aa794588ed2f
3
+ size 4106
checkpoints/combosciplex_unseen_combination/finetune_pearson_best-epoch=500-last/weights/metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sharded_backend": "torch_dist", "sharded_backend_version": 1, "common_backend": "torch", "common_backend_version": 1}
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/context/model.yaml ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ checkpoint_path: checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449
2
+ _target_: models.JIT_FP.nemo_model.JIT_FP_Nemo_Model
3
+ conf:
4
+ Interpolant:
5
+ prior_distribution: gaussian
6
+ prior_kwargs:
7
+ logits: 0.1
8
+ total_count: 1
9
+ zi_logits: 0
10
+ probability_path:
11
+ constant_noise: 0.0
12
+ cell_sentence_len: 512
13
+ cellmap_width: 512
14
+ dataset:
15
+ batch_size: 16
16
+ cell_sentence_len: 512
17
+ control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/control_train
18
+ control_pert: '[(''control'', 0.0, ''uM'')]'
19
+ fingerprint_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/drug_fingerprints_full.pkl
20
+ global_keys_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/global_keys.pkl
21
+ num_workers: 8
22
+ random_seed: 42
23
+ test_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/control_test
24
+ test_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/test
25
+ train_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/control_train
26
+ train_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/train
27
+ val_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/control_val
28
+ val_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_cell_line_MCF7/val
29
+ dataset_name: tahoe
30
+ experiment:
31
+ accumulate_grad_batches: 1
32
+ check_val_every_n_epoch: 5
33
+ ckpt_freq: 50
34
+ group: sciplex3-jit-fp-split_unseen_cell_line_MCF7
35
+ monitor: val_loss
36
+ num_epoch: 500
37
+ num_gpus: 2
38
+ num_nodes: 1
39
+ num_sanity_val_steps: 2
40
+ pipeline_model_parallel_size: 1
41
+ precision: 32
42
+ seed: 42
43
+ strategy: ddp
44
+ tensor_model_parallel_size: 1
45
+ time_step: 201
46
+ use_wandb: true
47
+ implementation_note: 2GPU Task03 split training config. Tahoe_Dataset fixed-celltype
48
+ sentence sampling; full fingerprints; split-specific controls.
49
+ layers_after_pool:
50
+ act_fn: SiLU
51
+ act_last_layer: true
52
+ dims:
53
+ - 512
54
+ - 512
55
+ dropout_rate: 0.0
56
+ input_dim: -1
57
+ layer_type: mlp
58
+ layers_before_pool:
59
+ batch_onehot:
60
+ dims:
61
+ - 256
62
+ - 256
63
+ dropout_rate: 0.0
64
+ input_dim: 256
65
+ layer_type: embedding
66
+ num_embeddings: 5
67
+ cell_type_onehot:
68
+ dims:
69
+ - 256
70
+ - 256
71
+ dropout_rate: 0.0
72
+ input_dim: 256
73
+ layer_type: embedding
74
+ num_embeddings: 3
75
+ pert_fingerprint:
76
+ dims:
77
+ - 256
78
+ - 256
79
+ dropout_rate: 0.0
80
+ fp_dim: 2048
81
+ fp_mode: concat
82
+ input_dim: 256
83
+ layer_type: film_continuous
84
+ layers_for_aggregation:
85
+ act_fn: SiLU
86
+ act_last_layer: true
87
+ dims:
88
+ - 1024
89
+ - 512
90
+ dropout_rate: 0.0
91
+ input_dim: 1024
92
+ layer_type: mlp
93
+ model_name: JIT_FP
94
+ prepare_model:
95
+ batch_class: 5
96
+ cell_class: 3
97
+ cell_dim: 2000
98
+ cell_sentence_len: 512
99
+ cellmap_decoder_kwargs:
100
+ act_fn: SiLU
101
+ dims:
102
+ - 512
103
+ - 1024
104
+ - 2000
105
+ dropout_rate: 0.0
106
+ gene_decoder_type: mlp
107
+ cellmap_encoder_kwargs:
108
+ act_fn: SiLU
109
+ dims:
110
+ - 2000
111
+ - 1024
112
+ - 512
113
+ dropout_rate: 0.0
114
+ gene_encoder_type: mlp
115
+ cellmap_width: 512
116
+ cond_output_dropout: 0.9
117
+ condition_embedding_dim: 256
118
+ condition_encoder_kwargs: {}
119
+ condition_mode: deterministic
120
+ conditioning: concatenation
121
+ dit_depth: 18
122
+ dit_kwargs:
123
+ mlp_ratio: 4.0
124
+ num_heads: 16
125
+ hidden_channels: 32
126
+ hidden_dims:
127
+ - 2048
128
+ - 2048
129
+ - 2048
130
+ hidden_dropout: 0.0
131
+ layer_norm_before_concatenation: false
132
+ layers_after_pool:
133
+ act_fn: SiLU
134
+ act_last_layer: true
135
+ dims:
136
+ - 512
137
+ - 512
138
+ dropout_rate: 0.0
139
+ input_dim: -1
140
+ layer_type: mlp
141
+ layers_before_pool:
142
+ batch_onehot:
143
+ dims:
144
+ - 256
145
+ - 256
146
+ dropout_rate: 0.0
147
+ input_dim: 256
148
+ layer_type: embedding
149
+ num_embeddings: 5
150
+ cell_type_onehot:
151
+ dims:
152
+ - 256
153
+ - 256
154
+ dropout_rate: 0.0
155
+ input_dim: 256
156
+ layer_type: embedding
157
+ num_embeddings: 3
158
+ pert_fingerprint:
159
+ dims:
160
+ - 256
161
+ - 256
162
+ dropout_rate: 0.0
163
+ fp_dim: 2048
164
+ fp_mode: concat
165
+ input_dim: 256
166
+ layer_type: film_continuous
167
+ layers_for_aggregation:
168
+ act_fn: SiLU
169
+ act_last_layer: true
170
+ dims:
171
+ - 1024
172
+ - 512
173
+ dropout_rate: 0.0
174
+ input_dim: 1024
175
+ layer_type: mlp
176
+ learning_rate: 0.001
177
+ linear_projection_before_concatenation: false
178
+ loss_type: x
179
+ match_fn: match_linear
180
+ megatron_ckpt: null
181
+ pool_sample_covariates: true
182
+ pooling: mean
183
+ pooling_kwargs:
184
+ num_heads: 8
185
+ pool_method: mean
186
+ pred_type: x
187
+ regularization: 0.0
188
+ seed: 0
189
+ solver_kwargs: {}
190
+ src_loss_weight: 1.0
191
+ time_encoder_dims:
192
+ - 1024
193
+ - 512
194
+ time_encoder_dropout: 0.0
195
+ time_freqs: 512
196
+ time_max_period: 10000
197
+ transformer_backbone_kwargs:
198
+ hidden_size: 512
199
+ configure_init_model_parallel: false
200
+ model_transform: null
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/.metadata ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:071c0a887d8827f27b843ec51f41b0db783866c7a5373260bc0a6fee691c1c23
3
+ size 307463
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b6d7d6b4a8132822257bb8ee94de056bbc1ac4fed07a7f63f3b2e923b6f07b64
3
+ size 304017444
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__0_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d6bc08098f6bb0cae85a299a5edf62d7c64a4ea6c0dec1872fcdd01da87b136
3
+ size 304031540
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3b0783e0c195674f811fee90883547aac751091d7f701c4762c607d1d05932e5
3
+ size 304014032
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/__1_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4adcdcbd0ee5f80e4e330bb3932055a6d34424dd9fb6f90ff77ec4aef9b794c
3
+ size 304030232
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/common.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:000116a9e276e6888bc15859cf02233b27d5f72c20e1ccf841f04f2cc2c540be
3
+ size 4110
checkpoints/sciplex3_unseen_cell_line/finetune_pearson_best-epoch=449/weights/metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sharded_backend": "torch_dist", "sharded_backend_version": 1, "common_backend": "torch", "common_backend_version": 1}
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/context/model.yaml ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ checkpoint_path: checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404
2
+ _target_: models.JIT_FP.nemo_model.JIT_FP_Nemo_Model
3
+ conf:
4
+ Interpolant:
5
+ prior_distribution: gaussian
6
+ prior_kwargs:
7
+ logits: 0.1
8
+ total_count: 1
9
+ zi_logits: 0
10
+ probability_path:
11
+ constant_noise: 0.0
12
+ cell_sentence_len: 512
13
+ cellmap_width: 512
14
+ dataset:
15
+ batch_size: 16
16
+ cell_sentence_len: 512
17
+ control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/control_train
18
+ control_pert: '[(''control'', 0.0, ''uM'')]'
19
+ fingerprint_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/drug_fingerprints_full.pkl
20
+ global_keys_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/global_keys.pkl
21
+ num_workers: 8
22
+ random_seed: 42
23
+ test_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/control_test
24
+ test_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/test
25
+ train_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/control_train
26
+ train_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/train
27
+ val_control_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/control_val
28
+ val_lmdb_path: resources/nemo_cellflow/analyzing/task03_single_drug_dose/data_processed/lmdb_by_split/split_unseen_dose_extrapolate_10nM_100nM_to_10uM/val
29
+ dataset_name: tahoe
30
+ experiment:
31
+ accumulate_grad_batches: 1
32
+ check_val_every_n_epoch: 5
33
+ ckpt_freq: 50
34
+ group: sciplex3-jit-fp-split_unseen_dose_extrapolate_10nM_100nM_to_10uM
35
+ monitor: val_loss
36
+ num_epoch: 500
37
+ num_gpus: 2
38
+ num_nodes: 1
39
+ num_sanity_val_steps: 2
40
+ pipeline_model_parallel_size: 1
41
+ precision: 32
42
+ seed: 42
43
+ strategy: ddp
44
+ tensor_model_parallel_size: 1
45
+ time_step: 201
46
+ use_wandb: true
47
+ implementation_note: 2GPU Task03 split training config. Tahoe_Dataset fixed-celltype
48
+ sentence sampling; full fingerprints; split-specific controls.
49
+ layers_after_pool:
50
+ act_fn: SiLU
51
+ act_last_layer: true
52
+ dims:
53
+ - 512
54
+ - 512
55
+ dropout_rate: 0.0
56
+ input_dim: -1
57
+ layer_type: mlp
58
+ layers_before_pool:
59
+ batch_onehot:
60
+ dims:
61
+ - 256
62
+ - 256
63
+ dropout_rate: 0.0
64
+ input_dim: 256
65
+ layer_type: embedding
66
+ num_embeddings: 5
67
+ cell_type_onehot:
68
+ dims:
69
+ - 256
70
+ - 256
71
+ dropout_rate: 0.0
72
+ input_dim: 256
73
+ layer_type: embedding
74
+ num_embeddings: 3
75
+ pert_fingerprint:
76
+ dims:
77
+ - 256
78
+ - 256
79
+ dropout_rate: 0.0
80
+ fp_dim: 2048
81
+ fp_mode: concat
82
+ input_dim: 256
83
+ layer_type: film_continuous
84
+ layers_for_aggregation:
85
+ act_fn: SiLU
86
+ act_last_layer: true
87
+ dims:
88
+ - 1024
89
+ - 512
90
+ dropout_rate: 0.0
91
+ input_dim: 1024
92
+ layer_type: mlp
93
+ model_name: JIT_FP
94
+ prepare_model:
95
+ batch_class: 5
96
+ cell_class: 3
97
+ cell_dim: 2000
98
+ cell_sentence_len: 512
99
+ cellmap_decoder_kwargs:
100
+ act_fn: SiLU
101
+ dims:
102
+ - 512
103
+ - 1024
104
+ - 2000
105
+ dropout_rate: 0.0
106
+ gene_decoder_type: mlp
107
+ cellmap_encoder_kwargs:
108
+ act_fn: SiLU
109
+ dims:
110
+ - 2000
111
+ - 1024
112
+ - 512
113
+ dropout_rate: 0.0
114
+ gene_encoder_type: mlp
115
+ cellmap_width: 512
116
+ cond_output_dropout: 0.9
117
+ condition_embedding_dim: 256
118
+ condition_encoder_kwargs: {}
119
+ condition_mode: deterministic
120
+ conditioning: concatenation
121
+ dit_depth: 18
122
+ dit_kwargs:
123
+ mlp_ratio: 4.0
124
+ num_heads: 16
125
+ hidden_channels: 32
126
+ hidden_dims:
127
+ - 2048
128
+ - 2048
129
+ - 2048
130
+ hidden_dropout: 0.0
131
+ layer_norm_before_concatenation: false
132
+ layers_after_pool:
133
+ act_fn: SiLU
134
+ act_last_layer: true
135
+ dims:
136
+ - 512
137
+ - 512
138
+ dropout_rate: 0.0
139
+ input_dim: -1
140
+ layer_type: mlp
141
+ layers_before_pool:
142
+ batch_onehot:
143
+ dims:
144
+ - 256
145
+ - 256
146
+ dropout_rate: 0.0
147
+ input_dim: 256
148
+ layer_type: embedding
149
+ num_embeddings: 5
150
+ cell_type_onehot:
151
+ dims:
152
+ - 256
153
+ - 256
154
+ dropout_rate: 0.0
155
+ input_dim: 256
156
+ layer_type: embedding
157
+ num_embeddings: 3
158
+ pert_fingerprint:
159
+ dims:
160
+ - 256
161
+ - 256
162
+ dropout_rate: 0.0
163
+ fp_dim: 2048
164
+ fp_mode: concat
165
+ input_dim: 256
166
+ layer_type: film_continuous
167
+ layers_for_aggregation:
168
+ act_fn: SiLU
169
+ act_last_layer: true
170
+ dims:
171
+ - 1024
172
+ - 512
173
+ dropout_rate: 0.0
174
+ input_dim: 1024
175
+ layer_type: mlp
176
+ learning_rate: 0.001
177
+ linear_projection_before_concatenation: false
178
+ loss_type: x
179
+ match_fn: match_linear
180
+ megatron_ckpt: null
181
+ pool_sample_covariates: true
182
+ pooling: mean
183
+ pooling_kwargs:
184
+ num_heads: 8
185
+ pool_method: mean
186
+ pred_type: x
187
+ regularization: 0.0
188
+ seed: 0
189
+ solver_kwargs: {}
190
+ src_loss_weight: 1.0
191
+ time_encoder_dims:
192
+ - 1024
193
+ - 512
194
+ time_encoder_dropout: 0.0
195
+ time_freqs: 512
196
+ time_max_period: 10000
197
+ transformer_backbone_kwargs:
198
+ hidden_size: 512
199
+ configure_init_model_parallel: false
200
+ model_transform: null
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/.metadata ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9b4a81360c37e385d4f60ccaa2b9d5b7470f8803b44d9244d42377643bb83021
3
+ size 307484
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa550d89c446263281652c493e35126a48c3da4795ba8951e98b8e37088ba9be
3
+ size 304017444
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__0_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da6074389e009918fcf77afd40674bfe14f4c5024bfe838b4b46d35893e43517
3
+ size 304031540
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4771c30893232d2205fcaf21e9f22a754fdd9576015db7d20c1f28bf18485daf
3
+ size 304014032
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/__1_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f62b541f0a52986b8db12cdeab84afd892d97f7389c73cf6b70e7b80c1895197
3
+ size 304030232
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/common.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9f00f5bbd4374d4d91abb7a603eeaeab237cdd203c730b4ffe484b06bb9c8e84
3
+ size 4110
checkpoints/sciplex3_unseen_dose/finetune_pearson_best-epoch=404/weights/metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sharded_backend": "torch_dist", "sharded_backend_version": 1, "common_backend": "torch", "common_backend_version": 1}
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/context/model.yaml ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ checkpoint_path: checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29
2
+ _target_: models.JIT_LLMSET.nemo_model.JIT_LLMSET_Nemo_Model
3
+ conf:
4
+ Interpolant:
5
+ prior_distribution: gaussian
6
+ prior_kwargs:
7
+ logits: 0.1
8
+ total_count: 1
9
+ zi_logits: 0
10
+ probability_path:
11
+ constant_noise: 0.0
12
+ cell_sentence_len: 32
13
+ cellmap_width: 512
14
+ dataset:
15
+ batch_size: 16
16
+ cell_sentence_len: 32
17
+ control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/control_train
18
+ control_pert: ctrl-inj
19
+ global_keys_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/global_keys.pkl
20
+ num_workers: 4
21
+ random_seed: 42
22
+ source_h5ad_path: resources/datasets/GSE202639_zscape_saunders_2023/processed/h5ad/GSE202639_zperturb_full_raw_counts.h5ad
23
+ test_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/control_test
24
+ test_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/test
25
+ train_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/control_train
26
+ train_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/train
27
+ val_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/control_val
28
+ val_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_ctrlinj_seen_gene_unseen_cells_v1/lmdb/val
29
+ dataset_name: tahoe
30
+ experiment:
31
+ accumulate_grad_batches: 1
32
+ check_val_every_n_epoch: 10
33
+ ckpt_freq: 50
34
+ group: zscape-jitllmset-r2-r2_lr2e-4_drop05_src025_xx
35
+ monitor: val_loss
36
+ num_epoch: 30
37
+ num_gpus: 1
38
+ num_nodes: 1
39
+ num_sanity_val_steps: 2
40
+ pipeline_model_parallel_size: 1
41
+ precision: 32
42
+ seed: 42
43
+ strategy: ddp
44
+ tensor_model_parallel_size: 1
45
+ time_step: 5
46
+ use_wandb: true
47
+ layers_after_pool:
48
+ act_fn: SiLU
49
+ act_last_layer: true
50
+ dims:
51
+ - 512
52
+ - 512
53
+ dropout_rate: 0.0
54
+ input_dim: -1
55
+ layer_type: mlp
56
+ layers_before_pool:
57
+ batch_onehot:
58
+ dims:
59
+ - 256
60
+ - 256
61
+ dropout_rate: 0.0
62
+ input_dim: 256
63
+ layer_type: embedding
64
+ num_embeddings: 5
65
+ cell_type_onehot:
66
+ dims:
67
+ - 256
68
+ - 256
69
+ dropout_rate: 0.0
70
+ input_dim: 256
71
+ layer_type: embedding
72
+ num_embeddings: 92
73
+ pert_onehot:
74
+ dims:
75
+ - 256
76
+ - 256
77
+ dropout_rate: 0.0
78
+ input_dim: 256
79
+ layer_type: embedding
80
+ num_embeddings: 28
81
+ layers_for_aggregation:
82
+ act_fn: SiLU
83
+ act_last_layer: true
84
+ dims:
85
+ - 1024
86
+ - 512
87
+ dropout_rate: 0.0
88
+ input_dim: 1024
89
+ layer_type: mlp
90
+ model_name: JIT_LLMSET
91
+ prepare_model:
92
+ batch_class: 5
93
+ cell_class: 92
94
+ cell_dim: 2000
95
+ cell_sentence_len: 32
96
+ cellmap_width: 512
97
+ cond_output_dropout: 0.5
98
+ condition_embedding_dim: 256
99
+ condition_encoder_kwargs: {}
100
+ condition_mode: deterministic
101
+ conditioning: concatenation
102
+ dit_depth: 8
103
+ dit_kwargs:
104
+ mlp_ratio: 4.0
105
+ num_heads: 16
106
+ hidden_channels: 32
107
+ hidden_dims:
108
+ - 2048
109
+ - 2048
110
+ - 2048
111
+ hidden_dropout: 0.0
112
+ layer_norm_before_concatenation: false
113
+ layers_after_pool:
114
+ act_fn: SiLU
115
+ act_last_layer: true
116
+ dims:
117
+ - 512
118
+ - 512
119
+ dropout_rate: 0.0
120
+ input_dim: -1
121
+ layer_type: mlp
122
+ layers_before_pool:
123
+ batch_onehot:
124
+ dims:
125
+ - 256
126
+ - 256
127
+ dropout_rate: 0.0
128
+ input_dim: 256
129
+ layer_type: embedding
130
+ num_embeddings: 5
131
+ cell_type_onehot:
132
+ dims:
133
+ - 256
134
+ - 256
135
+ dropout_rate: 0.0
136
+ input_dim: 256
137
+ layer_type: embedding
138
+ num_embeddings: 92
139
+ pert_onehot:
140
+ dims:
141
+ - 256
142
+ - 256
143
+ dropout_rate: 0.0
144
+ input_dim: 256
145
+ layer_type: embedding
146
+ num_embeddings: 28
147
+ layers_for_aggregation:
148
+ act_fn: SiLU
149
+ act_last_layer: true
150
+ dims:
151
+ - 1024
152
+ - 512
153
+ dropout_rate: 0.0
154
+ input_dim: 1024
155
+ layer_type: mlp
156
+ learning_rate: 0.0002
157
+ linear_projection_before_concatenation: false
158
+ loss_type: x
159
+ match_fn: match_linear
160
+ megatron_ckpt: null
161
+ n_decoder_layers: 4
162
+ pool_sample_covariates: true
163
+ pooling: attention_token
164
+ pooling_kwargs:
165
+ num_heads: 8
166
+ pool_method: mean
167
+ pred_type: x
168
+ regularization: 0.0
169
+ seed: 0
170
+ solver_kwargs: {}
171
+ src_loss_weight: 0.25
172
+ time_encoder_dims:
173
+ - 1024
174
+ - 512
175
+ time_encoder_dropout: 0.0
176
+ time_freqs: 512
177
+ time_max_period: 10000
178
+ transformer_backbone_kwargs:
179
+ backbone_name: llama
180
+ hidden_size: 512
181
+ intermediate_size: 1024
182
+ max_position_embeddings: 32
183
+ num_attention_heads: 8
184
+ num_hidden_layers: 4
185
+ num_key_value_heads: 8
186
+ vocab_size: 100
187
+ configure_init_model_parallel: false
188
+ model_transform: null
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/.metadata ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff86273c1484285b731677cc19facede90af0b9adff32186116bdac1696bd71b
3
+ size 224089
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b3f8347d13b8f3baa3fa0d31ad7a5707991012f8dc73a9127c6004f40f8e64d
3
+ size 343806872
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/__0_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5d0c512d4eb405c67a926ca7af3a4e24b859526373bf1a19fbf31fb87afae04
3
+ size 343809232
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/common.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:308bf8b0a776ce2def45f9034583c649fe3facaa22016e31fe6a09015756590b
3
+ size 4067
checkpoints/zscape_heldout_cells/finetune_pearson_best-epoch=29/weights/metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sharded_backend": "torch_dist", "sharded_backend_version": 1, "common_backend": "torch", "common_backend_version": 1}
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/context/model.yaml ADDED
@@ -0,0 +1,216 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ checkpoint_path: checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest
2
+ _target_: models.JIT_LLMSET_GENETIME.nemo_model.JIT_LLMSET_GENETIME_Nemo_Model
3
+ conf:
4
+ Interpolant:
5
+ prior_distribution: gaussian
6
+ prior_kwargs:
7
+ logits: 0.1
8
+ total_count: 1
9
+ zi_logits: 0
10
+ probability_path:
11
+ constant_noise: 0.0
12
+ cell_sentence_len: 32
13
+ cellmap_width: 512
14
+ dataset:
15
+ batch_size: 16
16
+ cell_sentence_len: 32
17
+ control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/control_train
18
+ control_pert: ctrl-inj
19
+ fingerprint_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/gene_target_fingerprints.pkl
20
+ global_keys_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/global_keys.pkl
21
+ num_workers: 4
22
+ random_seed: 42
23
+ test_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/control_test
24
+ test_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/test
25
+ train_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/control_train
26
+ train_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/train
27
+ val_control_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/control_val
28
+ val_lmdb_path: resources/nemo_cellflow/analyzing/zscape_develop/data_processed/zscape_seen_cell_unseen_gene_v1/lmdb/val
29
+ dataset_name: tahoe
30
+ experiment:
31
+ accumulate_grad_batches: 1
32
+ check_val_every_n_epoch: 5
33
+ ckpt_freq: 10
34
+ group: zscape-unseen-gene-genetime-gt-lr2e4-drop02-src010-dpcc005-20260706
35
+ monitor: val_loss
36
+ num_epoch: 60
37
+ num_gpus: 1
38
+ num_nodes: 1
39
+ num_sanity_val_steps: 2
40
+ pipeline_model_parallel_size: 1
41
+ precision: 32
42
+ seed: 42
43
+ strategy: ddp
44
+ tensor_model_parallel_size: 1
45
+ time_step: 5
46
+ use_wandb: true
47
+ layers_after_pool:
48
+ act_fn: SiLU
49
+ act_last_layer: true
50
+ dims:
51
+ - 512
52
+ - 512
53
+ dropout_rate: 0.0
54
+ input_dim: -1
55
+ layer_type: mlp
56
+ layers_before_pool:
57
+ batch_onehot:
58
+ dims:
59
+ - 256
60
+ - 256
61
+ dropout_rate: 0.0
62
+ input_dim: 256
63
+ layer_type: embedding
64
+ num_embeddings: 5
65
+ cell_type_onehot:
66
+ dims:
67
+ - 256
68
+ - 256
69
+ dropout_rate: 0.0
70
+ input_dim: 256
71
+ layer_type: embedding
72
+ num_embeddings: 92
73
+ pert_fingerprint:
74
+ continuous_dim: 256
75
+ dims:
76
+ - 256
77
+ - 256
78
+ dropout_rate: 0.0
79
+ input_dim: 256
80
+ layer_type: continuous
81
+ layers_for_aggregation:
82
+ act_fn: SiLU
83
+ act_last_layer: true
84
+ dims:
85
+ - 1024
86
+ - 512
87
+ dropout_rate: 0.0
88
+ input_dim: 1024
89
+ layer_type: mlp
90
+ model_name: JIT_LLMSET_GENETIME
91
+ prepare_model:
92
+ batch_class: 5
93
+ cell_class: 92
94
+ cell_dim: 2000
95
+ cell_sentence_len: 32
96
+ cellmap_width: 512
97
+ cond_output_dropout: 0.2
98
+ condition_embedding_dim: 256
99
+ condition_encoder_kwargs: {}
100
+ condition_mode: deterministic
101
+ conditioning: concatenation
102
+ delta_mae_loss_weight: 0.0
103
+ delta_mse_loss_weight: 0.0
104
+ delta_pearson_loss_weight: 0.05
105
+ dev_time_encoder_dims:
106
+ - 256
107
+ - 256
108
+ dev_time_freqs: 64
109
+ dev_time_max_period: 10000
110
+ dev_time_normalization: minmax
111
+ dev_time_values_by_index:
112
+ - 18
113
+ - 24
114
+ - 36
115
+ - 48
116
+ - 72
117
+ dit_depth: 8
118
+ dit_kwargs:
119
+ mlp_ratio: 4.0
120
+ num_heads: 16
121
+ gene_time_cell_key: cell_type_onehot
122
+ gene_time_condition_fusion_dims:
123
+ - 512
124
+ - 512
125
+ gene_time_dropout: 0.0
126
+ gene_time_gene_key: pert_fingerprint
127
+ gene_time_interaction_dims:
128
+ - 256
129
+ - 256
130
+ gene_time_time_fusion_dims:
131
+ - 256
132
+ - 256
133
+ gene_time_time_key: batch_onehot
134
+ hidden_channels: 32
135
+ hidden_dims:
136
+ - 2048
137
+ - 2048
138
+ - 2048
139
+ hidden_dropout: 0.0
140
+ layer_norm_before_concatenation: false
141
+ layers_after_pool:
142
+ act_fn: SiLU
143
+ act_last_layer: true
144
+ dims:
145
+ - 512
146
+ - 512
147
+ dropout_rate: 0.0
148
+ input_dim: -1
149
+ layer_type: mlp
150
+ layers_before_pool:
151
+ batch_onehot:
152
+ dims:
153
+ - 256
154
+ - 256
155
+ dropout_rate: 0.0
156
+ input_dim: 256
157
+ layer_type: embedding
158
+ num_embeddings: 5
159
+ cell_type_onehot:
160
+ dims:
161
+ - 256
162
+ - 256
163
+ dropout_rate: 0.0
164
+ input_dim: 256
165
+ layer_type: embedding
166
+ num_embeddings: 92
167
+ pert_fingerprint:
168
+ continuous_dim: 256
169
+ dims:
170
+ - 256
171
+ - 256
172
+ dropout_rate: 0.0
173
+ input_dim: 256
174
+ layer_type: continuous
175
+ layers_for_aggregation:
176
+ act_fn: SiLU
177
+ act_last_layer: true
178
+ dims:
179
+ - 1024
180
+ - 512
181
+ dropout_rate: 0.0
182
+ input_dim: 1024
183
+ layer_type: mlp
184
+ learning_rate: 0.0002
185
+ linear_projection_before_concatenation: false
186
+ loss_type: x
187
+ match_fn: match_linear
188
+ megatron_ckpt: null
189
+ n_decoder_layers: 4
190
+ pool_sample_covariates: true
191
+ pooling: attention_token
192
+ pooling_kwargs:
193
+ num_heads: 8
194
+ pool_method: mean
195
+ pred_type: x
196
+ regularization: 0.0
197
+ seed: 0
198
+ solver_kwargs: {}
199
+ src_loss_weight: 0.1
200
+ time_encoder_dims:
201
+ - 1024
202
+ - 512
203
+ time_encoder_dropout: 0.0
204
+ time_freqs: 512
205
+ time_max_period: 10000
206
+ transformer_backbone_kwargs:
207
+ backbone_name: llama
208
+ hidden_size: 512
209
+ intermediate_size: 1024
210
+ max_position_embeddings: 32
211
+ num_attention_heads: 8
212
+ num_hidden_layers: 4
213
+ num_key_value_heads: 8
214
+ vocab_size: 100
215
+ configure_init_model_parallel: false
216
+ model_transform: null
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/.metadata ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8e9e4d46e284f8dd365aad7cef291e7ea68a4e8c2810a128cbfb22ae09e89cf
3
+ size 242161
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_0.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3042a12dec59e9852d52e917828f8e6c2786c61e8b8ba88459ed59bddd6035a2
3
+ size 340351012
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/__0_1.distcp ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5febb198b08aace5ab38f3deb1be266d9d4ba0b4324453abb9fff7d76ae861e6
3
+ size 340349960
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/common.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:013f699b883c0a23a04dd73ba22a06b896efd04ad81e8ae75d4f71a66b88f618
3
+ size 4106
checkpoints/zscape_unseen_genes/genetime_gt_lr2e4_drop02_src010_dpcc005_ep44_valbest/weights/metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"sharded_backend": "torch_dist", "sharded_backend_version": 1, "common_backend": "torch", "common_backend_version": 1}