CIFAR-10 EDM2 large runs A and B: best EMA + final full checkpoints
Browse files- cifar_edm2/README.md +32 -0
- cifar_edm2/run_a_linear/args.json +133 -0
- cifar_edm2/run_a_linear/ema_step_0300000.pt +3 -0
- cifar_edm2/run_a_linear/latest_step_0402000.pt +3 -0
- cifar_edm2/run_a_linear/log_fid.jsonl +17 -0
- cifar_edm2/run_b_trunc_std1/args.json +143 -0
- cifar_edm2/run_b_trunc_std1/ema_step_0125000.pt +3 -0
- cifar_edm2/run_b_trunc_std1/latest_step_0402000.pt +3 -0
- cifar_edm2/run_b_trunc_std1/log_fid.jsonl +17 -0
cifar_edm2/README.md
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# EqFM on CIFAR-10, EDM2 U-Net, from scratch (DeltaAI, 4x GH200, global batch 512)
|
| 2 |
+
|
| 3 |
+
Code: github.com/justinlin26/EqFM, branch `mnist-smf-ablation` (`scripts/train_stage2_flowmap_mnist.py`,
|
| 4 |
+
`eqfm/models/edm2_flowmap.py`); recipe notes in `docs/RUN_ME_edm2_v2.md`.
|
| 5 |
+
Unconditional; the flow map is X_sigma(x) = x + sigma * v(x, sigma). Network: the flow-map paper's EDM2 CIFAR
|
| 6 |
+
U-Net (Boffi et al., arXiv 2505.18825), 54.67M params. Shifted MeanFlow loss with a live self-teacher.
|
| 7 |
+
|
| 8 |
+
| Folder | Run | Recipe |
|
| 9 |
+
|---|---|---|
|
| 10 |
+
| `run_b_trunc_std1/` | `edm2_large_b512_trunc_std1_s0` | truncated clock a=0.9, noise std 1 (sigma_data 1); diag weight 768 (256 from step 211277); LR 4e-3, 8e-4 at 130k, 1e-4 at 250k |
|
| 11 |
+
| `run_a_linear/` | `edm2_large_b512_s0` | linear clock, noise std 0.5 (sigma_data 0.5); diag weight 768 (3072 from step 60298); LR 4e-3, cosine to 1e-4 from 242k |
|
| 12 |
+
|
| 13 |
+
Files per run:
|
| 14 |
+
- `ema_step_XXXXXXX.pt`: best EMA checkpoint (EMA weights + training args). Run B: 125k. Run A: 300k.
|
| 15 |
+
- `latest_step_0402000.pt`: final full checkpoint (raw weights `model`, `ema`, AdamW `optimizer`, `step`) -- resumable.
|
| 16 |
+
- `args.json`: trainer arguments at the last launch; `log_fid.jsonl`: in-run FID-2k (EMA and raw) every 25k steps.
|
| 17 |
+
|
| 18 |
+
FID (CIFAR-10 train statistics, clean-fid legacy_tensorflow, expuniform sampler, EMA):
|
| 19 |
+
|
| 20 |
+
| Checkpoint | Metric | 1 | 2 | 4 | 8 | 16 steps |
|
| 21 |
+
|---|---|---|---|---|---|---|
|
| 22 |
+
| Run A, 300k | FID-50k | 11.30 | 12.81 | 10.54 | 7.50 | 7.04 |
|
| 23 |
+
| Run A, 402k | FID-50k | 14.24 | 15.77 | 13.27 | 8.41 | 7.10 |
|
| 24 |
+
| Run B, 125k | FID-2k (in-run) | 20.8 | 19.2 | 21.2 | 20.3 | – |
|
| 25 |
+
| Run B, 402k | FID-2k (in-run) | 25.0 | 20.7 | 22.1 | 20.9 | – |
|
| 26 |
+
|
| 27 |
+
(In run A, FID-2k read 13.1 +- 0.2 higher than FID-50k; run B's FID-50k is being measured.)
|
| 28 |
+
|
| 29 |
+
Evaluate: `python scripts/eval_fid_cifar.py --mode eqfm --ckpt <file> --nfe 1 2 4 8 16 --schedule expuniform
|
| 30 |
+
--num-gen 50000 --batch-size 500 --seed 2024` (reads clock and noise std from the checkpoint).
|
| 31 |
+
Resume: full checkpoint via `--resume-from latest_step_0402000.pt`; from an EMA snapshot via
|
| 32 |
+
`--resume-from ema_step_XXXXXXX.pt --rewarmup 1000` (optimizer reset), into a new run directory.
|
cifar_edm2/run_a_linear/args.json
ADDED
|
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"data_root": "/work/nvme/biqs/mdrieux/cifar-hth/data",
|
| 3 |
+
"dataset": "cifar",
|
| 4 |
+
"arch": "edm2",
|
| 5 |
+
"edm2_sigma_data": 0.5,
|
| 6 |
+
"edm2_dropout": 0.13,
|
| 7 |
+
"base_std": 0.5,
|
| 8 |
+
"init_fm_ckpt": null,
|
| 9 |
+
"results_dir": "/work/nvme/biqs/mdrieux/edm2-large",
|
| 10 |
+
"run_name": "edm2_large_b512_s0",
|
| 11 |
+
"self_teacher": true,
|
| 12 |
+
"diag_adaptive_p": 0.0,
|
| 13 |
+
"diag_blowup_factor": 0.0,
|
| 14 |
+
"offdiag_target_source": "live",
|
| 15 |
+
"self_teacher_source": "live",
|
| 16 |
+
"offdiag_b_source": "teacher",
|
| 17 |
+
"sigma_zero_frac": 0.0,
|
| 18 |
+
"ema_warmup": true,
|
| 19 |
+
"diag_c_eps": 0.01,
|
| 20 |
+
"diag_target": "eqm",
|
| 21 |
+
"diag_only_steps": 2000,
|
| 22 |
+
"teacher_ckpt": null,
|
| 23 |
+
"model": "SiT-S/4",
|
| 24 |
+
"image_size": 32,
|
| 25 |
+
"num_classes": 10,
|
| 26 |
+
"student_model": null,
|
| 27 |
+
"objective": "shifted_meanflow",
|
| 28 |
+
"semi_weight": 0.0,
|
| 29 |
+
"idem_weight": 0.0,
|
| 30 |
+
"idem_on": "both",
|
| 31 |
+
"idem_iters": 1,
|
| 32 |
+
"idem_target": "self",
|
| 33 |
+
"idem_polish_h": 0.5,
|
| 34 |
+
"idem_polish_steps": 4,
|
| 35 |
+
"idem_polish_weight": 0.1,
|
| 36 |
+
"idem_ema_h": 0.05,
|
| 37 |
+
"cd_delta": 0.05,
|
| 38 |
+
"cd_substeps": 1,
|
| 39 |
+
"meanflow_eps": 0.01,
|
| 40 |
+
"meanflow_clock": "clamp",
|
| 41 |
+
"interp_time_sampler": "uniform",
|
| 42 |
+
"interp_is_correct": false,
|
| 43 |
+
"interp_ref": "canonical",
|
| 44 |
+
"importance_k": 2.0,
|
| 45 |
+
"interp_logitnormal": [
|
| 46 |
+
-0.4,
|
| 47 |
+
1.0
|
| 48 |
+
],
|
| 49 |
+
"eps_train": 0.001,
|
| 50 |
+
"sigma_min": 0.0001,
|
| 51 |
+
"sigma_max": 1.0,
|
| 52 |
+
"sigma_sampler": "uniform",
|
| 53 |
+
"sigma_is_correct": false,
|
| 54 |
+
"sigma_ref": "canonical",
|
| 55 |
+
"sigma_power_k": 2.0,
|
| 56 |
+
"sigma_logitnormal": [
|
| 57 |
+
-0.4,
|
| 58 |
+
1.0
|
| 59 |
+
],
|
| 60 |
+
"time_weight_clip": 0.0,
|
| 61 |
+
"guidance_mode": "branch",
|
| 62 |
+
"teacher_cfg_scale": 1.0,
|
| 63 |
+
"cfg_rule": "standard",
|
| 64 |
+
"cfg_cnorm": false,
|
| 65 |
+
"cfg_match_uncond_norm": false,
|
| 66 |
+
"cfg_match_guided_norm": false,
|
| 67 |
+
"terminal_jacobian_weight": 0.0,
|
| 68 |
+
"terminal_jacobian_input": "interp",
|
| 69 |
+
"terminal_jacobian_max_batch": 64,
|
| 70 |
+
"terminal_invariance_weight": 0.0,
|
| 71 |
+
"terminal_invariance_step": 0.1,
|
| 72 |
+
"terminal_invariance_max_batch": 64,
|
| 73 |
+
"wandb_project": "eqfm-cifar-edm2",
|
| 74 |
+
"wandb_entity": null,
|
| 75 |
+
"wandb_mode": "online",
|
| 76 |
+
"wandb_tags": "edm2_large",
|
| 77 |
+
"bucket_grad_every": 0,
|
| 78 |
+
"max_nonfinite_window": 20,
|
| 79 |
+
"diag_weight": 3072.0,
|
| 80 |
+
"distill_weight": 1.0,
|
| 81 |
+
"terminal_weight": 0.0,
|
| 82 |
+
"terminal_teacher_norm_weight": 0.0,
|
| 83 |
+
"terminal_norm_max_batch": 64,
|
| 84 |
+
"adaptive_weight_p": 1.0,
|
| 85 |
+
"adaptive_weight_eps": 3.25521e-07,
|
| 86 |
+
"loss_reduction": "mean",
|
| 87 |
+
"adaptive_weight_mode": "sample",
|
| 88 |
+
"cfg_dropout_prob": 0.0,
|
| 89 |
+
"batch_size": 512,
|
| 90 |
+
"cuda_graph": true,
|
| 91 |
+
"micro_batch": 0,
|
| 92 |
+
"steps": 400000,
|
| 93 |
+
"lr": 0.004,
|
| 94 |
+
"warmup_steps": 1000,
|
| 95 |
+
"lr_decay_from": 240000,
|
| 96 |
+
"lr_final": 0.0001,
|
| 97 |
+
"ema_snapshot_every": 25000,
|
| 98 |
+
"rewarmup": 0,
|
| 99 |
+
"resume_from": null,
|
| 100 |
+
"grad_clip": 0.0,
|
| 101 |
+
"ema_decay": 0.9999,
|
| 102 |
+
"log_every": 100,
|
| 103 |
+
"val_every": 1000000,
|
| 104 |
+
"val_n": 256,
|
| 105 |
+
"ode_ref_steps": 400,
|
| 106 |
+
"save_every": 2000,
|
| 107 |
+
"offdiag_mode": "split",
|
| 108 |
+
"offdiag_q": [
|
| 109 |
+
0.25,
|
| 110 |
+
0.25
|
| 111 |
+
],
|
| 112 |
+
"offdiag_q_schedule": "const",
|
| 113 |
+
"offdiag_q_switch_at": 0.5,
|
| 114 |
+
"diag_only_frac": 0.0,
|
| 115 |
+
"compute_budget": 0.0,
|
| 116 |
+
"cost_offdiag": 0.0,
|
| 117 |
+
"cost_both": 0.0,
|
| 118 |
+
"budget_milestones": [
|
| 119 |
+
0.25,
|
| 120 |
+
0.5,
|
| 121 |
+
0.75,
|
| 122 |
+
1.0
|
| 123 |
+
],
|
| 124 |
+
"resume": true,
|
| 125 |
+
"flowmap_clock": "linear",
|
| 126 |
+
"flowmap_clock_a": 0.9,
|
| 127 |
+
"fid_every": 25000,
|
| 128 |
+
"fid_n": 2000,
|
| 129 |
+
"fid_ref_mnist": null,
|
| 130 |
+
"save_every_minutes": 10.0,
|
| 131 |
+
"ckpt_mode": "full",
|
| 132 |
+
"seed": 0
|
| 133 |
+
}
|
cifar_edm2/run_a_linear/ema_step_0300000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7ad18c9fc620f34e4a26eb67d9377b8994cba5fa33d170ea411738d273f2e77d
|
| 3 |
+
size 218756407
|
cifar_edm2/run_a_linear/latest_step_0402000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5fac2bb95a9465a309b13f30bbf0894f9d752ee331b1e513a97a3ab7fa509b3a
|
| 3 |
+
size 875040095
|
cifar_edm2/run_a_linear/log_fid.jsonl
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"step": 25000, "fid": {"EMA K1": 99.57170859589132, "EMA K2_expuniform": 107.91562602295306, "EMA K4_expuniform": 104.90291973702051, "EMA K8_expuniform": 105.32418132902461, "raw K1": 163.71937322709223, "raw K2_expuniform": 134.17373921843148, "raw K4_expuniform": 125.00918694945659, "raw K8_expuniform": 138.20905457598792, "raw K1_end0.999": 164.64843944357818, "raw K1_end0.97": 198.97684132036864, "raw K2_s0.5_end0.999": 134.71430560365263}, "fid_n": 2000}
|
| 2 |
+
{"step": 50000, "fid": {"EMA K1": 87.65014722978816, "EMA K2_expuniform": 96.30905293229159, "EMA K4_expuniform": 87.27831388286222, "EMA K8_expuniform": 90.66487669200234, "raw K1": 89.10271870715167, "raw K2_expuniform": 138.01065826067367, "raw K4_expuniform": 133.24007740485797, "raw K8_expuniform": 147.41846590731393, "raw K1_end0.999": 89.3863683692943, "raw K1_end0.97": 107.21376069755058, "raw K2_s0.5_end0.999": 138.1169053306212}, "fid_n": 2000}
|
| 3 |
+
{"step": 75000, "fid": {"EMA K1": 42.47347063458818, "EMA K2_expuniform": 47.387416298790185, "EMA K4_expuniform": 43.35081559296816, "EMA K8_expuniform": 40.25243847631452, "raw K1": 47.34455854136547, "raw K2_expuniform": 48.1624408969937, "raw K4_expuniform": 42.94177741891065, "raw K8_expuniform": 39.42930823944732, "raw K1_end0.999": 47.31421579811786, "raw K1_end0.97": 44.909103021053795, "raw K2_s0.5_end0.999": 48.07889450442053}, "fid_n": 2000}
|
| 4 |
+
{"step": 100000, "fid": {"EMA K1": 31.077150280243814, "EMA K2_expuniform": 32.68314271480466, "EMA K4_expuniform": 31.571310133617203, "EMA K8_expuniform": 28.58346249065937, "raw K1": 42.65270259662867, "raw K2_expuniform": 43.90629188343212, "raw K4_expuniform": 42.70450814973054, "raw K8_expuniform": 40.62379457997605, "raw K1_end0.999": 42.54288933264479, "raw K1_end0.97": 41.8149147859458, "raw K2_s0.5_end0.999": 43.82598360944962}, "fid_n": 2000}
|
| 5 |
+
{"step": 125000, "fid": {"EMA K1": 28.863616423691212, "EMA K2_expuniform": 28.952583711517832, "EMA K4_expuniform": 27.700162350418225, "EMA K8_expuniform": 24.73696241758256, "raw K1": 38.62562076510687, "raw K2_expuniform": 49.67263710286869, "raw K4_expuniform": 50.00454173299369, "raw K8_expuniform": 51.37390216848746, "raw K1_end0.999": 38.43085745480727, "raw K1_end0.97": 38.416031167162885, "raw K2_s0.5_end0.999": 49.438230062603225}, "fid_n": 2000}
|
| 6 |
+
{"step": 150000, "fid": {"EMA K1": 27.251091839492744, "EMA K2_expuniform": 26.919908457769623, "EMA K4_expuniform": 25.615438535371425, "EMA K8_expuniform": 22.679424985508376, "raw K1": 39.66995333672054, "raw K2_expuniform": 38.5313998861497, "raw K4_expuniform": 38.31283436177233, "raw K8_expuniform": 33.763891736624316, "raw K1_end0.999": 40.03318539793764, "raw K1_end0.97": 54.59735777532603, "raw K2_s0.5_end0.999": 38.36258112730701}, "fid_n": 2000}
|
| 7 |
+
{"step": 175000, "fid": {"EMA K1": 26.602911868699607, "EMA K2_expuniform": 26.603695739935347, "EMA K4_expuniform": 25.057282645730652, "EMA K8_expuniform": 22.0668074362913, "raw K1": 51.2160438219197, "raw K2_expuniform": 51.24736063942248, "raw K4_expuniform": 35.00691224780127, "raw K8_expuniform": 31.996554778665313, "raw K1_end0.999": 50.85698329513673, "raw K1_end0.97": 44.63510965768279, "raw K2_s0.5_end0.999": 50.85044162351545}, "fid_n": 2000}
|
| 8 |
+
{"step": 200000, "fid": {"EMA K1": 25.328334122564, "EMA K2_expuniform": 25.685694133692948, "EMA K4_expuniform": 24.94138866438442, "EMA K8_expuniform": 21.741834124095476, "raw K1": 35.63680057266396, "raw K2_expuniform": 44.89994995214886, "raw K4_expuniform": 40.40851013950345, "raw K8_expuniform": 35.36654034128185, "raw K1_end0.999": 35.510017833273594, "raw K1_end0.97": 34.5867153380068, "raw K2_s0.5_end0.999": 44.83079360996226}, "fid_n": 2000}
|
| 9 |
+
{"step": 225000, "fid": {"EMA K1": 26.984202921650194, "EMA K2_expuniform": 26.84131438194794, "EMA K4_expuniform": 24.822042947609873, "EMA K8_expuniform": 21.467546865488544, "raw K1": 32.86844619472515, "raw K2_expuniform": 37.311487554817234, "raw K4_expuniform": 37.985868415844095, "raw K8_expuniform": 40.93540440458696, "raw K1_end0.999": 32.782170774102724, "raw K1_end0.97": 43.708999029005156, "raw K2_s0.5_end0.999": 37.04666381417303}, "fid_n": 2000}
|
| 10 |
+
{"step": 250000, "fid": {"EMA K1": 26.62010753259449, "EMA K2_expuniform": 26.60540273423817, "EMA K4_expuniform": 24.51309324858289, "EMA K8_expuniform": 21.118741795505002, "raw K1": 35.81845942246446, "raw K2_expuniform": 40.94356989187844, "raw K4_expuniform": 46.36789660690886, "raw K8_expuniform": 53.126512779238624, "raw K1_end0.999": 35.542660322371034, "raw K1_end0.97": 34.185237931726306, "raw K2_s0.5_end0.999": 40.93275014691818}, "fid_n": 2000}
|
| 11 |
+
{"step": 275000, "fid": {"EMA K1": 24.70050354459221, "EMA K2_expuniform": 26.052629390316213, "EMA K4_expuniform": 24.08333487804009, "EMA K8_expuniform": 20.99821654055927, "raw K1": 69.5115186092379, "raw K2_expuniform": 64.33022969599602, "raw K4_expuniform": 52.73181382171026, "raw K8_expuniform": 47.05284423600381, "raw K1_end0.999": 69.21215805434474, "raw K1_end0.97": 64.39115665642453, "raw K2_s0.5_end0.999": 64.38669712155092}, "fid_n": 2000}
|
| 12 |
+
{"step": 300000, "fid": {"EMA K1": 24.290901189646092, "EMA K2_expuniform": 26.048586943673854, "EMA K4_expuniform": 23.51185194064408, "EMA K8_expuniform": 20.587931836702808, "raw K1": 29.621482798308477, "raw K2_expuniform": 30.429818299871158, "raw K4_expuniform": 30.941514521631575, "raw K8_expuniform": 29.07108756687751, "raw K1_end0.999": 29.613905242774536, "raw K1_end0.97": 35.60817815108328, "raw K2_s0.5_end0.999": 30.322723713953792}, "fid_n": 2000}
|
| 13 |
+
{"step": 325000, "fid": {"EMA K1": 25.23324602379563, "EMA K2_expuniform": 26.520671594469036, "EMA K4_expuniform": 24.143902781553777, "EMA K8_expuniform": 21.130842230453368, "raw K1": 29.960355365270686, "raw K2_expuniform": 33.310152389504196, "raw K4_expuniform": 32.36888712480754, "raw K8_expuniform": 30.04984632674916, "raw K1_end0.999": 29.863353906083205, "raw K1_end0.97": 37.79119877697349, "raw K2_s0.5_end0.999": 33.131422471957535}, "fid_n": 2000}
|
| 14 |
+
{"step": 350000, "fid": {"EMA K1": 25.547991051260908, "EMA K2_expuniform": 27.363201625371914, "EMA K4_expuniform": 25.020607280290392, "EMA K8_expuniform": 21.39528854045932, "raw K1": 29.234047490248656, "raw K2_expuniform": 30.755489301199532, "raw K4_expuniform": 28.107265519294913, "raw K8_expuniform": 24.13117552090796, "raw K1_end0.999": 29.029066378336324, "raw K1_end0.97": 30.054909925144443, "raw K2_s0.5_end0.999": 30.592447975162827}, "fid_n": 2000}
|
| 15 |
+
{"step": 375000, "fid": {"EMA K1": 26.97702698252408, "EMA K2_expuniform": 28.651478011504764, "EMA K4_expuniform": 25.867667849687507, "EMA K8_expuniform": 21.787012382316163, "raw K1": 31.99445606110544, "raw K2_expuniform": 34.38695893885307, "raw K4_expuniform": 31.292537602042216, "raw K8_expuniform": 27.574784353008965, "raw K1_end0.999": 31.914476242141745, "raw K1_end0.97": 38.40307867240074, "raw K2_s0.5_end0.999": 34.138321800591484}, "fid_n": 2000}
|
| 16 |
+
{"step": 400000, "fid": {"EMA K1": 27.813266666142226, "EMA K2_expuniform": 29.448669715531935, "EMA K4_expuniform": 26.078654996879322, "EMA K8_expuniform": 21.637462794096336, "raw K1": 31.749065928943196, "raw K2_expuniform": 33.37284466768335, "raw K4_expuniform": 29.998908575377584, "raw K8_expuniform": 25.937475282608148, "raw K1_end0.999": 31.460301909029, "raw K1_end0.97": 32.54949655822418, "raw K2_s0.5_end0.999": 33.110377754273145}, "fid_n": 2000}
|
| 17 |
+
{"step": 402000, "fid": {"EMA K1": 27.36981518867117, "EMA K2_expuniform": 29.05498301275935, "EMA K4_expuniform": 25.993837387061262, "EMA K8_expuniform": 21.536477778382277, "raw K1": 39.35758928214818, "raw K2_expuniform": 39.4223660925889, "raw K4_expuniform": 31.455260620401646, "raw K8_expuniform": 26.26750825508384, "raw K1_end0.999": 39.15179643355032, "raw K1_end0.97": 37.83680026021267, "raw K2_s0.5_end0.999": 39.050355333665095}, "fid_n": 2000}
|
cifar_edm2/run_b_trunc_std1/args.json
ADDED
|
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"data_root": "/work/nvme/biqs/mdrieux/cifar-hth/data",
|
| 3 |
+
"dataset": "cifar",
|
| 4 |
+
"arch": "edm2",
|
| 5 |
+
"edm2_sigma_data": 1.0,
|
| 6 |
+
"edm2_dropout": 0.13,
|
| 7 |
+
"base_std": 1.0,
|
| 8 |
+
"init_fm_ckpt": null,
|
| 9 |
+
"results_dir": "/work/nvme/biqs/mdrieux/edm2-large",
|
| 10 |
+
"run_name": "edm2_large_b512_trunc_std1_s0",
|
| 11 |
+
"self_teacher": true,
|
| 12 |
+
"diag_adaptive_p": 0.0,
|
| 13 |
+
"diag_blowup_factor": 0.0,
|
| 14 |
+
"offdiag_target_source": "live",
|
| 15 |
+
"self_teacher_source": "live",
|
| 16 |
+
"offdiag_b_source": "teacher",
|
| 17 |
+
"sigma_zero_frac": 0.0,
|
| 18 |
+
"ema_warmup": true,
|
| 19 |
+
"diag_c_eps": 0.01,
|
| 20 |
+
"diag_target": "eqm",
|
| 21 |
+
"diag_only_steps": 2000,
|
| 22 |
+
"teacher_ckpt": null,
|
| 23 |
+
"model": "SiT-S/4",
|
| 24 |
+
"image_size": 32,
|
| 25 |
+
"num_classes": 10,
|
| 26 |
+
"student_model": null,
|
| 27 |
+
"objective": "shifted_meanflow",
|
| 28 |
+
"semi_weight": 0.0,
|
| 29 |
+
"idem_weight": 0.0,
|
| 30 |
+
"idem_on": "both",
|
| 31 |
+
"idem_iters": 1,
|
| 32 |
+
"idem_target": "self",
|
| 33 |
+
"idem_polish_h": 0.5,
|
| 34 |
+
"idem_polish_steps": 4,
|
| 35 |
+
"idem_polish_weight": 0.1,
|
| 36 |
+
"idem_ema_h": 0.05,
|
| 37 |
+
"cd_delta": 0.05,
|
| 38 |
+
"cd_substeps": 1,
|
| 39 |
+
"meanflow_eps": 0.01,
|
| 40 |
+
"meanflow_clock": "clamp",
|
| 41 |
+
"interp_time_sampler": "uniform",
|
| 42 |
+
"interp_is_correct": false,
|
| 43 |
+
"interp_ref": "canonical",
|
| 44 |
+
"importance_k": 2.0,
|
| 45 |
+
"interp_logitnormal": [
|
| 46 |
+
-0.4,
|
| 47 |
+
1.0
|
| 48 |
+
],
|
| 49 |
+
"eps_train": 0.001,
|
| 50 |
+
"sigma_min": 0.0001,
|
| 51 |
+
"sigma_max": 1.0,
|
| 52 |
+
"sigma_sampler": "uniform",
|
| 53 |
+
"sigma_is_correct": false,
|
| 54 |
+
"sigma_ref": "canonical",
|
| 55 |
+
"sigma_power_k": 2.0,
|
| 56 |
+
"sigma_logitnormal": [
|
| 57 |
+
-0.4,
|
| 58 |
+
1.0
|
| 59 |
+
],
|
| 60 |
+
"time_weight_clip": 0.0,
|
| 61 |
+
"guidance_mode": "branch",
|
| 62 |
+
"teacher_cfg_scale": 1.0,
|
| 63 |
+
"cfg_rule": "standard",
|
| 64 |
+
"cfg_cnorm": false,
|
| 65 |
+
"cfg_match_uncond_norm": false,
|
| 66 |
+
"cfg_match_guided_norm": false,
|
| 67 |
+
"terminal_jacobian_weight": 0.0,
|
| 68 |
+
"terminal_jacobian_input": "interp",
|
| 69 |
+
"terminal_jacobian_max_batch": 64,
|
| 70 |
+
"terminal_invariance_weight": 0.0,
|
| 71 |
+
"terminal_invariance_step": 0.1,
|
| 72 |
+
"terminal_invariance_max_batch": 64,
|
| 73 |
+
"wandb_project": "eqfm-cifar-edm2",
|
| 74 |
+
"wandb_entity": null,
|
| 75 |
+
"wandb_mode": "online",
|
| 76 |
+
"wandb_tags": "edm2_large",
|
| 77 |
+
"bucket_grad_every": 0,
|
| 78 |
+
"max_nonfinite_window": 20,
|
| 79 |
+
"diag_weight": 256.0,
|
| 80 |
+
"distill_weight": 1.0,
|
| 81 |
+
"terminal_weight": 0.0,
|
| 82 |
+
"terminal_teacher_norm_weight": 0.0,
|
| 83 |
+
"terminal_norm_max_batch": 64,
|
| 84 |
+
"adaptive_weight_p": 1.0,
|
| 85 |
+
"adaptive_weight_eps": 3.25521e-07,
|
| 86 |
+
"loss_reduction": "mean",
|
| 87 |
+
"adaptive_weight_mode": "sample",
|
| 88 |
+
"cfg_dropout_prob": 0.0,
|
| 89 |
+
"batch_size": 512,
|
| 90 |
+
"cuda_graph": true,
|
| 91 |
+
"micro_batch": 0,
|
| 92 |
+
"steps": 400000,
|
| 93 |
+
"lr": 0.004,
|
| 94 |
+
"warmup_steps": 1000,
|
| 95 |
+
"lr_decay_from": 340000,
|
| 96 |
+
"lr_final": 0.0001,
|
| 97 |
+
"lr_milestones": [
|
| 98 |
+
130000,
|
| 99 |
+
250000
|
| 100 |
+
],
|
| 101 |
+
"lr_milestone_values": [
|
| 102 |
+
0.0008,
|
| 103 |
+
0.0001
|
| 104 |
+
],
|
| 105 |
+
"diag_weight_milestones": [],
|
| 106 |
+
"diag_weight_milestone_values": [],
|
| 107 |
+
"ema_snapshot_every": 25000,
|
| 108 |
+
"rewarmup": 0,
|
| 109 |
+
"resume_from": null,
|
| 110 |
+
"grad_clip": 0.0,
|
| 111 |
+
"ema_decay": 0.9999,
|
| 112 |
+
"log_every": 100,
|
| 113 |
+
"val_every": 1000000,
|
| 114 |
+
"val_n": 64,
|
| 115 |
+
"ode_ref_steps": 50,
|
| 116 |
+
"save_every": 2000,
|
| 117 |
+
"offdiag_mode": "split",
|
| 118 |
+
"offdiag_q": [
|
| 119 |
+
0.25,
|
| 120 |
+
0.25
|
| 121 |
+
],
|
| 122 |
+
"offdiag_q_schedule": "const",
|
| 123 |
+
"offdiag_q_switch_at": 0.5,
|
| 124 |
+
"diag_only_frac": 0.0,
|
| 125 |
+
"compute_budget": 0.0,
|
| 126 |
+
"cost_offdiag": 0.0,
|
| 127 |
+
"cost_both": 0.0,
|
| 128 |
+
"budget_milestones": [
|
| 129 |
+
0.25,
|
| 130 |
+
0.5,
|
| 131 |
+
0.75,
|
| 132 |
+
1.0
|
| 133 |
+
],
|
| 134 |
+
"resume": true,
|
| 135 |
+
"flowmap_clock": "trunc",
|
| 136 |
+
"flowmap_clock_a": 0.9,
|
| 137 |
+
"fid_every": 25000,
|
| 138 |
+
"fid_n": 2000,
|
| 139 |
+
"fid_ref_mnist": null,
|
| 140 |
+
"save_every_minutes": 10.0,
|
| 141 |
+
"ckpt_mode": "full",
|
| 142 |
+
"seed": 0
|
| 143 |
+
}
|
cifar_edm2/run_b_trunc_std1/ema_step_0125000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d294fa1be69130dd65af66102b44ef640eae6caa22ae76bc4695aa1f5faa2fe8
|
| 3 |
+
size 218756471
|
cifar_edm2/run_b_trunc_std1/latest_step_0402000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bccfaa63fee59269bcc0799d74ab1f5306fc18b12f5e906737996e8d183ec4bd
|
| 3 |
+
size 875040287
|
cifar_edm2/run_b_trunc_std1/log_fid.jsonl
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"step": 25000, "fid": {"EMA K1": 30.522196443902885, "EMA K2_expuniform": 30.860533285430165, "EMA K4_expuniform": 42.34053853129598, "EMA K8_expuniform": 35.278785953093404, "raw K1": 34.936391674354525, "raw K2_expuniform": 37.49230073246065, "raw K4_expuniform": 49.363982547156354, "raw K8_expuniform": 36.919501546558934, "raw K1_end0.999": 35.29459674870384, "raw K1_end0.97": 52.819364664719615, "raw K2_s0.5_end0.999": 37.29445428114798}, "fid_n": 2000}
|
| 2 |
+
{"step": 50000, "fid": {"EMA K1": 22.765231551507895, "EMA K2_expuniform": 25.300260907020515, "EMA K4_expuniform": 28.132661331649103, "EMA K8_expuniform": 24.80841802417342, "raw K1": 32.25871820950755, "raw K2_expuniform": 34.74468000369973, "raw K4_expuniform": 45.5742390769376, "raw K8_expuniform": 41.94434029164864, "raw K1_end0.999": 32.512697633598464, "raw K1_end0.97": 50.458239614296076, "raw K2_s0.5_end0.999": 34.55626525412271}, "fid_n": 2000}
|
| 3 |
+
{"step": 75000, "fid": {"EMA K1": 22.54701221741368, "EMA K2_expuniform": 22.32699183218631, "EMA K4_expuniform": 24.55531965776106, "EMA K8_expuniform": 22.790999179028802, "raw K1": 26.331058720130272, "raw K2_expuniform": 25.744949764935996, "raw K4_expuniform": 32.236180492585504, "raw K8_expuniform": 30.590385179612667, "raw K1_end0.999": 26.668459561952318, "raw K1_end0.97": 48.34806918320612, "raw K2_s0.5_end0.999": 25.65448239149771}, "fid_n": 2000}
|
| 4 |
+
{"step": 100000, "fid": {"EMA K1": 21.29323615572821, "EMA K2_expuniform": 21.3616668316605, "EMA K4_expuniform": 22.805087207491795, "EMA K8_expuniform": 20.77242131286306, "raw K1": 25.201742646965442, "raw K2_expuniform": 26.624309519491476, "raw K4_expuniform": 32.21994043585215, "raw K8_expuniform": 33.84933981217938, "raw K1_end0.999": 25.347258522932123, "raw K1_end0.97": 41.29536221138835, "raw K2_s0.5_end0.999": 26.525815735326375}, "fid_n": 2000}
|
| 5 |
+
{"step": 125000, "fid": {"EMA K1": 20.776623020701606, "EMA K2_expuniform": 19.179882381727396, "EMA K4_expuniform": 21.205542346611537, "EMA K8_expuniform": 20.287036424170765, "raw K1": 42.123322671586266, "raw K2_expuniform": 36.77890225550868, "raw K4_expuniform": 35.23719165757814, "raw K8_expuniform": 28.385569841607264, "raw K1_end0.999": 42.28614824572139, "raw K1_end0.97": 56.174180231516175, "raw K2_s0.5_end0.999": 36.64573567553367}, "fid_n": 2000}
|
| 6 |
+
{"step": 150000, "fid": {"EMA K1": 20.9117342664278, "EMA K2_expuniform": 19.788667409766163, "EMA K4_expuniform": 21.522045201471656, "EMA K8_expuniform": 20.392390697463554, "raw K1": 28.05090644303351, "raw K2_expuniform": 24.041528703191148, "raw K4_expuniform": 23.877063042964664, "raw K8_expuniform": 21.565648596184587, "raw K1_end0.999": 27.470517008365675, "raw K1_end0.97": 40.14900813999594, "raw K2_s0.5_end0.999": 23.91144164449912}, "fid_n": 2000}
|
| 7 |
+
{"step": 175000, "fid": {"EMA K1": 21.64050867286602, "EMA K2_expuniform": 20.606872978569754, "EMA K4_expuniform": 22.032925420901506, "EMA K8_expuniform": 20.906919930164065, "raw K1": 23.40478834900881, "raw K2_expuniform": 23.03902088626353, "raw K4_expuniform": 26.127714645302376, "raw K8_expuniform": 25.298558309772034, "raw K1_end0.999": 23.545632382157123, "raw K1_end0.97": 44.31914679061401, "raw K2_s0.5_end0.999": 22.94034238534971}, "fid_n": 2000}
|
| 8 |
+
{"step": 200000, "fid": {"EMA K1": 21.83312267758788, "EMA K2_expuniform": 20.52414311707298, "EMA K4_expuniform": 22.395691274095952, "EMA K8_expuniform": 21.00223738093041, "raw K1": 22.943934853141116, "raw K2_expuniform": 22.08319078657587, "raw K4_expuniform": 24.305031408935577, "raw K8_expuniform": 22.28491025395624, "raw K1_end0.999": 22.86619659368057, "raw K1_end0.97": 40.17960775802135, "raw K2_s0.5_end0.999": 21.93118977902276}, "fid_n": 2000}
|
| 9 |
+
{"step": 225000, "fid": {"EMA K1": 21.197790923140133, "EMA K2_expuniform": 20.932524911674875, "EMA K4_expuniform": 23.51528248096548, "EMA K8_expuniform": 21.589355812797294, "raw K1": 25.105310842884137, "raw K2_expuniform": 26.67904537823108, "raw K4_expuniform": 30.76553935363063, "raw K8_expuniform": 27.53999038533874, "raw K1_end0.999": 25.35553031186214, "raw K1_end0.97": 50.30988042007539, "raw K2_s0.5_end0.999": 26.511061003858686}, "fid_n": 2000}
|
| 10 |
+
{"step": 250000, "fid": {"EMA K1": 21.524883558180022, "EMA K2_expuniform": 21.04550109697817, "EMA K4_expuniform": 23.343791402449483, "EMA K8_expuniform": 21.509484215556654, "raw K1": 21.992022381563515, "raw K2_expuniform": 21.33692979987802, "raw K4_expuniform": 23.93654044284574, "raw K8_expuniform": 24.783339312759324, "raw K1_end0.999": 22.299128751351418, "raw K1_end0.97": 41.827522144041836, "raw K2_s0.5_end0.999": 21.268009923216255}, "fid_n": 2000}
|
| 11 |
+
{"step": 275000, "fid": {"EMA K1": 24.71720765764428, "EMA K2_expuniform": 21.12916509294479, "EMA K4_expuniform": 23.20584799235803, "EMA K8_expuniform": 21.186094123055568, "raw K1": 25.290247135321806, "raw K2_expuniform": 20.660852554136113, "raw K4_expuniform": 22.989606287621257, "raw K8_expuniform": 20.943066115989836, "raw K1_end0.999": 23.999828717752393, "raw K1_end0.97": 40.35261520108537, "raw K2_s0.5_end0.999": 20.50780039169888}, "fid_n": 2000}
|
| 12 |
+
{"step": 300000, "fid": {"EMA K1": 24.521354598394396, "EMA K2_expuniform": 20.9844966485781, "EMA K4_expuniform": 22.74540603198409, "EMA K8_expuniform": 20.87254226311461, "raw K1": 27.3087541638497, "raw K2_expuniform": 21.92752183438199, "raw K4_expuniform": 23.270188916015968, "raw K8_expuniform": 21.77846635433349, "raw K1_end0.999": 25.57709101304596, "raw K1_end0.97": 41.842692952021594, "raw K2_s0.5_end0.999": 21.728630665282935}, "fid_n": 2000}
|
| 13 |
+
{"step": 325000, "fid": {"EMA K1": 24.664765028168517, "EMA K2_expuniform": 20.707139094648085, "EMA K4_expuniform": 22.390713742944627, "EMA K8_expuniform": 20.84633306132241, "raw K1": 25.782927718130395, "raw K2_expuniform": 21.447216333953406, "raw K4_expuniform": 22.657482816750303, "raw K8_expuniform": 21.155356159738403, "raw K1_end0.999": 24.174309087598203, "raw K1_end0.97": 41.04424647088325, "raw K2_s0.5_end0.999": 21.24573590176726}, "fid_n": 2000}
|
| 14 |
+
{"step": 350000, "fid": {"EMA K1": 24.608757339232795, "EMA K2_expuniform": 20.455940764229354, "EMA K4_expuniform": 22.120770495974398, "EMA K8_expuniform": 20.68331605382417, "raw K1": 20.495150725806013, "raw K2_expuniform": 20.5532910512502, "raw K4_expuniform": 22.485500892691277, "raw K8_expuniform": 21.428729808831804, "raw K1_end0.999": 20.496643674239976, "raw K1_end0.97": 43.489373889280955, "raw K2_s0.5_end0.999": 20.408733174643714}, "fid_n": 2000}
|
| 15 |
+
{"step": 375000, "fid": {"EMA K1": 24.423529660900556, "EMA K2_expuniform": 20.719326581891607, "EMA K4_expuniform": 22.26016652246244, "EMA K8_expuniform": 20.92108936601818, "raw K1": 31.21414571013827, "raw K2_expuniform": 22.558717902984085, "raw K4_expuniform": 23.981500932152187, "raw K8_expuniform": 22.54257114228062, "raw K1_end0.999": 28.914918960635532, "raw K1_end0.97": 44.30327975592127, "raw K2_s0.5_end0.999": 22.333495304885787}, "fid_n": 2000}
|
| 16 |
+
{"step": 400000, "fid": {"EMA K1": 24.870509147723908, "EMA K2_expuniform": 20.67628021737812, "EMA K4_expuniform": 22.01211953540178, "EMA K8_expuniform": 20.844137248301877, "raw K1": 23.620060211321743, "raw K2_expuniform": 21.57324846467293, "raw K4_expuniform": 23.2881603369986, "raw K8_expuniform": 22.381274276057297, "raw K1_end0.999": 22.770201796773392, "raw K1_end0.97": 44.13017522342841, "raw K2_s0.5_end0.999": 21.407349781276366}, "fid_n": 2000}
|
| 17 |
+
{"step": 402000, "fid": {"EMA K1": 25.017519336848125, "EMA K2_expuniform": 20.69426218116638, "EMA K4_expuniform": 22.060584316407926, "EMA K8_expuniform": 20.891062502277578, "raw K1": 44.440251815348745, "raw K2_expuniform": 23.196997505597153, "raw K4_expuniform": 22.583885514179542, "raw K8_expuniform": 20.651422047664937, "raw K1_end0.999": 41.862552674693234, "raw K1_end0.97": 38.03355638965968, "raw K2_s0.5_end0.999": 22.92367040505161}, "fid_n": 2000}
|