mdrieux's picture
CIFAR-10 EDM2 large runs A and B: best EMA + final full checkpoints
b148cf6 verified
Raw History Blame Contribute Delete
3.21 kB
{
"data_root": "/work/nvme/biqs/mdrieux/cifar-hth/data",
"dataset": "cifar",
"arch": "edm2",
"edm2_sigma_data": 0.5,
"edm2_dropout": 0.13,
"base_std": 0.5,
"init_fm_ckpt": null,
"results_dir": "/work/nvme/biqs/mdrieux/edm2-large",
"run_name": "edm2_large_b512_s0",
"self_teacher": true,
"diag_adaptive_p": 0.0,
"diag_blowup_factor": 0.0,
"offdiag_target_source": "live",
"self_teacher_source": "live",
"offdiag_b_source": "teacher",
"sigma_zero_frac": 0.0,
"ema_warmup": true,
"diag_c_eps": 0.01,
"diag_target": "eqm",
"diag_only_steps": 2000,
"teacher_ckpt": null,
"model": "SiT-S/4",
"image_size": 32,
"num_classes": 10,
"student_model": null,
"objective": "shifted_meanflow",
"semi_weight": 0.0,
"idem_weight": 0.0,
"idem_on": "both",
"idem_iters": 1,
"idem_target": "self",
"idem_polish_h": 0.5,
"idem_polish_steps": 4,
"idem_polish_weight": 0.1,
"idem_ema_h": 0.05,
"cd_delta": 0.05,
"cd_substeps": 1,
"meanflow_eps": 0.01,
"meanflow_clock": "clamp",
"interp_time_sampler": "uniform",
"interp_is_correct": false,
"interp_ref": "canonical",
"importance_k": 2.0,
"interp_logitnormal": [
-0.4,
1.0
],
"eps_train": 0.001,
"sigma_min": 0.0001,
"sigma_max": 1.0,
"sigma_sampler": "uniform",
"sigma_is_correct": false,
"sigma_ref": "canonical",
"sigma_power_k": 2.0,
"sigma_logitnormal": [
-0.4,
1.0
],
"time_weight_clip": 0.0,
"guidance_mode": "branch",
"teacher_cfg_scale": 1.0,
"cfg_rule": "standard",
"cfg_cnorm": false,
"cfg_match_uncond_norm": false,
"cfg_match_guided_norm": false,
"terminal_jacobian_weight": 0.0,
"terminal_jacobian_input": "interp",
"terminal_jacobian_max_batch": 64,
"terminal_invariance_weight": 0.0,
"terminal_invariance_step": 0.1,
"terminal_invariance_max_batch": 64,
"wandb_project": "eqfm-cifar-edm2",
"wandb_entity": null,
"wandb_mode": "online",
"wandb_tags": "edm2_large",
"bucket_grad_every": 0,
"max_nonfinite_window": 20,
"diag_weight": 3072.0,
"distill_weight": 1.0,
"terminal_weight": 0.0,
"terminal_teacher_norm_weight": 0.0,
"terminal_norm_max_batch": 64,
"adaptive_weight_p": 1.0,
"adaptive_weight_eps": 3.25521e-07,
"loss_reduction": "mean",
"adaptive_weight_mode": "sample",
"cfg_dropout_prob": 0.0,
"batch_size": 512,
"cuda_graph": true,
"micro_batch": 0,
"steps": 400000,
"lr": 0.004,
"warmup_steps": 1000,
"lr_decay_from": 240000,
"lr_final": 0.0001,
"ema_snapshot_every": 25000,
"rewarmup": 0,
"resume_from": null,
"grad_clip": 0.0,
"ema_decay": 0.9999,
"log_every": 100,
"val_every": 1000000,
"val_n": 256,
"ode_ref_steps": 400,
"save_every": 2000,
"offdiag_mode": "split",
"offdiag_q": [
0.25,
0.25
],
"offdiag_q_schedule": "const",
"offdiag_q_switch_at": 0.5,
"diag_only_frac": 0.0,
"compute_budget": 0.0,
"cost_offdiag": 0.0,
"cost_both": 0.0,
"budget_milestones": [
0.25,
0.5,
0.75,
1.0
],
"resume": true,
"flowmap_clock": "linear",
"flowmap_clock_a": 0.9,
"fid_every": 25000,
"fid_n": 2000,
"fid_ref_mnist": null,
"save_every_minutes": 10.0,
"ckpt_mode": "full",
"seed": 0
}