| """Phase-2 RAEv2 ISIC configs: extend the FM ladder under RAEv2. |
| dinov3l-k7 : DINOv3-L multi-layer sum (7 layers) — completes K1<K7<K23 dose-response |
| mael : MAE-L (reconstruction-objective FM; old best family) |
| dinov2l : DINOv2-L single-layer (generic discriminative FM) |
| All latent 1024x16x16 -> same DiT as dinov3l-k1. Same budget (gb256/accum4, 200ep) = fair vs Phase-1. |
| """ |
| import yaml, os |
|
|
| REPO = "/data/temp/qinshengqian/c3/Code/RAEv2" |
| DATA = "/data/temp/qinshengqian/c3/data/isic-arrow-256" |
| NUM_CLASSES = 8 |
| ENCODERS = { |
| "dinov3l-k7": "dinov3mls-vit-l16[layers=11.13.15.17.19.21.23]", |
| "mael": "mae-vit-l", |
| "dinov2l": "dinov2-vit-l", |
| } |
|
|
| def dataset_block(): |
| return {"target": "imagenet", "type": "hf", "data_dir": DATA, |
| "split": "train", "condition_type": "label", "shared_tmpdir": "~/tmp"} |
|
|
| for tag, enc in ENCODERS.items(): |
| s1dir = f"results/stage1/ISIC_{tag}" |
| |
| s1 = { |
| "stage_1": {"target": "stage1.RAE", "params": { |
| "encoder_name": enc, "resolution": 256, |
| "decoder_config_path": "configs/decoder/ViTXL", "noise_tau": 0.8}}, |
| "dataset": dataset_block(), |
| "training": {"epochs": 40, "ema_decay": 0.9978, "global_batch_size": 64, |
| "num_workers": 4, "clip_grad": 0.0, "log_interval": 50, |
| "checkpoint_interval": 5, "sample_every": 100000, "image_size": 256, |
| "optimizer": {"lr": 2.0e-4, "betas": [0.9, 0.95], "weight_decay": 0.0}, |
| "scheduler": {"type": "cosine", "warmup_epochs": 1, "decay_end_epoch": 40, |
| "base_lr": 2.0e-4, "final_lr": 2.0e-5, "warmup_from_zero": True}}, |
| "eval": {"eval_interval": 100000, "eval_model": False, "eval_dir": "results/stage1/eval", "datasets": {}}, |
| "gan": {"arch": {"dino_ckpt_path": REPO + "/pretrained_models/encoders/dino/dino_vit_small_patch8_224.pth", |
| "ks": 9, "norm_type": "bn", "using_spec_norm": True, "recipe": "S_8"}, |
| "optimizer": {"lr": 2.0e-4, "betas": [0.9, 0.95], "weight_decay": 0.0}, |
| "scheduler": {"type": "cosine", "warmup_epochs": 1, "decay_end_epoch": 40, |
| "base_lr": 2.0e-4, "final_lr": 2.0e-5, "warmup_from_zero": True}, |
| "augment": {"prob": 1.0, "cutout": 0.0}, |
| "loss": {"disc_loss": "hinge", "gen_loss": "vanilla", "disc_weight": 0.75, |
| "perceptual_weight": 1.0, "disc_start": 20, "disc_upd_start": 15, |
| "lpips_start": 0, "max_d_weight": 10000.0, "disc_updates": 1}}, |
| } |
| os.makedirs(f"{REPO}/configs/stage1/training/ISIC", exist_ok=True) |
| yaml.safe_dump(s1, open(f"{REPO}/configs/stage1/training/ISIC/{tag}.yaml", "w"), sort_keys=False) |
| |
| s2 = { |
| "stage_1": {"target": "stage1.RAE", "params": { |
| "encoder_name": enc, "resolution": 256, |
| "decoder_config_path": "configs/decoder/ViTXL", |
| "pretrained_decoder_path": f"{s1dir}/decoder_ema.pt", |
| "noise_tau": 0.0, |
| "normalization_stat_path": f"{s1dir}/stats.pt"}}, |
| "stage_2": {"target": "stage2.models.DDT.DiTwDDTHeadIG", "params": { |
| "input_size": 16, "patch_size": [1, 1], "in_channels": 1024, |
| "hidden_size": [1440, 2048], "depth": [28, 2], "num_heads": [20, 16], |
| "mlp_ratio": 4.0, "base_model_depth": 8}}, |
| "conditioning": {"type": "label", "cfg_dropout_prob": 0.1, |
| "arch": {"num_t_tokens": 4, "num_c_tokens": 8}}, |
| "transport": {"prediction": "x", "time_dist_type": "logit-normal_0_1"}, |
| "sampler": {"num_steps": 50}, |
| "guidance": {"cfg": {"scale": 1.0, "t_min": 0.0, "t_max": 1.0}}, |
| "dataset": dataset_block(), |
| "training": {"epochs": 200, "global_batch_size": 256, "grad_accum_steps": 4, |
| "ema_decay": 0.9995, "num_workers": 4, "log_interval": 50, |
| "checkpoint_interval": 10, "sample_every": 100000, "clip_grad": 1.0, "global_seed": 42, |
| "optimizer": {"type": "gmuon", "lr": 0.0002, "momentum": 0.95, "nesterov": True, |
| "ns_coefficients_preset": "POLAR_EXPRESS_COEFFICIENTS", |
| "ns_use_kernels": False, "weight_decay": 0.0}, |
| "scheduler": {"type": "linear", "warmup_epochs": 10, "decay_end_epoch": 150, |
| "base_lr": 0.0002, "final_lr": 2.0e-05, "warmup_from_zero": False}, |
| "image_size": 256}, |
| "eval": {"eval_interval": 100000, "eval_model": False, "eval_dir": "results/stage2/eval", "datasets": {}}, |
| "misc": {"latent_size": [1024, 16, 16], "num_classes": NUM_CLASSES, |
| "time_dist_shift_dim": 1024 * 16 * 16, "time_dist_shift_base": 4096}, |
| } |
| os.makedirs(f"{REPO}/configs/stage2/training/ISIC", exist_ok=True) |
| yaml.safe_dump(s2, open(f"{REPO}/configs/stage2/training/ISIC/{tag}.yaml", "w"), sort_keys=False) |
| print("wrote phase2 configs:", tag, "->", enc) |
| print("PHASE2_CONFIGS_DONE") |
|
|
