| # SpatialDiffuseSlot — Semanticist DiT-L/ViT (tok_L init) + OUR 85 multi-res | |
| # spatial tokens + OUR spatial-align attention mask. (2026-07-15 설계 확정본) | |
| # Phase 1 (this config): DiT trunk FROZEN (freeze_dit) — encoder/pool/cond-embedder | |
| # adapt first with LONG warmup. Phase 2: resume with freeze_dit=false + low blr. | |
| trainer: | |
| target: semanticist.engine.diffusion_trainer.DiffusionTrainer | |
| params: | |
| num_epoch: 400 | |
| valid_size: 64 | |
| blr: 2.5e-5 # Semanticist recipe (batch 256 -> lr 2.5e-5) | |
| cosine_lr: True | |
| warmup_epochs: 100 # 25% warmup (Semanticist recipe, 우리 기존 2%의 교훈) | |
| batch_size: 32 # per-gpu; 4 GPU -> eff 128 (large-batch regime 회피) | |
| num_workers: 8 | |
| pin_memory: True | |
| grad_accum_steps: 2 # eff 256 == Semanticist | |
| precision: 'bf16' | |
| max_grad_norm: 3.0 | |
| save_every: 10000 | |
| sample_every: 5000 | |
| fid_every: 50000 | |
| result_folder: "./output/tokenizer/models_l_spatial" | |
| log_dir: "./output/tokenizer/models_l_spatial/logs" | |
| cfg: 3.0 | |
| compile: False | |
| enable_ema: True | |
| model: | |
| target: semanticist.stage1.spatial_diffuse_slot.SpatialDiffuseSlot | |
| params: | |
| encoder: 'vit_base_patch16' | |
| enc_img_size: 256 | |
| enc_causal: True | |
| num_slots: 256 # KEEP 256: tok_L encoder loads 100% (slots unused) | |
| slot_dim: 16 | |
| norm_slots: True | |
| enable_nest: False | |
| enable_nest_after: -1 | |
| dit_model: 'DiT-L-2' | |
| vae: 'xwen99/mar-vae-kl16' | |
| num_sampling_steps: '250' | |
| use_repa: True | |
| repa_encoder_depth: 8 | |
| repa_loss_weight: 1.0 | |
| eval_fid: True | |
| fid_stats: 'fid_stats/adm_in256_stats.npz' | |
| # --- SpatialDiffuseSlot 전용 --- | |
| level_sizes: [8, 4, 2, 1] # 85 multi-res cond tokens | |
| pool_depth: 2 # 우리 L_repa attn-pool depth | |
| freeze_dit: True # Phase-1: pretrained DiT trunk frozen | |
| init_from: /NHNHOME/WORKSPACE/0226010398_A/sr_diffusion/clevr_sudoku/new_eval_spatial_reasoning0430/eval_assets/semanticist_pretrained/semanticist_tok_L.pkl | |
| ckpt_path: null # trainer가 무조건 참조 (resume용; init은 init_from) | |
| dataset: | |
| target: semanticist.utils.datasets.ImageNet | |
| params: | |
| root: ./dataset/imagenet/ | |
| split: train | |
| aug: 'randcrop' | |
| img_size: 256 | |
| test_dataset: | |
| target: semanticist.utils.datasets.ImageNet | |
| params: | |
| root: ./dataset/imagenet/ | |
| split: val | |
| aug: 'centercrop' | |
| img_size: 256 | |