Upload small.yaml with huggingface_hub

small.yaml

@@ -0,0 +1,87 @@
+# Small config: fast experiments, single GPU
+input:
+  num_surface_points: 4096
+
+tokenizer:
+  latent:
+    latent_shape: [24, 24, 24]
+    token_dim: 128
+  neighbor:
+    multiscale_radii: [0.05, 0.1, 0.2]
+  agno:
+    num_heads: 4
+
+processor:
+  hidden_size: 128
+  num_layers: 3
+  num_heads: 4
+  num_kv_heads: 2
+  patch_size: 6
+
+heads:
+  embedding_dim: 128
+  pooling: attention
+  symmetry:
+    hidden_dim: 128
+  primitive:
+    hidden_dim: 128
+  part:
+    hidden_dim: 128
+  caption:
+    hidden_dim: 128
+  reduction:
+    hidden_dim: 128
+
+train:
+  epochs: 50
+  batch_size: 16
+  gradient_accumulation_steps: 2
+  mixed_precision: bf16
+  num_workers: 4
+  pin_memory: true
+  # Heavy outputs live on /data to keep the root disk free.
+  # Can also be overridden via SHAPE_CHECKPOINT_DIR / SHAPE_LOG_DIR env vars.
+  checkpoint_dir: /data/shape-v2/checkpoints
+  log_dir: /data/shape-v2/runs
+  optimizer:
+    lr: 3.0e-4
+    warmup_steps: 500
+  loss:
+    inpainting:
+      enabled: false
+    # Regression loss used by masked-token reconstruction, inpainting
+    # reconstruction, and symmetry plane/axis regression heads.
+    # Classification losses (symmetry class, primitive, part, reduction)
+    # are unaffected.
+    regression:
+      kind: smooth_l1   # "mse" | "smooth_l1"
+      beta: 1.0
+    # Self-supervised only. Supervised heads inherit weight=0.0 from the
+    # defaults because the stock synthetic labels overfit catastrophically
+    # (val/symmetry ≈ 2.5 while train/symmetry ≈ 1e-4). Re-enable in a
+    # follow-up config once the labels in data/synthetic_labels.py are
+    # fixed to generalize across unseen meshes.
+    weights:
+      masked_token: 1.0
+      contrastive: 0.2
+      inpainting: 0.0
+  wandb:
+    enabled: true
+    project: shape-foundation
+    tags: [small, pretrain, v1-fix]
+
+data:
+  # Deterministic hash-based train/val split applied across every source.
+  # Every file's assignment is stable across runs and ranks (md5 of path).
+  # Set to 0.0 to opt out and fall back to per-source `split:` entries.
+  val_fraction: 0.05
+  sources:
+    - name: thingi10k
+      root: data_cache/thingi10k/thingi10k
+      weight: 1.0
+    - name: mfcad
+      root: data_cache/mfcad/mfcad
+      weight: 1.0
+    - name: fusion360
+      root: data_cache/fusion360/fusion360
+      weight: 0.5