Update README.md

README.md

@@ -45,48 +45,67 @@ Supported backbones: **CLIP RN50**, **CLIP ViT-B/32**, **CLIP ViT-L/14**, **SigL
 
 ## Repository Layout
 
+Each dataset follows the same two-level structure: top-level `embeddings/` holds the **full** feature dump (all classes), and `embeddings/splits/` holds the **seen/unseen partition** used throughout the paper.
+
 ```
 data/
 ├── CIFAR-100/
-│   ├── config/
-│   │   ├── cifar100_classes.txt      # class names
-│   │   └── cifar100_filtered.txt     # concept vocabulary
 │   └── embeddings/
+│       ├── RN50_train_embeddings.pt                       # (N_train, d) - all classes
+│       ├── RN50_test_embeddings.pt                        # (N_test,  d) - all classes
+│       ├── ViT-B-32_train_embeddings.pt
+│       ├── ViT-B-32_test_embeddings.pt
+│       ├── ViT-L-14_train_embeddings.pt
+│       ├── ViT-L-14_test_embeddings.pt
+│       ├── siglip-so400m-patch14-384_train_embeddings.pt
+│       ├── siglip-so400m-patch14-384_test_embeddings.pt
+│       ├── train_ids.pt                                   # (N_train,)  - class indices
+│       ├── test_ids.pt                                    # (N_test,)   - class indices
 │       └── splits/
-│           ├── class_split.pt                     # {seen_classes, unseen_classes}
-│           ├── RN50_seen_train_embs.pt            # (N_train_seen, d)
-│           ├── RN50_seen_test_embs.pt             # (N_test_seen, d)
-│           ├── RN50_unseen_test_embs.pt           # (N_test_unseen, d)
-│           ├── seen_test_ids.pt                   # (N_test_seen,)
-│           ├── unseen_test_ids.pt                 # (N_test_unseen,)
+│           ├── class_split.pt                             # {seen_classes, unseen_classes}
+│           ├── RN50_seen_train_embs.pt
+│           ├── RN50_seen_test_embs.pt
+│           ├── RN50_unseen_train_embs.pt
+│           ├── RN50_unseen_test_embs.pt
 │           ├── ViT-B-32_seen_train_embs.pt
-│           ├── ...
-│           └── ViT-SO400M-14-SigLIP-384_unseen_test_embs.pt
+│           ├── ViT-B-32_seen_test_embs.pt
+│           ├── ViT-B-32_unseen_train_embs.pt
+│           ├── ViT-B-32_unseen_test_embs.pt
+│           ├── ViT-L-14_seen_train_embs.pt
+│           ├── ViT-L-14_seen_test_embs.pt
+│           ├── ViT-L-14_unseen_train_embs.pt
+│           ├── ViT-L-14_unseen_test_embs.pt
+│           ├── siglip-so400m-patch14-384_seen_train_embs.pt
+│           ├── siglip-so400m-patch14-384_seen_test_embs.pt
+│           ├── siglip-so400m-patch14-384_unseen_train_embs.pt
+│           ├── siglip-so400m-patch14-384_unseen_test_embs.pt
+│           ├── seen_train_ids.pt
+│           ├── seen_test_ids.pt
+│           ├── unseen_train_ids.pt
+│           └── unseen_test_ids.pt
 ├── CUB-200-2011/
-│   ├── config/
-│   │   ├── cub_classes.txt
-│   │   └── cub_filtered.txt
-│   └── embeddings/splits/ ...
+│   └── embeddings/ ...
 ├── ImageNet/
-│   ├── config/
-│   │   ├── class_mapping.json
-│   │   ├── imagenet_classes.txt
-│   │   └── imagenet_filtered.txt
-│   └── embeddings/splits/ ...
+│   └── embeddings/ ...
 ├── ImageNet-100/
-│   ├── config/
-│   │   ├── Labels.json
-│   │   ├── imagenet100_classes.txt
-│   │   └── imagenet_filtered.txt
-│   └── embeddings/splits/ ...
+│   └── embeddings/ ...
 └── Places365/
-    ├── config/
-    │   ├── categories_places365_clean.txt
-    │   └── places365_filtered.txt
-    └── embeddings/splits/ ...
+    └── embeddings/ ...
 ```
 
-All `.pt` files are standard PyTorch tensors saved with `torch.save`. Embedding tensors have shape `(N, d)` where `d` is the backbone output dimension (1024 for RN50, 512 for ViT-B/32, 768 for ViT-L/14, 1152 for SigLIP SO400M). Target tensors are `(N,)` `int64` class indices.
+**File conventions:**
+
+| File | Shape | Dtype | Description |
+|:--|:--|:--|:--|
+| `{backbone}_train_embeddings.pt` / `{backbone}_test_embeddings.pt` | `(N, d)` | `float32` | Full feature dump across all classes |
+| `train_ids.pt` / `test_ids.pt` | `(N,)` | `int64` | Class indices aligned with the full embeddings above |
+| `splits/class_split.pt` | `dict` | - | `{"seen_classes": [...], "unseen_classes": [...]}` |
+| `splits/{backbone}_seen_train_embs.pt` etc. | `(N', d)` | `float32` | Embeddings restricted to seen/unseen classes |
+| `splits/seen_train_ids.pt` etc. | `(N',)` | `int64` | Class indices for the corresponding split |
+
+Embedding dimensions: **1024** (RN50), **512** (ViT-B/32), **768** (ViT-L/14), **1152** (SigLIP SO400M/14).
+
+All `.pt` files are standard PyTorch tensors saved with `torch.save` and can be loaded with `torch.load(..., weights_only=True)`.
 
 ## Quickstart