| """ |
| Extract the CT-Chat CTViT image encoder from the upstream CT-CLIP checkpoint. |
| |
| The CT-Chat loader expects: |
| |
| CT_CHAT_MODELS / "models/CT-CLIP-Related/encoder.pth" |
| |
| to contain a bare CTViT state dict. The upstream CT-CLIP_v2.pt checkpoint stores |
| those weights under the "visual_transformer." prefix, alongside text encoder and |
| projection weights. This script strips that prefix and writes the encoder-only |
| checkpoint. |
| """ |
|
|
| import argparse |
| from pathlib import Path |
|
|
| import torch |
|
|
| from constants_and_path_utils import CT_CHAT_MODELS |
|
|
|
|
| REQUIRED_ENCODER_PREFIXES = ( |
| "spatial_rel_pos_bias.", |
| "to_patch_emb.", |
| "enc_spatial_transformer.", |
| ) |
|
|
|
|
| def looks_like_bare_encoder(state: dict) -> bool: |
| return all( |
| any(isinstance(key, str) and key.startswith(prefix) for key in state) |
| for prefix in REQUIRED_ENCODER_PREFIXES |
| ) |
|
|
|
|
| def has_visual_transformer(state: dict) -> bool: |
| return any( |
| isinstance(key, str) and key.startswith("visual_transformer.") |
| for key in state |
| ) |
|
|
|
|
| def extract_visual_transformer(state: dict) -> dict: |
| extracted = { |
| key.removeprefix("visual_transformer."): value |
| for key, value in state.items() |
| if isinstance(key, str) and key.startswith("visual_transformer.") |
| } |
|
|
| if not extracted: |
| raise ValueError( |
| "No visual_transformer.* keys found in CT-CLIP_v2.pt. " |
| "Inspect the checkpoint keys before using it as the CT-Chat encoder." |
| ) |
|
|
| if not looks_like_bare_encoder(extracted): |
| raise ValueError( |
| "Extracted visual_transformer.* keys, but the result does not look " |
| "like the CTViT encoder expected by CT-Chat." |
| ) |
|
|
| return extracted |
|
|
|
|
| def load_checkpoint(path: Path) -> dict: |
| if not path.is_file(): |
| raise FileNotFoundError(f"Checkpoint not found: {path}") |
|
|
| state = torch.load(path, map_location="cpu", weights_only=True) |
| if not hasattr(state, "keys"): |
| raise TypeError(f"Checkpoint is not dict-like: {path}") |
| return state |
|
|
|
|
| def parse_args() -> argparse.Namespace: |
| parser = argparse.ArgumentParser( |
| description="Create CT-Chat encoder.pth from CT-CLIP_v2.pt." |
| ) |
| parser.add_argument( |
| "--overwrite", |
| action="store_true", |
| help="Overwrite encoder.pth if it already exists and is not a bare encoder.", |
| ) |
| return parser.parse_args() |
|
|
|
|
| def main() -> None: |
| args = parse_args() |
|
|
| ct_chat_models = Path(CT_CHAT_MODELS) |
| checkpoint_dir = ct_chat_models / "models/CT-CLIP-Related" |
| source_path = checkpoint_dir / "CT-CLIP_v2.pt" |
| output_path = checkpoint_dir / "encoder.pth" |
|
|
| existing_state = None |
| if output_path.exists(): |
| existing_state = load_checkpoint(output_path) |
| if looks_like_bare_encoder(existing_state): |
| print(f"encoder.pth already exists and looks valid: {output_path}") |
| return |
| if not args.overwrite: |
| raise FileExistsError( |
| f"encoder.pth already exists but does not look like a bare CTViT " |
| f"encoder: {output_path}\n" |
| "Move it aside or rerun with --overwrite." |
| ) |
|
|
| if existing_state is not None and has_visual_transformer(existing_state): |
| source_state = existing_state |
| source_description = output_path |
| else: |
| source_state = load_checkpoint(source_path) |
| source_description = source_path |
|
|
| encoder_state = extract_visual_transformer(source_state) |
|
|
| output_path.parent.mkdir(parents=True, exist_ok=True) |
| torch.save(encoder_state, output_path) |
|
|
| print(f"Read CT-CLIP checkpoint: {source_description}") |
| print(f"Extracted encoder keys: {len(encoder_state)}") |
| print(f"Wrote CT-Chat encoder: {output_path}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
