Delete examples/export_to_hf.py

examples/export_to_hf.py

@@ -1,103 +0,0 @@
-#!/usr/bin/env python
-"""Export an existing checkpoint to Hugging Face format."""
-
-import sys
-from pathlib import Path
-sys.path.insert(0, str(Path(__file__).parent.parent))
-
-from LWMTemporal.models.lwm import LWMBackbone, LWMConfig
-from LWMTemporal.models.hf import LWMHFModel, LWMHFConfig
-from LWMTemporal.utils.logging import setup_logging
-
-logger = setup_logging("export_to_hf", log_dir=Path("logs"))
-
-# Path to your existing checkpoint directory (with config.json and pytorch_model.bin)
-checkpoint_dir = Path("checkpoints")  # Directory containing config.json and pytorch_model.bin
-checkpoint_path = checkpoint_dir / "pytorch_model.bin"  # Or use m18_cp.pth if that's your file
-
-# Output directory for HF export
-hf_export_dir = Path("models/hf_export")
-
-logger.info("Loading checkpoint from %s", checkpoint_path)
-# Load the LWM model
-lwm_model = LWMBackbone.from_pretrained(checkpoint_path)
-
-# Load config from checkpoint directory if it exists
-config_path = checkpoint_dir / "config.json"
-if config_path.exists():
-    import json
-    with open(config_path) as f:
-        config_dict = json.load(f)
-    lwm_config = LWMConfig.from_dict(config_dict)
-else:
-    lwm_config = lwm_model.config
-
-# Ensure max_seq_len matches checkpoint positional embeddings
-if lwm_config.max_seq_len is None and hasattr(lwm_model, "pos_embed"):
-    pos_len = int(lwm_model.pos_embed.shape[1])
-    cls_tokens = 1 if lwm_config.global_cls else 0
-    inferred = max(0, pos_len - cls_tokens)
-    if inferred > 0:
-        lwm_config.max_seq_len = inferred
-        logger.info("Inferred max_seq_len=%d from checkpoint positional embeddings", inferred)
-
-# Convert to HF format
-logger.info("Converting to Hugging Face format...")
-hf_config = LWMHFConfig(**lwm_config.to_dict())
-hf_model = LWMHFModel(hf_config)
-hf_model.backbone.load_state_dict(lwm_model.state_dict())
-
-logger.info("Exporting to Hugging Face format at %s", hf_export_dir)
-hf_model.save_pretrained(hf_export_dir)
-
-# Copy the modeling files so HF can load it with trust_remote_code=True
-# HF expects the files to match the auto_map import path
-import shutil
-base_dir = Path(__file__).parent.parent
-modeling_dir = hf_export_dir / "LWMTemporal" / "models"
-modeling_dir.mkdir(parents=True, exist_ok=True)
-
-# Copy hf.py (the HF wrapper)
-hf_file = base_dir / "LWMTemporal" / "models" / "hf.py"
-if hf_file.exists():
-    shutil.copy2(hf_file, modeling_dir / "hf.py")
-    logger.info("✓ Copied hf.py")
-else:
-    logger.warning("hf.py not found at %s", hf_file)
-
-# Copy lwm.py (dependency)
-lwm_file = base_dir / "LWMTemporal" / "models" / "lwm.py"
-if lwm_file.exists():
-    shutil.copy2(lwm_file, modeling_dir / "lwm.py")
-    logger.info("✓ Copied lwm.py")
-else:
-    logger.warning("lwm.py not found at %s", lwm_file)
-
-# Create __init__.py files for proper imports
-(hf_export_dir / "LWMTemporal" / "__init__.py").touch()
-(modeling_dir / "__init__.py").touch()
-
-logger.info("✓ Exported to %s", hf_export_dir)
-logger.info("Files created:")
-for f in sorted(hf_export_dir.glob("*")):
-    logger.info("  - %s", f.name)
-
-# Optional: Upload directly to HF Hub
-# Uncomment to automatically push:
-# try:
-#     from huggingface_hub import HfApi
-#     api = HfApi()
-#     api.upload_folder(
-#         folder_path=hf_export_dir,
-#         repo_id="wi-lab/lwm-temporal",
-#         repo_type="model",
-#         commit_message="Export existing checkpoint to HF format",
-#     )
-#     logger.info("✓ Uploaded to Hugging Face Hub: wi-lab/lwm-temporal")
-# except ImportError:
-#     logger.warning("huggingface_hub not installed; skipping upload")
-#     logger.info("To upload manually:")
-#     logger.info("  1. git clone https://huggingface.co/wi-lab/lwm-temporal")
-#     logger.info("  2. cp -r %s/* lwm-temporal/", hf_export_dir)
-#     logger.info("  3. cd lwm-temporal && git add . && git commit -m 'Add checkpoint' && git push")
-