(調査記録) HuggingFaceのライブラリ

HuggingFaceのライブラリの雑多な調査記録

Transformers

Q. model = transformers.CLIPModel.from_pretrained("openai/clip-vit-large-patch14")でPyTorch, or TensorFlow or Jax or Safetensorsのモデルのどれがロードされるの？

• (cite): https://huggingface.co/openai/clip-vit-large-patch14/blob/main/README.md
  • README.mdによると 次のようにしてpretrainedモデルのロードする。

  • from PIL import Image
    import requests
    from transformers import CLIPProcessor, CLIPModel
    model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
    

• (cite): https://huggingface.co/openai/clip-vit-large-patch14/tree/main
  • しかしながら、HuggingFace Hub Modelsで配布されているファイル PyTorch(pytorch_model.bin), or TensorFlow(tf_model.h5) or Jax(flax_model.msgpack) or Safetensors(model.safetensors) のどれがロードされるかわからない

A. from_tf=Trueを渡せばTensorflow、from_flax =Trueを渡せばJax、use_safetensors=Trueを渡せばSafetensors、デフォルトはPyTorch

• class CLIPModel(CLIPPreTrainedModel): (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/models/clip/modeling_clip.py#L938
  • ↓
• class CLIPPreTrainedModel(PreTrainedModel): (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/models/clip/modeling_clip.py#L399C1-L399C44
  • ↓
• class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin): (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L1223
  • ↓

• def from_pretrained( (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L2724

  • コメント:
    • from_tf (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L2793C13-L2793C20
    • from_flax (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L2796
    • use_safetensors (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L2907C13-L2907C29
  • コード:
    • 3200行目~3485行目まで (cite): https://github.com/huggingface/transformers/blob/91d155ea92da372b319a79dd4eef69533ee15170/src/transformers/modeling_utils.py#L3200

• 参考までにhuggingface/transformersが想定するファイル名は下記の通り。(cite): src/transformers/utils/__init__.py#L220C1-L234C35

  • WEIGHTS_NAME = "pytorch_model.bin"
    WEIGHTS_INDEX_NAME = "pytorch_model.bin.index.json"
    TF2_WEIGHTS_NAME = "tf_model.h5"
    TF2_WEIGHTS_INDEX_NAME = "tf_model.h5.index.json"
    TF_WEIGHTS_NAME = "model.ckpt"
    FLAX_WEIGHTS_NAME = "flax_model.msgpack"
    FLAX_WEIGHTS_INDEX_NAME = "flax_model.msgpack.index.json"
    SAFE_WEIGHTS_NAME = "model.safetensors"
    SAFE_WEIGHTS_INDEX_NAME = "model.safetensors.index.json"