scripts/convert_to_pytorch.py

"""Convert ViT and non-distilled DeiT checkpoints from the timm library."""

import argparse
from pathlib import Path

import requests
import timm
import torch
from PIL import Image
from timm.data import ImageNetInfo, infer_imagenet_subset

from transformers import DeiTImageProcessor, ViTConfig, ViTForImageClassification, ViTImageProcessor, ViTModel
from transformers.utils import logging


logging.set_verbosity_info()
logger = logging.get_logger(__name__)


# here we list all keys to be renamed (original name on the left, our name on the right)
def create_rename_keys(config, base_model=False):
    rename_keys = []
    for i in range(config.num_hidden_layers):
        # encoder layers: output projection, 2 feedforward neural networks and 2 layernorms
        rename_keys.append((f"blocks.{i}.norm1.weight", f"vit.encoder.layer.{i}.layernorm_before.weight"))
        rename_keys.append((f"blocks.{i}.norm1.bias", f"vit.encoder.layer.{i}.layernorm_before.bias"))
        rename_keys.append((f"blocks.{i}.attn.proj.weight", f"vit.encoder.layer.{i}.attention.output.dense.weight"))
        rename_keys.append((f"blocks.{i}.attn.proj.bias", f"vit.encoder.layer.{i}.attention.output.dense.bias"))
        rename_keys.append((f"blocks.{i}.norm2.weight", f"vit.encoder.layer.{i}.layernorm_after.weight"))
        rename_keys.append((f"blocks.{i}.norm2.bias", f"vit.encoder.layer.{i}.layernorm_after.bias"))
        rename_keys.append((f"blocks.{i}.mlp.fc1.weight", f"vit.encoder.layer.{i}.intermediate.dense.weight"))
        rename_keys.append((f"blocks.{i}.mlp.fc1.bias", f"vit.encoder.layer.{i}.intermediate.dense.bias"))
        rename_keys.append((f"blocks.{i}.mlp.fc2.weight", f"vit.encoder.layer.{i}.output.dense.weight"))
        rename_keys.append((f"blocks.{i}.mlp.fc2.bias", f"vit.encoder.layer.{i}.output.dense.bias"))

    # projection layer + position embeddings
    rename_keys.extend(
        [
            ("cls_token", "vit.embeddings.cls_token"),
            ("patch_embed.proj.weight", "vit.embeddings.patch_embeddings.projection.weight"),
            ("patch_embed.proj.bias", "vit.embeddings.patch_embeddings.projection.bias"),
            ("pos_embed", "vit.embeddings.position_embeddings"),
        ]
    )

    if base_model:
        # layernorm
        rename_keys.extend(
            [
                ("norm.weight", "layernorm.weight"),
                ("norm.bias", "layernorm.bias"),
            ]
        )

        # if just the base model, we should remove "vit" from all keys that start with "vit"
        rename_keys = [(pair[0], pair[1][4:]) if pair[1].startswith("vit") else pair for pair in rename_keys]
    else:
        # layernorm + classification head
        rename_keys.extend(
            [
                ("norm.weight", "vit.layernorm.weight"),
                ("norm.bias", "vit.layernorm.bias"),
                ("head.weight", "classifier.weight"),
                ("head.bias", "classifier.bias"),
            ]
        )

    return rename_keys


# we split up the matrix of each encoder layer into queries, keys and values
def read_in_q_k_v(state_dict, config, base_model=False):
    for i in range(config.num_hidden_layers):
        if base_model:
            prefix = ""
        else:
            prefix = "vit."
        # read in weights + bias of input projection layer (in timm, this is a single matrix + bias)
        in_proj_weight = state_dict.pop(f"blocks.{i}.attn.qkv.weight")
        in_proj_bias = state_dict.pop(f"blocks.{i}.attn.qkv.bias")
        # next, add query, keys and values (in that order) to the state dict
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.query.weight"] = in_proj_weight[
            : config.hidden_size, :
        ]
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.query.bias"] = in_proj_bias[: config.hidden_size]
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.key.weight"] = in_proj_weight[
            config.hidden_size : config.hidden_size * 2, :
        ]
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.key.bias"] = in_proj_bias[
            config.hidden_size : config.hidden_size * 2
        ]
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.value.weight"] = in_proj_weight[
            -config.hidden_size :, :
        ]
        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.value.bias"] = in_proj_bias[-config.hidden_size :]


def remove_classification_head_(state_dict):
    ignore_keys = ["head.weight", "head.bias"]
    for k in ignore_keys:
        state_dict.pop(k, None)


def rename_key(dct, old, new):
    val = dct.pop(old)
    dct[new] = val


# We will verify our results on an image of cute cats
def prepare_img():
    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
    im = Image.open(requests.get(url, stream=True).raw)
    return im


@torch.no_grad()
def convert_vit_checkpoint(vit_name, pytorch_dump_folder_path):
    """
    Copy/paste/tweak model's weights to our ViT structure.
    """

    # define default ViT configuration
    config = ViTConfig()
    base_model = False

    # load original model from timm
    timm_model = timm.create_model(vit_name, pretrained=True)
    timm_model.eval()

    # detect unsupported ViT models in transformers
    # fc_norm is present
    if not isinstance(getattr(timm_model, "fc_norm", None), torch.nn.Identity):
        raise ValueError(f"{vit_name} is not supported in transformers because of the presence of fc_norm.")

    # use of global average pooling in combination (or without) class token
    if getattr(timm_model, "global_pool", None) == "avg":
        raise ValueError(f"{vit_name} is not supported in transformers because of use of global average pooling.")

    # CLIP style vit with norm_pre layer present
    if "clip" in vit_name and not isinstance(getattr(timm_model, "norm_pre", None), torch.nn.Identity):
        raise ValueError(
            f"{vit_name} is not supported in transformers because it's a CLIP style ViT with norm_pre layer."
        )

    # SigLIP style vit with attn_pool layer present
    if "siglip" in vit_name and getattr(timm_model, "global_pool", None) == "map":
        raise ValueError(
            f"{vit_name} is not supported in transformers because it's a SigLIP style ViT with attn_pool."
        )

    # use of layer scale in ViT model blocks
    if not isinstance(getattr(timm_model.blocks[0], "ls1", None), torch.nn.Identity) or not isinstance(
        getattr(timm_model.blocks[0], "ls2", None), torch.nn.Identity
    ):
        raise ValueError(f"{vit_name} is not supported in transformers because it uses a layer scale in its blocks.")

    # Hybrid ResNet-ViTs
    if not isinstance(timm_model.patch_embed, timm.layers.PatchEmbed):
        raise ValueError(f"{vit_name} is not supported in transformers because it is a hybrid ResNet-ViT.")

    # get patch size and image size from the patch embedding submodule
    config.patch_size = timm_model.patch_embed.patch_size[0]
    config.image_size = timm_model.patch_embed.img_size[0]

    # retrieve architecture-specific parameters from the timm model
    config.hidden_size = timm_model.embed_dim
    config.intermediate_size = timm_model.blocks[0].mlp.fc1.out_features
    config.num_hidden_layers = len(timm_model.blocks)
    config.num_attention_heads = timm_model.blocks[0].attn.num_heads

    # check whether the model has a classification head or not
    if timm_model.num_classes != 0:
        config.num_labels = timm_model.num_classes
        # infer ImageNet subset from timm model
        imagenet_subset = infer_imagenet_subset(timm_model)
        dataset_info = ImageNetInfo(imagenet_subset)
        config.id2label = {i: dataset_info.index_to_label_name(i) for i in range(dataset_info.num_classes())}
        config.label2id = {v: k for k, v in config.id2label.items()}
    else:
        print(f"{vit_name} is going to be converted as a feature extractor only.")
        base_model = True

    # load state_dict of original model
    state_dict = timm_model.state_dict()

    # remove and rename some keys in the state dict
    if base_model:
        remove_classification_head_(state_dict)
    rename_keys = create_rename_keys(config, base_model)
    for src, dest in rename_keys:
        rename_key(state_dict, src, dest)
    read_in_q_k_v(state_dict, config, base_model)

    # load HuggingFace model
    if base_model:
        model = ViTModel(config, add_pooling_layer=False).eval()
    else:
        model = ViTForImageClassification(config).eval()
    model.load_state_dict(state_dict)

    # Check outputs on an image, prepared by ViTImageProcessor/DeiTImageProcessor
    if "deit" in vit_name:
        image_processor = DeiTImageProcessor(size=config.image_size)
    else:
        image_processor = ViTImageProcessor(size=config.image_size)
    encoding = image_processor(images=prepare_img(), return_tensors="pt")
    pixel_values = encoding["pixel_values"]
    outputs = model(pixel_values)

    if base_model:
        timm_pooled_output = timm_model.forward_features(pixel_values)
        assert timm_pooled_output.shape == outputs.last_hidden_state.shape
        assert torch.allclose(timm_pooled_output, outputs.last_hidden_state, atol=1e-1)
    else:
        timm_logits = timm_model(pixel_values)
        assert timm_logits.shape == outputs.logits.shape
        assert torch.allclose(timm_logits, outputs.logits, atol=1e-3)

    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
    print(f"Saving model {vit_name} to {pytorch_dump_folder_path}")
    model.save_pretrained(pytorch_dump_folder_path)
    print(f"Saving image processor to {pytorch_dump_folder_path}")
    image_processor.save_pretrained(pytorch_dump_folder_path)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    # Required parameters
    parser.add_argument(
        "--vit_name",
        default="vit_base_patch16_224",
        type=str,
        help="Name of the ViT timm model you'd like to convert.",
    )
    parser.add_argument(
        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
    )

    args = parser.parse_args()
    convert_vit_checkpoint(args.vit_name, args.pytorch_dump_folder_path)