modeling_resnet.py

from transformers import PreTrainedModel
from timm.models.resnet import BasicBlock, Bottleneck, ResNet
from .configuration_resnet import ResnetConfig
import torch
from typing import Dict
import timm
import os.path as osp
import os
file_path = osp.abspath(__file__)
dir_path = osp.dirname(file_path)

BLOCK_MAPPING = {
    "basic": BasicBlock,
    "bottleneck": Bottleneck,
}

class ResnetModel(PreTrainedModel):
    config_class = ResnetConfig # 用于register 不可与transformers内部的其他config_class冲突

    def __init__(self, config):
        super().__init__(config)
        block_layer = BLOCK_MAPPING[config.block_type]
        self.model = ResNet(
            block_layer,
            config.layers,
            num_classes=config.num_classes,
            in_chans=config.input_channels,
            cardinality=config.cardinality,
            base_width=config.base_width,
            stem_width=config.stem_width,
            stem_type=config.stem_type,
            avg_down=config.avg_down,
        )

    def forward(self, tensor) -> torch.Tensor: # 直接返回hidden states张量
        return self.model.forward_features(tensor)

class ResnetModelForImageClassification(PreTrainedModel):
    config_class = ResnetConfig # 用于register 不可与transformers内部的其他config_class冲突

    def __init__(self, config):
        super().__init__(config)
        block_layer = BLOCK_MAPPING[config.block_type]
        self.model = ResNet(
            block_layer,
            config.layers,
            num_classes=config.num_classes,
            in_chans=config.input_channels,
            cardinality=config.cardinality,
            base_width=config.base_width,
            stem_width=config.stem_width,
            stem_type=config.stem_type,
            avg_down=config.avg_down,
        )

    def forward(self, tensor, labels=None) -> Dict: # 返回字典类型的输出使得模型兼容transformers的Trainer API
        logits = self.model(tensor)
        if labels is not None:
            loss = torch.nn.functional.cross_entropy(logits, labels)
            return {"loss": loss, "logits": logits}
        return {"logits": logits} 


if __name__ == "__main__":
    resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)

    baremodel = ResnetModel(resnet50d_config)
    dummy_input = torch.randn(1, 3, 224, 224)
    output = baremodel(dummy_input)
    print(output.shape) 

    model = ResnetModelForImageClassification(resnet50d_config)
    dummy_input = torch.randn(1, 3, 224, 224)
    output = model(dummy_input)
    print(output)

    pretrained_model = timm.create_model('resnet50d', pretrained=True)
    model.model.load_state_dict(pretrained_model.state_dict())
    dummy_input = torch.randn(1, 3, 224, 224)
    output = model(dummy_input)
    print(output)

    # 上传(前提已使用hf auth login)
    ResnetConfig.register_for_auto_class()
    ResnetModel.register_for_auto_class("AutoModel")
    ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
    model.save_pretrained(dir_path, max_shard_size="50MB") # 默认是5GB 只有sharded后，才会生成model.safetensors.index.json文件 (会自动生成PretrainedConfig中的config.json文件)
    model.push_to_hub("custom_resnet")