image_processing_resnet.py

import os

import albumentations as A
from albumentations.pytorch.transforms import ToTensorV2
import PIL.Image
import numpy as np
from functools import partial
from typing import Dict, List, Optional, Union
from datasets import load_dataset, DatasetDict, Image  # pip install datasets
from torch.utils.data import DataLoader
import torch

# 自定义 ImageProcessor 为了与 pipeline使用
from transformers import ViTImageProcessor
from transformers.image_utils import PILImageResampling, ChannelDimension
from transformers.image_processing_utils import get_size_dict

class ResnetImageProcessor(ViTImageProcessor):
    """
    >>> # tfs = A.Compose([A.Resize(256, 256), A.CenterCrop(224, 224)])
    >>> # 如果传入 参数 tfs=tfs 在调用save_pretrained会报错
    >>> # 本地使用
    >>> mean = [0.485, 0.456, 0.406];std = [0.229, 0.224, 0.225]
    >>> image_processor = ResnetImageProcessor(size=(224, 224), image_mean=mean, image_std=std)
    >>> image_processor.save_pretrained("custom-resnet")
    >>> image_processor = ResnetImageProcessor.from_pretrained("custom-resnet")

    >>> # push_to_hub
    >>> # hub登录
    >>> from huggingface_hub import notebook_login;notebook_login()
    >>> # or huggingface-cli login

    >>> ResnetImageProcessor.register_for_auto_class()
    >>> mean = [0.485, 0.456, 0.406];std = [0.229, 0.224, 0.225]
    >>> image_processor = ResnetImageProcessor(size=(224, 224), image_mean=mean, image_std=std)
    >>> image_processor.save_pretrained("custom-resnet")
    >>> # image_processor = ResnetImageProcessor.from_pretrained("custom-resnet")
    >>> # 如果要执行 push_to_hub 需要将 custom-resnet/preprocessor_config.json 中的 "image_processor_type" 改成 "ViTImageProcessor"
    >>> # 默认的 ResnetImageProcessor 没有注册到 AutoImageProcessor
    >>> # 否则从 使用 AutoImageProcessor 加载 会报错了
    >>> image_processor.push_to_hub('custom-resnet')

    >>> # 从 huggingface hub 加载
    >>> from transformers import AutoImageProcessor
    >>> AutoImageProcessor.register(config_class='wucng/custom-resnet/config.json',image_processor_class=ResnetImageProcessor)
    >>> image_processor = AutoImageProcessor.from_pretrained('wucng/custom-resnet', trust_remote_code=True)
    """

    def resize(
            self,
            image: np.ndarray,
            size: Dict[str, int],
            resample: PILImageResampling = PILImageResampling.BILINEAR,
            data_format: Optional[Union[str, ChannelDimension]] = None,
            input_data_format: Optional[Union[str, ChannelDimension]] = None,
            **kwargs,
    ) -> np.ndarray:
        size = get_size_dict(size)
        output_size = (size["height"], size["width"])
        height, width = size["height"], size["width"]

        tfs = kwargs.get('tfs', None)
        if tfs is None:
            ratio = 256 / 224
            tfs = A.Compose([A.Resize(int(ratio * height), int(ratio * width)), A.CenterCrop(height, width)])
        return tfs(image=image)['image']