pair_classification.py

from torchvision import transforms
from transformers import ImageClassificationPipeline
import torch


class PairClassificationPipeline(ImageClassificationPipeline):
    pipe_to_tensor = transforms.ToTensor()
    pipe_to_pil = transforms.ToPILImage()

    def preprocess(self, image):       
        left_image, right_image = self.horizontal_split_image(image)
        model_inputs = self.extract_split_feature(left_image, right_image)
        # model_inputs = super().preprocess(image)
        # print(model_inputs['pixel_values'].shape)
        return model_inputs

    def horizontal_split_image(self, image):
        # image = image.resize((448,224))
        w, h = image.size
        half_w = w//2
        left_image = image.crop([0,0,half_w,h])
        right_image = image.crop([half_w,0,2*half_w,h])
        return left_image, right_image
    
    def extract_split_feature(self, left_image, right_image):
        model_inputs = self.feature_extractor(images=left_image, return_tensors=self.framework)
        right_inputs = self.feature_extractor(images=right_image, return_tensors=self.framework)
        model_inputs['pixel_values'] = torch.cat([model_inputs['pixel_values'],right_inputs['pixel_values']], dim=1)
        return model_inputs