def __getitem__(self, index):
"""Return dict of PyTorch tensors for preprocessed images and text and tensor of labels"""
# Text processing
x_text = self.text_inputs[index].replace('\n', ' ').replace('\r', ' ')
tokenized_x = (
['[CLS]'] +
self.tokenizer.tokenize(x_text)[:self.max_seq_length - 2] +
['[SEP]'])
input_ids = self.tokenizer.convert_tokens_to_ids(tokenized_x)
padding = [0] * (self.max_seq_length - len(input_ids))
input_ids += padding
assert len(input_ids) == self.max_seq_length
bert_text = torch.from_numpy(np.array(input_ids))
# Image processing
full_img = Image.open(self.img_inputs[index]).convert('RGB')
cropped_img = center_crop_pil_image(full_img)
full_img = self.transform(full_img)
cropped_img = self.crop_transform(cropped_img)
y_output = torch.from_numpy(np.array(self.y[index])).long()
return {
'bert_text': bert_text,
'full_img': full_img,
'crop_img': cropped_img
}, y_output
def __getitem__(self, index):
"""Return tuple of images as PyTorch tensors and and tensor of labels"""
label = self.y[index]
full_img = Image.open(self.x[index]).convert('RGB')
cropped_img = center_crop_pil_image(full_img)
full_img = self.transform(full_img)
cropped_img = self.crop_transform(cropped_img)
label = torch.FloatTensor(label)
return {'full_img': full_img, 'crop_img': cropped_img}, label
def __getitem__(self, index):
"""Return tuple of images as PyTorch tensors and and tensor of labels"""
label = self.y[index]
label = self.label_encoder.transform([label])[0]
full_img = Image.open(self.x[index]).convert('RGB')
cropped_img = center_crop_pil_image(full_img)
full_img = self.transform(full_img)
cropped_img = self.crop_transform(cropped_img)
label = torch.from_numpy(np.array(label)).long()
return {'full_img': full_img, 'crop_img': cropped_img}, label