def func(batch):
text, imgs, y = [], [], []
for (_text, _id), _y in batch:
text.append([_text])
imgs.append(load_vgg(_id)) # _imgs: List[4096, 4096, 4096]
y.append(_y)
padded_text, lens, sents_num = padding_text_2d(text) # padding文本
imgs = torch.tensor(imgs) # 转为张量
y = torch.tensor(y) #
return (padded_text, lens, sents_num, imgs), y
def func(batch):
text, imgs, y = [], [], []
for (_text, _imgs), _y in batch:
text.append(_text)
imgs.append(mean.copy()) # 空白图片
imgs += _imgs # _imgs: List[4096, 4096, 4096]
y.append(_y)
padded_text, lens, sents_num = padding_text_2d(text) # padding文本
imgs = torch.tensor(imgs) # 转为张量
y = torch.tensor(y) #
return (padded_text, lens, sents_num, imgs), y
def func(batch):
text, imgs, y = [], [], []
for (_text, _imgs), _y in batch:
text.append(_text)
imgs += _imgs # _imgs: List[4096, 4096, 4096]
y.append(_y)
padded_text, lens, sents_num = padding_text_2d(
text, config["max_tokens"], config["max_sents"]) # padding文本
imgs = torch.tensor(imgs)
y = torch.tensor(y)
return (padded_text, lens, sents_num, imgs), y
def func(batch):
text, imgs, y = [], [], []
for (_text, _imgs), _y in batch:
text.append(_text)
imgs.append(mean.copy()) # 空白图片
for i in _imgs:
feature = load_vgg_features(i)
if feature is not None:
imgs.append(feature)
y.append(_y)
padded_text, lens, sents_num = padding_text_2d(text) # padding文本
imgs = torch.tensor(imgs) # 转为张量
y = torch.tensor(y) #
return (padded_text, lens, sents_num, imgs), y
def electra_vgg_collate_2d(batch):
text, imgs, y = [], [], []
for (_text, _imgs), _y in batch:
text.append(_text)
imgs += _imgs # _imgs: List[4096, 4096, 4096]
y.append(_y)
padded_text, lens, sents_num = padding_text_2d(
text, config["max_tokens"], config["max_sents"]) # padding文本
mask = torch.zeros(padded_text.shape, dtype=torch.float32) #
for i, l in enumerate(lens): # mask
mask[i][:l] = 1
token_type_ids = torch.zeros(padded_text.shape, dtype=torch.long)
position_ids = torch.zeros(padded_text.shape, dtype=torch.long)
imgs = torch.tensor(imgs) # 转为张量
y = torch.tensor(y) #
return (padded_text, mask, token_type_ids, position_ids, sents_num,
imgs), y