def __init__(self, data_path, tokenizer, transforms, vocab, args):
df = pd.read_csv(data_path, sep='\t')[[
'gold_label', 'sentence1', 'sentence2', 'image'
]]
print(f'{data_path}, number of rows: {len(df)}')
if df['sentence2'].isnull().sum() > 0:
print(
f" drop number of lines because of missing sentence2: {df['sentence2'].isnull().sum()}"
)
df = df.loc[df['sentence2'].isnull() != True]
df = df.rename({'gold_label': 'label', 'image': 'img'}, axis=1)
df['img'] = args.img_path + '/' + df['img']
self.data = df.to_dict('records')
self.data_dir = str(os.path.dirname(data_path))
self.tokenizer = tokenizer
self.args = args
self.vocab = vocab
self.n_classes = len(args.labels)
self.text_start_token = ["[CLS]"
] if args.model != "mmbt" else ["[SEP]"]
with numpy_seed(0):
for row in self.data:
if np.random.random() < args.drop_img_percent:
row["img"] = None
self.max_seq_len = args.max_seq_len
if args.model == "mmbt":
self.max_seq_len -= args.num_image_embeds
self.transforms = transforms
def __init__(self, data_path, tokenizer, transforms, vocab, args):
self.data = [json.loads(l) for l in open(data_path)]
self.data_dir = os.path.dirname(data_path)
self.tokenizer = tokenizer
self.args = args
self.vocab = vocab
self.n_classes = len(args.labels)
self.text_start_token = ["[CLS]"
] if args.model != "mmbt" else ["[SEP]"]
with numpy_seed(0):
for row in self.data:
if np.random.random() < args.drop_img_percent:
row["img"] = None
self.max_seq_len = args.max_seq_len
if args.model == "mmbt":
self.max_seq_len -= args.num_image_embeds
self.transforms = transforms