def big_model(self):
"""Cached property means this code will only be executed once."""
checkpoint_path = "microsoft/layoutlm-large-uncased"
model = LayoutLMForMaskedLM.from_pretrained(checkpoint_path).to(
torch_device
) # test whether AutoModel can determine your model_class from checkpoint name
if torch_device == "cuda":
model.half()
def create_and_check_for_masked_lm(
self, config, input_ids, bbox, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = LayoutLMForMaskedLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, bbox, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
args = Args(args)
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
logging.basicConfig(
filename=os.path.join(args.output_dir, "train.log"),
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.addHandler(logging.StreamHandler())
if not args.test_only:
if args.load_pretrain:
model = LayoutLMForMaskedLM.from_pretrained(args.layoutlm_model,
return_dict=True)
tokenizer = LayoutLMTokenizer.from_pretrained(args.layoutlm_model)
print('Loading pre-trained model from', args.layoutlm_model)
else:
config = LayoutLMConfig.from_pretrained(args.model_name_or_path,
return_dict=True)
if args.bert_model is not None:
tokenizer = AutoTokenizer.from_pretrained(args.bert_model)
config.vocab_size = tokenizer.vocab_size
model = LayoutLMForMaskedLM(config)
if args.bert_model is None:
tokenizer = LayoutLMTokenizer.from_pretrained(args.layoutlm_model,
do_lower_case=True)
else: