def test_forward_pass_token_classification(self):
# initialize model with randomly initialized token classification head
model = LayoutLMForTokenClassification.from_pretrained(
"microsoft/layoutlm-base-uncased", num_labels=13).to(torch_device)
input_ids, attention_mask, bbox, token_type_ids, labels = prepare_layoutlm_batch_inputs(
)
# forward pass
outputs = model(input_ids=input_ids,
bbox=bbox,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
labels=labels)
# test the loss calculation to be around 2.65
# expected_loss = torch.tensor(2.65, device=torch_device)
# The loss is currently somewhat random and can vary between 0.1-0.3 atol.
# self.assertTrue(torch.allclose(outputs.loss, expected_loss, atol=0.1))
# test the shape of the logits
logits = outputs.logits
expected_shape = torch.Size((2, 25, 13))
self.assertEqual(logits.shape, expected_shape)
def _load_model(self):
self.model = LayoutLMForTokenClassification.from_pretrained(
'microsoft/layoutlm-base-uncased',
config=self.config,
)
self.model.resize_token_embeddings(len(self.tokenizer))
self.model.to(self.device)
def model_init():
"""Returns an initialized model for use in a Hugging Face Trainer."""
model = LayoutLMForTokenClassification.from_pretrained(
"microsoft/layoutlm-base-uncased",
num_labels=5,
from_tf = True
)
model = model.to('cuda')
return model
def create_and_check_for_token_classification(
self, config, input_ids, bbox, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = LayoutLMForTokenClassification(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.num_labels))
def __init__(self,
image_path,
model_path,
config_path,
num_labels=13,
args=None):
super(LayoutLM, self).__init__()
self.image = openImage(image_path)
self.args = args
self.tokenizer = LayoutLMTokenizer.from_pretrained(
"microsoft/layoutlm-base-uncased")
config = LayoutLMConfig.from_pretrained(config_path)
self.model = LayoutLMForTokenClassification.from_pretrained(
model_path, config=config)
self.model.to(device)
self.input_ids = None
self.attention_mask = None
self.token_type_ids = None
self.bboxes = None
self.token_actual_boxes = None
def train(dataset_path,
loader_type='combined_loader',
batch_size=4,
num_workers=16,
dataset_rand_seq=True,
dataset_rand_seq_prob=0.5,
dataset_exlude_image_mask=True,
state_dict_path=None,
weight_path='weights/extract/',
max_epoch=5,
lr=0.001,
valcheck_interval=2000,
num_gpus=1,
log_freq=100,
resume_checkpoint_path=None,
checkpoint_saved_path="checkpoints/v3/",
logs_path="logs/v3/",
prefix_name='layoutlm-v2',
manual_seed=1261):
logging.basicConfig(level=logging.INFO)
#load tokensizer
logging.info("Load BertTokenizer with indobenchmark/indobert-base-p2")
tokenizer = BertTokenizer.from_pretrained(
"indobenchmark/indobert-base-p2",
do_lower_case=True,
cache_dir=None,
)
path = dataset_path
if loader_type == 'combined_loader':
logging.info(f"Load Combined Loader with path {dataset_path}")
train_loader, valid_loader = loader.get_loader(
path,
tokenizer=tokenizer,
batch_size=batch_size,
num_workers=num_workers,
rand_seq=dataset_rand_seq,
rand_seq_prob=dataset_rand_seq_prob,
excluce_image_mask=dataset_exlude_image_mask)
else:
logging.info(f"Load Base Loader with path {dataset_path}")
train_loader, valid_loader = loader.get_base_loader(
path,
tokenizer=tokenizer,
batch_size=batch_size,
num_workers=num_workers,
rand_seq=dataset_rand_seq,
rand_seq_prob=dataset_rand_seq_prob,
excluce_image_mask=dataset_exlude_image_mask)
logging.info(f"Load LayoutLMConfig for LayoutLMForTokenClassification")
config = LayoutLMConfig.from_pretrained("microsoft/layoutlm-base-uncased",
num_labels=label_cfg.num_labels,
cache_dir=None)
logging.info(
f"Load LayoutLMForTokenClassification from_pretrained microsoft/layoutlm-base-uncased"
)
model = LayoutLMForTokenClassification.from_pretrained(
'microsoft/layoutlm-base-uncased',
config=config,
# return_dict=True
)
model.resize_token_embeddings(len(tokenizer))
if state_dict_path:
logging.info(f"Load state_dict from path {state_dict_path}")
state_dict = torch.load(state_dict_path,
map_location=torch.device("cpu"))
model.load_state_dict(state_dict)
# model = model.to(device)
#prepare the task
task = TaskLayoutLM(model, tokenizer)
# DEFAULTS used by the Trainer
checkpoint_callback = pl.callbacks.ModelCheckpoint(
dirpath=checkpoint_saved_path,
save_top_k=1,
verbose=True,
monitor='val_loss',
mode='min',
prefix=prefix_name)
tb_logger = pl_loggers.TensorBoardLogger(logs_path)
pl.trainer.seed_everything(manual_seed)
trainer = pl.Trainer(
weights_summary="top",
max_epochs=max_epoch,
val_check_interval=valcheck_interval,
gpus=num_gpus,
log_every_n_steps=log_freq,
deterministic=True,
benchmark=True,
logger=tb_logger,
checkpoint_callback=checkpoint_callback,
resume_from_checkpoint=resume_checkpoint_path,
)
trainer.fit(task, train_loader, valid_loader)
# metrics = trainer.test(task, valid_loader)
metrics = trainer.logged_metrics
#prepare to save result
# print(task._results.keys())
vacc, vloss = metrics['val_acc'], metrics['val_loss']
# tacc, tloss = metrics['trn_acc'], metrics['trn_loss']
last_epoch = metrics['epoch']
dirname = Path(dataset_path).name
filename = f'layoutlm_v2_ktp_{dirname}_vacc{vacc:.4}_vloss{vloss:.4}_epoch{last_epoch}_cli.pth'
saved_filename = str(Path(weight_path).joinpath(filename))
logging.info(f"Prepare to save training results to path {saved_filename}")
torch.save(model.state_dict(), saved_filename)