def __init__(self, config, pretrained=True):
self.config = config
self.model, self.vocab = build_model(config)
self.device = config['device']
self.num_iters = config['trainer']['iters']
self.beamsearch = config['predictor']['beamsearch']
self.data_root = config['dataset']['data_root']
self.train_annotation = config['dataset']['train_annotation']
self.valid_annotation = config['dataset']['valid_annotation']
self.dataset_name = config['dataset']['name']
self.batch_size = config['trainer']['batch_size']
self.print_every = config['trainer']['print_every']
self.valid_every = config['trainer']['valid_every']
self.checkpoint = config['trainer']['checkpoint']
self.export_weights = config['trainer']['export']
self.metrics = config['trainer']['metrics']
logger = config['trainer']['log']
if logger:
self.logger = Logger(logger)
if pretrained:
weight_file = download_weights(**config['pretrain'],
quiet=config['quiet'])
self.load_weights(weight_file)
self.iter = 0
self.optimizer = AdamW(self.model.parameters(),
betas=(0.9, 0.98),
eps=1e-09)
self.scheduler = OneCycleLR(self.optimizer, **config['optimizer'])
# self.optimizer = ScheduledOptim(
# Adam(self.model.parameters(), betas=(0.9, 0.98), eps=1e-09),
# #config['transformer']['d_model'],
# 512,
# **config['optimizer'])
self.criterion = LabelSmoothingLoss(len(self.vocab),
padding_idx=self.vocab.pad,
smoothing=0.1)
transforms = ImgAugTransform()
self.train_gen = self.data_gen('train_{}'.format(self.dataset_name),
self.data_root,
self.train_annotation,
transform=transforms)
if self.valid_annotation:
self.valid_gen = self.data_gen(
'valid_{}'.format(self.dataset_name), self.data_root,
self.valid_annotation)
self.train_losses = []
def __init__(self, config, pretrained=True, augmentor=ImgAugTransform()):
self.config = config
self.model, self.vocab = build_model(config)
self.device = config['device']
self.num_iters = config['trainer']['iters']
self.beamsearch = config['predictor']['beamsearch']
self.data_root = config['dataset']['data_root']
self.train_annotation = config['dataset']['train_annotation']
self.valid_annotation = config['dataset']['valid_annotation']
self.train_lmdb = config['dataset']['train_lmdb']
self.valid_lmdb = config['dataset']['valid_lmdb']
self.dataset_name = config['dataset']['name']
self.batch_size = config['trainer']['batch_size']
self.print_every = config['trainer']['print_every']
self.valid_every = config['trainer']['valid_every']
self.image_aug = config['aug']['image_aug']
self.masked_language_model = config['aug']['masked_language_model']
self.metrics = config['trainer']['metrics']
self.is_padding = config['dataset']['is_padding']
self.tensorboard_dir = config['monitor']['log_dir']
if not os.path.exists(self.tensorboard_dir):
os.makedirs(self.tensorboard_dir, exist_ok=True)
self.writer = SummaryWriter(self.tensorboard_dir)
# LOGGER
self.logger = Logger(config['monitor']['log_dir'])
self.logger.info(config)
self.iter = 0
self.best_acc = 0
self.scheduler = None
self.is_finetuning = config['trainer']['is_finetuning']
if self.is_finetuning:
self.logger.info("Finetuning model ---->")
if self.model.seq_modeling == 'crnn':
self.optimizer = Adam(lr=0.0001,
params=self.model.parameters(),
betas=(0.5, 0.999))
else:
self.optimizer = AdamW(lr=0.0001,
params=self.model.parameters(),
betas=(0.9, 0.98),
eps=1e-09)
else:
self.optimizer = AdamW(self.model.parameters(),
betas=(0.9, 0.98),
eps=1e-09)
self.scheduler = OneCycleLR(self.optimizer,
total_steps=self.num_iters,
**config['optimizer'])
if self.model.seq_modeling == 'crnn':
self.criterion = torch.nn.CTCLoss(self.vocab.pad,
zero_infinity=True)
else:
self.criterion = LabelSmoothingLoss(len(self.vocab),
padding_idx=self.vocab.pad,
smoothing=0.1)
# Pretrained model
if config['trainer']['pretrained']:
self.load_weights(config['trainer']['pretrained'])
self.logger.info("Loaded trained model from: {}".format(
config['trainer']['pretrained']))
# Resume
elif config['trainer']['resume_from']:
self.load_checkpoint(config['trainer']['resume_from'])
for state in self.optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(torch.device(self.device))
self.logger.info("Resume training from {}".format(
config['trainer']['resume_from']))
# DATASET
transforms = None
if self.image_aug:
transforms = augmentor
train_lmdb_paths = [
os.path.join(self.data_root, lmdb_path)
for lmdb_path in self.train_lmdb
]
self.train_gen = self.data_gen(
lmdb_paths=train_lmdb_paths,
data_root=self.data_root,
annotation=self.train_annotation,
masked_language_model=self.masked_language_model,
transform=transforms,
is_train=True)
if self.valid_annotation:
self.valid_gen = self.data_gen(
lmdb_paths=[os.path.join(self.data_root, self.valid_lmdb)],
data_root=self.data_root,
annotation=self.valid_annotation,
masked_language_model=False)
self.train_losses = []
self.logger.info("Number batch samples of training: %d" %
len(self.train_gen))
self.logger.info("Number batch samples of valid: %d" %
len(self.valid_gen))
config_savepath = os.path.join(self.tensorboard_dir, "config.yml")
if not os.path.exists(config_savepath):
self.logger.info("Saving config file at: %s" % config_savepath)
Cfg(config).save(config_savepath)
def __init__(self, config, pretrained=True):
self.config = config
self.model, self.vocab = build_model(config)
self.device = config['device']
self.num_iters = config['trainer']['iters']
self.beamsearch = config['predictor']['beamsearch']
self.data_root = config['dataset']['data_root']
self.train_annotation = config['dataset']['train_annotation']
self.valid_annotation = config['dataset']['valid_annotation']
self.dataset_name = config['dataset']['name']
self.batch_size = config['trainer']['batch_size']
self.print_every = config['trainer']['print_every']
self.valid_every = config['trainer']['valid_every']
self.checkpoint = config['trainer']['checkpoint']
self.export_weights = config['trainer']['export']
self.metrics = config['trainer']['metrics']
logger = config['trainer']['log']
if logger:
self.logger = Logger(logger)
if pretrained:
download_weights(**config['pretrain'], quiet=config['quiet'])
state_dict = torch.load(config['pretrain']['cached'],
map_location=torch.device(self.device))
for name, param in self.model.named_parameters():
if state_dict[name].shape != param.shape:
print('{} missmatching shape'.format(name))
del state_dict[name]
self.model.load_state_dict(state_dict, strict=False)
self.iter = 0
self.optimizer = ScheduledOptim(
Adam(self.model.parameters(), betas=(0.9, 0.98), eps=1e-09),
config['transformer']['d_model'], **config['optimizer'])
# self.criterion = nn.CrossEntropyLoss(ignore_index=0)
self.criterion = LabelSmoothingLoss(len(self.vocab),
padding_idx=self.vocab.pad,
smoothing=0.1)
# transforms = torchvision.transforms.Compose([
# torchvision.transforms.ColorJitter(brightness=.1, contrast=.1, hue=.1, saturation=.1),
# torchvision.transforms.RandomAffine(degrees=0, scale=(3/4, 4/3))
# ])
transforms = ImgAugTransform()
self.train_gen = self.data_gen('train_{}'.format(self.dataset_name),
self.data_root,
self.train_annotation,
transform=transforms)
if self.valid_annotation:
self.valid_gen = self.data_gen(
'valid_{}'.format(self.dataset_name), self.data_root,
self.valid_annotation)
self.train_losses = []
def __init__(self, config, pretrained=False, augmentor=ImgAugTransform()):
self.config = config
self.model, self.vocab = build_model(config)
# ipdb.set_trace()
self.device = config['device']
self.num_iters = config['trainer']['iters']
self.beamsearch = config['predictor']['beamsearch']
self.data_root = config['dataset']['data_root']
self.train_annotation = config['dataset']['train_annotation']
self.valid_annotation = config['dataset']['valid_annotation']
self.dataset_name = config['dataset']['name']
self.batch_size = config['trainer']['batch_size']
self.print_every = config['trainer']['print_every']
self.valid_every = config['trainer']['valid_every']
self.image_aug = config['aug']['image_aug']
self.masked_language_model = config['aug']['masked_language_model']
# self.checkpoint = config['trainer']['checkpoint']
self.export_weights = config['trainer']['export']
self.metrics = config['trainer']['metrics']
logger = config['trainer']['log']
if logger:
self.logger = Logger(logger)
if pretrained:
logging.info(f"Load weights from {config['weights']}....")
# weight_file = download_weights(**config['pretrain'], quiet=config['quiet'])
self.load_weights(config['weights'])
self.iter = 0
self.optimizer = AdamW(self.model.parameters(),
betas=(0.9, 0.98),
eps=1e-09)
self.scheduler = OneCycleLR(self.optimizer,
total_steps=self.num_iters,
**config['optimizer'])
# self.optimizer = ScheduledOptim(
# Adam(self.model.parameters(), betas=(0.9, 0.98), eps=1e-09),
# #config['transformer']['d_model'],
# 512,
# **config['optimizer'])
self.criterion = LabelSmoothingLoss(len(self.vocab),
padding_idx=self.vocab.pad,
smoothing=0.1)
transforms = None
if self.image_aug:
transforms = augmentor
self.train_gen = self.data_gen('train_{}'.format(self.dataset_name),
self.data_root,
self.train_annotation,
self.masked_language_model,
transform=transforms)
if self.valid_annotation:
self.valid_gen = self.data_gen('valid_{}'.format(
self.dataset_name),
self.data_root,
self.valid_annotation,
masked_language_model=False)
self.train_losses = []