def test_filename(tmpdir):
with tmpdir.as_cwd():
filepath = "test"
log.config(filepath=filepath)
log.info("test!")
log.clear()
assert tmpdir.join(filepath).exists()
def common_main(args: Dict[str, Any]) -> Dict[str, Any]:
del args["config"]
# configure logging
logging = args.pop("logging")
if logging["filepath"] is not None:
logging["filepath"] = join(
args["common"].experiment_dirpath, logging["filepath"]
)
log.config(**logging)
log.info(f"Arguments: {args}")
versions = get_installed_versions()
if versions:
log.info("Installed:", versions)
return args
def run(
syms: str,
fixed_input_height: Optional[NonNegativeInt] = 0,
adaptive_pooling: str = "avgpool-16",
common: CommonArgs = CommonArgs(),
crnn: CreateCRNNArgs = CreateCRNNArgs(),
save_model: bool = False,
) -> LaiaCRNN:
seed_everything(common.seed)
crnn.num_output_labels = len(SymbolsTable(syms))
if crnn is not None:
if fixed_input_height:
conv_output_size = LaiaCRNN.get_conv_output_size(
size=(fixed_input_height, fixed_input_height),
cnn_kernel_size=crnn.cnn_kernel_size,
cnn_stride=crnn.cnn_stride,
cnn_dilation=crnn.cnn_dilation,
cnn_poolsize=crnn.cnn_poolsize,
)
fixed_size_after_conv = conv_output_size[1 if crnn.
vertical_text else 0]
assert (fixed_size_after_conv >
0), "The image size is too small for the CNN architecture"
crnn.image_sequencer = f"none-{fixed_size_after_conv}"
else:
crnn.image_sequencer = adaptive_pooling
crnn.rnn_type = getattr(nn, crnn.rnn_type)
crnn.cnn_activation = [getattr(nn, act) for act in crnn.cnn_activation]
model = LaiaCRNN(**vars(crnn))
log.info(
"Model has {} parameters",
sum(param.numel() for param in model.parameters()),
)
if save_model:
ModelSaver(common.train_path,
common.model_filename).save(LaiaCRNN, **vars(crnn))
return model
default=[1, 2, 3, 4, 5],
nargs='+',
help='PHOC levels used to encode the transcript')
add_argument('syms', help='Symbols table mapping from strings to integers')
add_argument('img_dir', help='Directory containing word images')
add_argument('queries', help='Transcription of each query image')
add_argument('model_checkpoint', help='Filepath of the model checkpoint')
add_argument('output',
type=argparse.FileType('w'),
help='Filepath of the output file')
args = args()
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = build_dortmund_model(phoc_size)
log.info('Model has {} parameters',
sum(param.data.numel() for param in model.parameters()))
model.load_state_dict(torch.load(args.model_checkpoint))
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
model.eval()
queries_dataset = TextImageFromTextTableDataset(
args.queries, args.img_dir, img_transform=ImageToTensor())
queries_loader = DataLoader(queries_dataset)
def process_image(sample):
sample = Variable(sample, requires_grad=False)
sample = sample.cuda(args.gpu - 1) if args.gpu > 0 else sample.cpu()
phoc = torch.nn.functional.logsigmoid(model(sample))
return phoc.data.cpu().squeeze()
# Predict PHOC vectors
if __name__ == '__main__':
add_defaults('gpu')
add_argument('--phoc_levels', type=int, default=[1, 2, 3, 4, 5], nargs='+',
help='PHOC levels used to encode the transcript')
add_argument('syms', help='Symbols table mapping from strings to integers')
add_argument('img_dir', help='Directory containing word images')
add_argument('gt_txt', help='Transcription of each image')
add_argument('model_checkpoint', help='Filepath of the model checkpoint')
add_argument('output', type=argparse.FileType('w'),
help='Filepath of the output file')
args = args()
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = build_dortmund_model(phoc_size)
log.info('Model has {} parameters',
sum(param.data.numel() for param in model.parameters()))
model.load_state_dict(torch.load(args.model_checkpoint))
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
model.eval()
dataset = TextImageFromTextTableDataset(
args.gt_txt, args.img_dir, img_transform=ImageToTensor())
loader = DataLoader(dataset)
def process_image(sample):
sample = Variable(sample, requires_grad=False)
sample = sample.cuda(args.gpu - 1) if args.gpu > 0 else sample.cpu()
phoc = torch.nn.functional.sigmoid(model(sample))
return phoc.data.cpu().numpy()
# Predict PHOC vectors
default=[1, 2, 3, 4, 5],
nargs='+',
help='PHOC levels used to encode the transcript')
add_argument('--distractors',
help='Transcription of each distractor image')
add_argument('syms', help='Symbols table mapping from strings to integers')
add_argument('img_dir', help='Directory containing word images')
add_argument('queries', help='Transcription of each query image')
add_argument('model_checkpoint', help='Filepath of the model checkpoint')
add_argument('output', type=FileType('w'), help='Output file')
args = args()
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = build_dortmund_model(phoc_size)
log.info('Model has {} parameters',
sum(param.data.numel() for param in model.parameters()))
model.load_state_dict(torch.load(args.model_checkpoint))
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
model.eval()
def process_image(sample):
sample = Variable(sample, requires_grad=False)
sample = sample.cuda(args.gpu - 1) if args.gpu > 0 else sample.cpu()
phoc = torch.nn.functional.sigmoid(model(sample))
return phoc.data.cpu().numpy()
def process_dataset(filename):
dataset = TextImageFromTextTableDataset(filename,
args.img_dir,
img_transform=ImageToTensor())
data_loader = DataLoader(dataset)
def run(
syms: str,
img_dirs: List[str],
tr_txt_table: str,
va_txt_table: str,
common: CommonArgs = CommonArgs(),
train: TrainArgs = TrainArgs(),
optimizer: OptimizerArgs = OptimizerArgs(),
scheduler: SchedulerArgs = SchedulerArgs(),
data: DataArgs = DataArgs(),
trainer: TrainerArgs = TrainerArgs(),
):
pl.seed_everything(common.seed)
loader = ModelLoader(common.train_path,
filename=common.model_filename,
device="cpu")
# maybe load a checkpoint
checkpoint = None
if train.resume:
checkpoint = loader.prepare_checkpoint(common.checkpoint,
common.experiment_dirpath,
common.monitor)
trainer.max_epochs = torch.load(checkpoint)["epoch"] + train.resume
log.info(f'Using checkpoint "{checkpoint}"')
log.info(f"Max epochs set to {trainer.max_epochs}")
# load the non-pytorch_lightning model
model = loader.load()
assert (
model is not None
), "Could not find the model. Have you run pylaia-htr-create-model?"
# prepare the symbols
syms = SymbolsTable(syms)
for d in train.delimiters:
assert d in syms, f'The delimiter "{d}" is not available in the symbols file'
# prepare the engine
engine_module = HTREngineModule(
model,
[syms[d] for d in train.delimiters],
optimizer=optimizer,
scheduler=scheduler,
batch_input_fn=Compose([ItemFeeder("img"),
ImageFeeder()]),
batch_target_fn=ItemFeeder("txt"),
batch_id_fn=ItemFeeder("id"), # Used to print image ids on exception
)
# prepare the data
data_module = DataModule(
syms=syms,
img_dirs=img_dirs,
tr_txt_table=tr_txt_table,
va_txt_table=va_txt_table,
batch_size=data.batch_size,
color_mode=data.color_mode,
shuffle_tr=not bool(trainer.limit_train_batches),
augment_tr=train.augment_training,
stage="fit",
)
# prepare the training callbacks
# TODO: save on lowest_va_wer and every k epochs https://github.com/PyTorchLightning/pytorch-lightning/issues/2908
checkpoint_callback = pl.callbacks.ModelCheckpoint(
dirpath=common.experiment_dirpath,
filename="{epoch}-lowest_" + common.monitor,
monitor=common.monitor,
verbose=True,
save_top_k=train.checkpoint_k,
mode="min",
save_last=True,
)
checkpoint_callback.CHECKPOINT_NAME_LAST = "{epoch}-last"
early_stopping_callback = pl.callbacks.EarlyStopping(
monitor=common.monitor,
patience=train.early_stopping_patience,
verbose=True,
mode="min",
strict=False, # training_step may return None
)
callbacks = [
ProgressBar(refresh_rate=trainer.progress_bar_refresh_rate),
checkpoint_callback,
early_stopping_callback,
checkpoint_callback,
]
if train.gpu_stats:
callbacks.append(ProgressBarGPUStats())
if scheduler.active:
callbacks.append(LearningRate(logging_interval="epoch"))
# prepare the trainer
trainer = pl.Trainer(
default_root_dir=common.train_path,
resume_from_checkpoint=checkpoint,
callbacks=callbacks,
logger=EpochCSVLogger(common.experiment_dirpath),
checkpoint_callback=True,
**vars(trainer),
)
# train!
trainer.fit(engine_module, datamodule=data_module)
# training is over
if early_stopping_callback.stopped_epoch:
log.info(
"Early stopping triggered after epoch"
f" {early_stopping_callback.stopped_epoch + 1} (waited for"
f" {early_stopping_callback.wait_count} epochs). The best score was"
f" {early_stopping_callback.best_score}")
log.info(f"Model has been trained for {trainer.current_epoch + 1} epochs"
f" ({trainer.global_step + 1} steps)")
log.info(
f"Best {checkpoint_callback.monitor}={checkpoint_callback.best_model_score} "
f"obtained with model={checkpoint_callback.best_model_path}")
nargs='+',
help='PHOC levels used to encode the transcript')
add_argument('syms', help='Symbols table mapping from strings to integers')
add_argument('img_dir', help='Directory containing word images')
add_argument('candidates', help='Transcription of each candidate image')
add_argument('queries', help='Transcription of each query image')
add_argument('model_checkpoint', help='Filepath of the model checkpoint')
add_argument('output',
type=argparse.FileType('w'),
help='Filepath of the output file')
args = args()
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = build_dortmund_model(phoc_size)
log.info('Model has {} parameters',
sum(param.data.numel() for param in model.parameters()))
model.load_state_dict(torch.load(args.model_checkpoint))
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
model.eval()
candidates_dataset = TextImageFromTextTableDataset(
args.candidates, args.img_dir, img_transform=ImageToTensor())
candidates_loader = DataLoader(candidates_dataset)
queries_dataset = TextImageFromTextTableDataset(
args.queries, args.img_dir, img_transform=ImageToTensor())
queries_loader = DataLoader(queries_dataset)
def process_image(sample):
sample = Variable(sample, requires_grad=False)
sample = sample.cuda(args.gpu - 1) if args.gpu > 0 else sample.cpu()
def test_filepath(tmpdir):
filepath = tmpdir / "test"
log.config(filepath=filepath)
log.info("test!")
log.clear()
assert filepath.exists()
help='PHOC levels used to encode the transcript')
add_argument('--distractors',
help='Transcription of each distractor image')
add_argument('syms', help='Symbols table mapping from strings to integers')
add_argument('img_dir', help='Directory containing word images')
add_argument('queries', help='Transcription of each query image')
add_argument('model_checkpoint', help='Filepath of the model checkpoint')
add_argument('output',
type=argparse.FileType('w'),
help='Filepath of the output file')
args = args()
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = build_dortmund_model(phoc_size)
log.info('Model has {} parameters',
sum(param.data.numel() for param in model.parameters()))
model.load_state_dict(torch.load(args.model_checkpoint))
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
model.eval()
def process_image(sample):
sample = Variable(sample, requires_grad=False)
sample = sample.cuda(args.gpu - 1) if args.gpu > 0 else sample.cpu()
phoc = torch.nn.functional.logsigmoid(model(sample))
return phoc.data.cpu().squeeze()
def process_dataset(filename):
dataset = TextImageFromTextTableDataset(filename,
args.img_dir,
img_transform=ImageToTensor())
data_loader = DataLoader(dataset)
