def process_dataset(filename):
dataset = TextImageFromTextTableDataset(
filename, args.img_dir, img_transform=ImageToTensor())
data_loader = DataLoader(dataset)
phocs = []
samples = []
for sample in tqdm(data_loader):
phocs.append(process_image(sample['img']))
samples.append(sample['id'][0])
return np.concatenate(phocs), samples
def process_dataset(filename):
dataset = TextImageFromTextTableDataset(
filename, args.img_dir, img_transform=ImageToTensor()
)
data_loader = DataLoader(dataset)
phocs = []
samples = []
for sample in tqdm(data_loader):
phocs.append(process_image(sample["img"]))
samples.append(sample["id"][0])
return torch.stack(phocs).type("torch.DoubleTensor"), samples
ckpt = torch.load(args.checkpoint)
if 'model' in ckpt and 'optimizer' in ckpt:
model.load_state_dict(ckpt['model'])
else:
model.load_state_dict(ckpt)
# Ensure parameters are in the correct device
model.eval()
if args.gpu > 0:
model = model.cuda(args.gpu - 1)
else:
model = model.cpu()
dataset = TextImageFromTextTableDataset(args.gt_file,
args.img_dir,
img_transform=ImageToTensor(),
txt_transform=TextToTensor(syms))
dataset_loader = ImageDataLoader(dataset=dataset,
image_channels=1,
num_workers=8)
with torch.cuda.device(args.gpu - 1):
for batch in dataset_loader:
if args.gpu > 0:
x = batch['img'].data.cuda(args.gpu - 1)
else:
x = batch['img'].data.cpu()
y = model(torch.autograd.Variable(x)).data
if args.add_softmax:
y = torch.nn.functional.log_softmax(y, dim=-1)
dump_output_matrix(batch['id'], [y], args.output,
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
phocs = []
labels = []
samples = []
for query in tqdm(queries_loader):
phocs.append(process_image(query['img']))
labels.append(query['txt'][0])
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
for query in tqdm(loader):
phoc = process_image(query['img'])
print(query['id'][0], file=args.output, end='')
for j in range(phoc.shape[1]):
print(' %.12g' % phoc[0, j], file=args.output, end='')
print('', file=args.output)
sampler=FixedSizeSampler(tr_ds, args.train_samples_per_epoch)
if args.train_samples_per_epoch else None)
# Set all these separately because they might change between executions
trainer.iterations_per_update = args.iterations_per_update
trainer.set_data_loader(tr_ds_loader)
trainer.set_batch_input_fn(
ImageFeeder(
device=args.gpu,
# keep_padded_tensors=False,
parent_feeder=ItemFeeder('img')))
va_ds = TextImageFromTextTableDataset(
args.va_txt_table,
args.img_dir,
img_transform=ImageToTensor(fixed_height=args.fixed_height),
txt_transform=TextToTensor(syms))
va_ds_loader = ImageDataLoader(dataset=va_ds,
image_channels=1,
batch_size=args.batch_size,
num_workers=mp.cpu_count())
evaluator = Evaluator(
model=model,
data_loader=va_ds_loader,
batch_input_fn=ImageFeeder(device=args.gpu,
parent_feeder=ItemFeeder('img')),
batch_target_fn=ItemFeeder('txt'),
progress_bar='Valid' if args.show_progress_bar else None)
engine_wrapper = HtrEngineWrapper(trainer, evaluator)