add_argument("--use_new_phoc", action="store_true")
add_argument("syms", help="Symbols table mapping from strings to integers")
add_argument("tr_img_dir", help="Directory containing word images")
add_argument("tr_txt_table",
help="Character transcriptions of each training image")
add_argument("va_txt_table",
help="Character transcriptions of each validation image")
args = args()
laia.common.random.manual_seed(args.seed)
syms = laia.utils.SymbolsTable(args.syms)
phoc_size = sum(args.phoc_levels) * len(syms)
model = DortmundPHOCNet(phoc_size=phoc_size,
tpp_levels=args.tpp_levels,
spp_levels=args.spp_levels)
if args.load_checkpoint:
model_ckpt = torch.load(args.load_checkpoint)
model.load_state_dict(model_ckpt)
model = model.cuda(args.gpu - 1) if args.gpu > 0 else model.cpu()
logger.info("PHOC embedding size = {}", phoc_size)
logger.info(
"Model has {} parameters",
sum(param.data.numel() for param in model.parameters()),
)
if args.use_adam_optim:
logger.info(
"Using ADAM optimizer with learning rate = {:g} and weight decay = {:g}",
args.learning_rate,
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 = DortmundPHOCNet(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 test_batch_output_size(self):
m = DortmundPHOCNet(phoc_size=125, tpp_levels=[1, 2, 3])
x = torch.randn(3, 1, 93, 30)
y = m(x)
self.assertEqual([3, 125], list(y.size()))
def test_padded_batch_output_size_tpp_and_spp(self):
m = DortmundPHOCNet(phoc_size=40, tpp_levels=[1, 2, 3], spp_levels=[1, 2])
x = torch.randn(3, 1, 93, 30)
xs = torch.tensor([[93, 30], [40, 30], [93, 20]])
y = m(PaddedTensor(x, xs))
self.assertEqual([3, 40], list(y.size()))
def test_number_parameters(self):
m = DortmundPHOCNet(phoc_size=540, tpp_levels=[1, 2, 3, 4, 5])
self.assertEqual(59859420, sum(p.numel() for p in m.parameters()))