def get_crnn_output(boxes):
crnn = model.CRNN()()
crnn.load_weights(config.crnn_weights_path)
for i in range(len(boxes)):
image = boxes[i]['img']
image = utils_image.resizeNormalize(image, 32)
image = image.astype(np.float32)
image = np.array([[image]])
preds = crnn.predict(image)
preds = np.argmax(preds, axis=-1).reshape((-1,))
boxes[i]['text'] = utils.strLabelConverter(preds, config.alphabet)
return boxes
'print_word': train_dataset_print_word,
'print_num': train_dataset_print_num,
'symbol': train_dataset_symbol,
'seal': train_dataset_seal,
'catword': train_dataset_catword,
}
valid_datasets = {
'hand_word': valid_dataset_hand_word,
'hand_num': valid_dataset_hand_num,
'print_word': valid_dataset_print_word,
'print_num': valid_dataset_print_num,
'symbol': valid_dataset_symbol,
'seal': valid_dataset_seal,
'catword': valid_dataset_catword,
}
alphabets = alphabets.alphabet_char
n_class = len(alphabets) + 1
converter = convert.strLabelConverter(alphabets)
criterion = CTCLoss()
model = model.CRNN(class_num=n_class, backbone='resnet', pretrain=False)
train(model,
criterion,
converter,
device,
train_datasets,
valid_datasets,
pretrain=True)