parser.add_argument("-e",
"--epochs",
type=int,
default=100,
help="Num of epochs to train.")
args = parser.parse_args()
print('args', args)
train_dl = OCRDataLoader(args.train_annotation_paths, args.train_parse_funcs,
args.image_width, args.table_path, args.batch_size,
True)
with open(r'E:\tsl_file\python_project\CRNN.tf2\example\table.txt', "r") as f:
inv_table = [char.strip() for char in f]
decoder = Decoder(inv_table)
print("Num of training samples: {}".format(len(train_dl)))
localtime = time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime())
# saved_model_path = ("saved_models/{}/".format(localtime) +
# "{epoch:03d}_{word_accuracy:.4f}.h5")
if args.val_annotation_paths:
val_dl = OCRDataLoader(args.val_annotation_paths, args.val_parse_funcs,
args.image_width, args.table_path, args.batch_size)
print("Num of val samples: {}".format(len(val_dl)))
# saved_model_path = ("saved_models/{}/".format(localtime) +
# "{epoch:03d}_{word_accuracy:.4f}_{val_word_accuracy:.4f}.h5")
else:
val_dl = lambda: None
print("Start at {}".format(localtime))
def read_img_and_preprocess(path):
img = tf.io.read_file(path)
img = tf.io.decode_jpeg(img, channels=args.img_channels)
img = tf.image.convert_image_dtype(img, tf.float32)
img = tf.image.resize(img, (32, args.img_width))
return img
if os.path.isdir(args.images):
img_paths = os.listdir(args.images)
img_paths = [os.path.join(args.images, path) for path in img_paths]
imgs = list(map(read_img_and_preprocess, img_paths))
imgs = tf.stack(imgs)
else:
img_paths = [args.images]
img = read_img_and_preprocess(args.images)
imgs = tf.expand_dims(img, 0)
with open(args.table_path, 'r') as f:
inv_table = [char.strip() for char in f]
model = keras.models.load_model(args.model, compile=False)
decoder = Decoder(inv_table)
y_pred = model.predict(imgs)
for path, g_pred, b_pred in zip(img_paths,
decoder.decode(y_pred, method='greedy'),
decoder.decode(y_pred, method='beam_search')):
print('Path: {}, greedy: {}, beam search: {}'.format(path, g_pred, b_pred))