Ошибка измерения в UNet с использованием Keras. Обрезка не происходит

Question

Я кодирую UNet для изображений разных размеров и использую data generator. UNet выглядит следующим образом:

class UNet():
    def __init__(self):
        print ('build UNet ...')

    def get_crop_shape(self, target, refer):
        # width, the 3rd dimension
        cw = (target.get_shape()[2] - refer.get_shape()[2]).value
        # height, the 2nd dimension
        ch = (target.get_shape()[1] - refer.get_shape()[1]).value
        if(cw==None and ch==None):
            print(cw, ch)
            return (0,0),(0,0)
        else:
            print(ch,cw)
            assert (cw >= 0)
            if cw % 2 != 0:
                cw1, cw2 = int(cw/2), int(cw/2) + 1
            else:
                cw1, cw2 = int(cw/2), int(cw/2)
            assert (ch >= 0)
            if ch % 2 != 0:
                ch1, ch2 = int(ch/2), int(ch/2) + 1
            else:
                ch1, ch2 = int(ch/2), int(ch/2)

            return (ch1, ch2), (cw1, cw2)

    def create_model(self, img_shape, num_class):

        concat_axis = 3
        inputs = keras.layers.Input(shape = (None, None, 1))
        # inputs = keras.layers.Input(shape = img_shape)

        conv1 = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same', name='conv1_1')(inputs)
        conv1 = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(conv1)
        pool1 = keras.layers.MaxPooling2D(pool_size=(2, 2))(conv1)
        conv2 = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(pool1)
        conv2 = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(conv2)
        pool2 = keras.layers.MaxPooling2D(pool_size=(2, 2))(conv2)

        conv3 = keras.layers.Conv2D(128, (3, 3), activation='relu', padding='same')(pool2)
        conv3 = keras.layers.Conv2D(128, (3, 3), activation='relu', padding='same')(conv3)
        pool3 = keras.layers.MaxPooling2D(pool_size=(2, 2))(conv3)

        conv4 = keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(pool3)
        conv4 = keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(conv4)
        pool4 = keras.layers.MaxPooling2D(pool_size=(2, 2))(conv4)

        conv5 = keras.layers.Conv2D(512, (3, 3), activation='relu', padding='same')(pool4)
        conv5 = keras.layers.Conv2D(512, (3, 3), activation='relu', padding='same')(conv5)

        up_conv5 = keras.layers.UpSampling2D(size=(2, 2))(conv5)
        ch, cw = self.get_crop_shape(conv4, up_conv5)
        crop_conv4 = keras.layers.Cropping2D(cropping=(ch,cw))(conv4)

        up6 = keras.layers.concatenate([up_conv5, crop_conv4], axis=concat_axis)

        conv6 = keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(up6)
        conv6 = keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(conv6)

        up_conv6 = keras.layers.UpSampling2D(size=(2, 2))(conv6)
        ch, cw = self.get_crop_shape(conv3, up_conv6)
        crop_conv3 = keras.layers.Cropping2D(cropping=(ch,cw))(conv3)
        up7 = keras.layers.concatenate([up_conv6, crop_conv3], axis=concat_axis) 

        conv7 = keras.layers.Conv2D(128, (3, 3), activation='relu', padding='same')(up7)
        conv7 = keras.layers.Conv2D(128, (3, 3), activation='relu', padding='same')(conv7)

        up_conv7 = keras.layers.UpSampling2D(size=(2, 2))(conv7)
        ch, cw = self.get_crop_shape(conv2, up_conv7)
        crop_conv2 = keras.layers.Cropping2D(cropping=(ch,cw))(conv2)
        up8 = keras.layers.concatenate([up_conv7, crop_conv2], axis=concat_axis)

        conv8 = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(up8)
        conv8 = keras.layers.Conv2D(64, (3, 3), activation='relu', padding='same')(conv8)

        up_conv8 = keras.layers.UpSampling2D(size=(2, 2))(conv8)
        ch, cw = self.get_crop_shape(conv1, up_conv8)
        crop_conv1 = keras.layers.Cropping2D(cropping=(ch,cw))(conv1)
        up9 = keras.layers.concatenate([up_conv8, crop_conv1], axis=concat_axis)

        conv9 = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(up9)
        conv9 = keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(conv9)

        ch, cw = self.get_crop_shape(inputs, conv9)
        conv9 = keras.layers.ZeroPadding2D(padding=((ch[0], ch[1]), (cw[0], cw[1])))(conv9)
        conv10 = keras.layers.Conv2D(1, (1, 1), activation='sigmoid')(conv9)

        model = keras.models.Model(inputs=inputs, outputs=conv10)

        return model

Когда я запускаю код, я получаю следующую ошибку:

тензор потока. python .framework.errors_impl.InvalidArgumentError: 2 root найдены ошибки. (0) Неверный аргумент: ConcatOp: Размеры входных данных должны совпадать: shape [0] = [16,512,30,52] и shape [1] = [16,256,30,53] [[{{node concatenate_1 / concat}} ]]
[[metrics / acc / Mean_1 / _345]] (1) Недопустимый аргумент: ConcatOp: размеры входных данных должны совпадать: shape [0] = [16,512,30,52] против формы [1] = [ 16,256,30,53] [[{{node concatenate_1 / concat}}]]

, что, вероятно, означает cropping из conv4. Более того, это единственный тензор с размерностью 256.

Однако, если я укажу форму входного изображения, ie, запустите inputs = keras.layers.Input(shape = img_shape) вместо inputs = keras.layers.Input(shape = (None, None, 1)), тогда код будет работать идеально.

Почему это происходит?