模型输入必须来自“tf.keras.Input”...，它们不能是先前非输入层的输出

Question

我正在使用Python 3.7.7。和张量流 2.1.0。我有一个预训练的 U-Net 网络，我想获取它的编码器和解码器。如下图所示：您可以看到卷积编码器-解码器架构。我想要获取编码器部分，即出现在图像左侧的图层：以及解码器部分：我从这个函数中得到了 U-Net 模型：def get_unet_uncompiled(img_shape = (200,200,1)):    inputs = Input(shape=img_shape)    conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_1')(inputs)    conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_2')(conv1)    pool1 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool1')(conv1)    conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_1')(pool1)    conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_2')(conv2)    pool2 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool2')(conv2)    conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_1')(pool2)    conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_2')(conv3)    pool3 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool3')(conv3)    conv4 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv4_1')(pool3)    conv4 = Conv2D(256, (4, 4), activation='relu', padding='same', data_format="channels_last", name='conv4_2')(conv4)    pool4 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool4')(conv4)

Answer 1

我的建议是定义编码器和解码器的结构（get_encoder，get_decoder）。在整个模型的训练之后，我们的想法是创建一个新的解码器架构（通过get_decoder），我们可以用解码器训练的权重来填充它

pythonic 来说你可以用这种方式做到这一点......

def get_crop_shape(target, refer):

    

    # width, the 3rd dimension

    cw = (target.get_shape()[2] - refer.get_shape()[2])

    assert (cw >= 0)

    if cw % 2 != 0:

        cw1, cw2 = cw // 2, cw // 2 + 1

    else:

        cw1, cw2 = cw // 2, cw // 2

    # height, the 2nd dimension

    ch = (target.get_shape()[1] - refer.get_shape()[1])

    assert (ch >= 0)

    if ch % 2 != 0:

        ch1, ch2 = ch // 2, ch // 2 + 1

    else:

        ch1, ch2 = ch // 2, ch // 2

    return (ch1, ch2), (cw1, cw2)

def get_encoder(img_shape):

    

    inp = Input(shape=img_shape)

    conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_1')(inp)

    conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_2')(conv1)

    pool1 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool1')(conv1)

    conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_1')(pool1)

    conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_2')(conv2)

    pool2 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool2')(conv2)

    conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_1')(pool2)

    conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_2')(conv3)

    pool3 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool3')(conv3)

    conv4 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv4_1')(pool3)

    conv4 = Conv2D(256, (4, 4), activation='relu', padding='same', data_format="channels_last", name='conv4_2')(conv4)

    pool4 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool4')(conv4)

    conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_1')(pool4)

    conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_2')(conv5)

    

    return conv5,conv4,conv3,conv2,conv1,inp

def get_decoder(convs):

    

    conv5,conv4,conv3,conv2,conv1,inputs = convs

    

    up_conv5 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv5')(conv5)

    ch, cw = get_crop_shape(conv4, up_conv5)

    crop_conv4 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv4')(conv4)

    up6 = concatenate([up_conv5, crop_conv4])

    conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_1')(up6)

    conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_2')(conv6)

    up_conv6 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv6')(conv6)

    ch, cw = get_crop_shape(conv3, up_conv6)

    crop_conv3 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv3')(conv3)

    up7 = concatenate([up_conv6, crop_conv3])

    conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_1')(up7)

    conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_2')(conv7)

    up_conv7 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv7')(conv7)

    ch, cw = get_crop_shape(conv2, up_conv7)

    crop_conv2 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv2')(conv2)

    up8 = concatenate([up_conv7, crop_conv2])

    conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_1')(up8)

    conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_2')(conv8)

    up_conv8 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv8')(conv8)

    ch, cw = get_crop_shape(conv1, up_conv8)

    crop_conv1 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv1')(conv1)

    up9 = concatenate([up_conv8, crop_conv1])

    conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_1')(up9)

    conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_2')(conv9)

    ch, cw = get_crop_shape(inputs, conv9)

    conv9 = ZeroPadding2D(padding=(ch, cw), data_format="channels_last", name='conv9_3')(conv9)

    conv10 = Conv2D(1, (1, 1), activation='sigmoid', data_format="channels_last", name='conv10_1')(conv9)

    

    return conv10

    

def get_unet(img_shape = (200,200,1)):

    enc = get_encoder(img_shape)

    

    dec = get_decoder(enc)

    

    model = Model(inputs=enc[-1], outputs=dec)

    return model

创建整个模型并拟合

img_shape = (200,200,1)

old_model = get_unet(img_shape)

# old_model.compile(...)

# old_model.fit(...)

一如既往地提取编码器

# extract encoder

first_encoder_layer = 0

last_encoder_layer = 14

encoder_output_layer = [14, 11, 8, 5, 2, 0]

encoder = Model(inputs=old_model.layers[first_encoder_layer].input,

                outputs=[old_model.layers[l].output for l in encoder_output_layer],

                name='encoder')

encoder.summary()

创建解码器结构并分配训练后的权重

# extract decoder fitted weights

restored_w = []

for w in old_model.layers[last_encoder_layer + 1:]:

    restored_w.extend(w.get_weights())

  

# reconstruct decoder architecture setting the fitted weights

new_inp = [Input(l.shape[1:]) for l in get_encoder(img_shape)]

new_dec = get_decoder(new_inp)

decoder = Model(new_inp, new_dec)

decoder.set_weights(restored_w)

decoder.summary()

返回预测

# generate random images

n_images = 20

X = np.random.uniform(0,1, (n_images,200,200,1)).astype('float32')

# get encoder predictions 

pred_encoder = encoder.predict(X)

print([p.shape for p in pred_encoder])

# get decoder predictions

pred_decoder = decoder.predict(pred_encoder)

print(pred_decoder.shape)