如果网络未按模型馈电，如何馈电。适合

Question

我在keras中有一个简单的网络，我定义了一个自定义层，该层对输入张量进行一些操作，然后将其返回到网络，但当我想要实现它时，它会产生以下错误，并表示输入尚未反馈，而我想当我们使用fit函数时，它会反馈网络。你能帮我解决这个问题吗？我找不到合适的答案来解决我的问题。我也把代码放在这里了。非常感谢。

def C(u):
    if u == 0:
        return 1. / np.sqrt(2.)
    else:
        return 1.

def DCT(a, b):
    for u in range(8):
        for v in range(8):
            for x in range(8):
                for y  in range(8):
                    b[u,v] = b[u, v] + 0.25 * C(u) * C(v) * a[x, y]* np.cos((2 * x+1) * (u) * np.pi / 16) * np.cos((2 * y+1) * (v) * np.pi / 16)

def IDCT(a, b):
    for u in range(8):
        for v in range(8):
            for x in range(8):
                for y  in range(8):
                    b[x,y] = b[x, y] + 0.25 * C(u) * C(v) * a[u,v] * np.cos((2 * x+1) * (u) * np.pi / 16) * np.cos((2 * y+1) * (v) * np.pi / 16)

def quntize_mask(window_size: int, keep_count: int):
    mask = np.zeros((window_size, window_size), dtype=np.uint8)

    index_order = sorted(((x, y) for x in range(window_size) for y in range(window_size)),
                         key=lambda p: (p[0] + p[1], -p[1] if (p[0] + p[1]) % 2 else p[1]))

    for i, j in index_order[0:keep_count]:
        mask[i, j] = 1
    return mask
def slicAndJpeg(img):
    for i in range (int(img.shape[1].value/8)):
        for j in range(int(img.shape[2].value/8)):
            temp=(img[:,i*8:i*8+8,j*8:j*8+8])
            tempb=np.zeros((8,8))
            DCT(temp,tempb)
            mask=quntize_mask(8,9)
            qunz=Kr.layers.multiply(mask,tempb)
            tempc=K.zeros((8,8))
            IDCT(qunz,tempc)
            img[:,i*8:i*8+8,j*8:j*8+8]=tempc

class JPEGLayer(Layer):

    def __init__(self,**kwargs):
        super(JPEGLayer, self).__init__(**kwargs)
        self.supports_masking = True
    def call(self, noised_image, training=True):
        def noise():
#            noised_image = noised_and_cover
            # pad the image so that we can do dct on 8x8 blocks
            pad_height = (8 - noised_image.shape[1] % 8) % 8
            pad_width = (8 - noised_image.shape[2] % 8) % 8

            noised_image_pad = Kr.layers.ZeroPadding2D(padding=(( pad_width, 0),( pad_height,0)))(noised_image)
            slicAndJpeg(K.eval(noised_image_pad))


        # un-pad
            noised_and_cover = noised_image_pad[ :, :noised_image_pad.shape[1]-pad_height, :noised_image_pad.shape[2]-pad_width]
            return noised_and_cover 
        return noise()

#-----------------building w train---------------------------------------------
wt_random=np.random.randint(2, size=(49999,4,4))
w_expand=wt_random.astype(np.float32)
wv_random=np.random.randint(2, size=(9999,4,4))
wv_expand=wv_random.astype(np.float32)
x,y,z=w_expand.shape
w_expand=w_expand.reshape((x,y,z,1))
x,y,z=wv_expand.shape
wv_expand=wv_expand.reshape((x,y,z,1))

#-----------------building w test---------------------------------------------
w_test = np.random.randint(2,size=(1,4,4))
w_test=w_test.astype(np.float32)
w_test=w_test.reshape((1,4,4,1))


#-----------------------encoder------------------------------------------------
#------------------------------------------------------------------------------
image = Input((28, 28, 1))
conv1 = Conv2D(64, (5, 5),activation='relu',padding='same', name='convl1e')(image)
wtm=Input((4,4,1))
#--------------------------------------------------------------
wpad=Kr.layers.Lambda(lambda xy: xy[0] + Kr.backend.spatial_2d_padding(xy[1], padding=((0, 24), (0, 24))))
encoded_merged=wpad([conv1,wtm])#-----------------------decoder------------------------------------------------
#------------------------------------------------------------------------------
decoded = Conv2D(1, (5, 5),activation='relu', padding='same', name='decoder_output')(encoded_merged) 
model=Model(inputs=[image,wtm],outputs=decoded)
model.summary()
decoded_noise=JPEGLayer()(decoded)#16

#----------------------w extraction------------------------------------
convw1 = Conv2D(64, (5,5),activation='relu' , name='conl1w')(decoded_noise)#24
convw2 = Conv2D(64, (5,5),activation='relu' , name='conl2w')(convw1)#20
#Avw1=AveragePooling2D(pool_size=(2,2))(convw2)
convw3 = Conv2D(64, (5,5),activation='relu' ,name='conl3w')(convw2)#16
convw4 = Conv2D(64, (5,5), activation='relu' ,name='conl4w')(convw3)#12
#Avw2=AveragePooling2D(pool_size=(2,2))(convw4)
convw5 = Conv2D(64, (5,5), activation='relu' ,name='conl5w')(convw4)#8

convw6 = Conv2D(64, (5,5), activation='relu' ,name='conl6w')(convw5)#4
pred_w = Conv2D(1, (1, 1),activation='relu' ,padding='same', name='reconstructed_W')(convw6)

model1=Model(inputs=[image,wtm],outputs=[decoded,pred_w])

model1.summary()
#----------------------training the model--------------------------------------
#------------------------------------------------------------------------------
#----------------------Data preparesion----------------------------------------

(x_train, _), (x_test, _) = mnist.load_data()
x_validation=x_train[1:10000,:,:]
x_train=x_train[10001:60000,:,:]
#
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_validation = x_validation.astype('float32') / 255.
x_train = np.reshape(x_train, (len(x_train), 28, 28, 1))  # adapt this if using `channels_first` image data format
x_test = np.reshape(x_test, (len(x_test), 28, 28, 1))  # adapt this if using `channels_first` image data format
x_validation = np.reshape(x_validation, (len(x_validation), 28, 28, 1))

#---------------------compile and train the model------------------------------
opt=SGD(momentum=0.99,lr=0.0001)
model1.compile(optimizer='adam', loss={'imageprim':'mse','wprim':'binary_crossentropy'}, loss_weights={'imageprim': 0.5, 'wprim': 1.0},metrics=['mae'])
es = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=40)
#rlrp = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=20, min_delta=1E-4, verbose=1)
mc = ModelCheckpoint('sendAct.h5', monitor='val_loss', mode='min', verbose=1, save_best_only=True)
history=model1.fit([x_train,w_expand], [x_train,w_expand],
          epochs=4000,
          batch_size=32, 
          validation_data=([x_validation,wv_expand], [x_validation,wv_expand]),
          callbacks=[TensorBoard(log_dir='/home/jamalm8/tensorboardGNWLoss/', histogram_freq=0, write_graph=False),es,mc])
model1.summary()

回溯（最近一次呼叫最后一次）：

文件“”，第124行，已解码_noise=JPEGLayer（）（已解码）#16

html" target="_blank">文件"D：\Software\Anaconda3\envs\py36\lib\site-包\keras\Engine\base_layer.py"，第457行，在调用输出=self.call（输入，**kwargs）

文件"，第94行，调用返回噪声（）

文件“”，第88行，采用噪声滑动JPEG格式（K.eval（噪声图像垫））

文件"D：\Software\Anaconda3\envs\py36\lib\site-包\keras\backend\tensorflow_backend.py"，第673行，在值返回to_dense（x）。

文件"D：\Software\Anaconda3\envs\py36\lib\site-包\tenstorflow\python\框架\ops.py"，第713行，在返回_eval_using_default_session（自我，feed_dict，self.graph，会话）

文件“D:\software\Anaconda3\envs\py36\lib\site packages\tensorflow\python\framework\ops.py”，第5157行，使用默认会话返回会话进行评估。运行（张量、进给量）

文件“D:\software\Anaconda3\envs\py36\lib\site packages\tensorflow\python\client\session.py”，第929行，在运行元数据中

文件“D:\software\Anaconda3\envs\py36\lib\site packages\tensorflow\python\client\session.py”，第1152行，在运行提要中

文件“D:\software\Anaconda3\envs\py36\lib\site packages\tensorflow\python\client\session.py”，第1328行，在元数据中）

文件“D:\software\Anaconda3\envs\py36\lib\site packages\tensorflow\python\client\session.py”，第1348行，位于调用提升类型（e）（节点定义，操作，消息）

InvalidArgumentError：您必须为占位符张量'input_1'提供一个值，该值具有dtype浮点和形状[？，28,28,1][[节点input_1（定义于D：\Software\Anaconda3\envs\py36\lib\site-包\keras\backend\tensorflow_backend.py:517）=Placeholderdtype=DT_FLOAT，形状=[？，28,28,1]，_device="/作业：localhost/replica： 0/任务： 0/设备： GPU： 0"]] [[{{节点jpeg_layer_1/zero_padding2d_1/Pad/_9}}=_Recvclient_terminated=假，recv_device="/作业：localhost/replica： 0/任务： 0/设备： CPU： 0"，send_device="/作业：localhost/replica： 0/任务： 0/设备： GPU： 0"，send_device_incarnation=1，tensor_name="edge_39_jpeg_layer_1/zero_padding2d_1/Pad"，tensor_type=DT_FLOAT，_device="/作业：localhost/replica： 0/任务： 0/设备： CPU： 0"]]

InvalidArgumentError（回溯见上文）：您必须为占位符张量'input_1'输入一个值，该值带有dtype float和shape[？，28,28,1][[node input_1（定义在D:\software\Anaconda3\envs\py36\lib\site packages\keras\backend\tensorflow_backend.py:517]=占位符dtype=dtu float，shape=[？，28,28,1]，\u设备=“/job:localhost/replica:0/task:0/device:GPU:0”][{{node-jpeg\u layer\u 1/zero\u padding2d\u 1/Pad/\u 9}}=\u-Recvclient\u-terminated=false，recv\u-device=“/job:localhost/replica:0/task:0/device:CPU:0/task:0/device:GPU:0”，send\u-device\u-device\u-environment\u-environment=1，tensor\u-name=“edge\u-39\jpeg\u-paddingu-2d\paddingu-layer\paddu-1”“，tensor_type=DT_FLOAT，_device=“/job:localhost/replica:0/task:0/device:CPU:0”]]

Answer 1

它是由JPEGLayer类中的行slicAndJpeg（K. ava（noised_image_pad））引起的。基本上，您试图通过调用K. ava（）来计算张量，而不向其提供任何数据。你不能计算空张量，对吗？这可以通过完全删除噪声（）函数并在预处理中执行填充/切片和其他任务来修复。

用以下方式替换JPEGLayer类，它应该可以工作（假设输入数据已经填充）

class JPEGLayer(Layer):

    def __init__(self,**kwargs):
        super(JPEGLayer, self).__init__(**kwargs)
        self.supports_masking = True
    def call(self, noised_image, training=True):
        return noised_image