python - FailedPreconditionError：尝试使用未初始化的值 lambda/Variable

Question

我一直在尝试执行 UNET CNN 代码，但我得到了同样的错误：FailedPreconditionError: Attempting to use uninitialized value lambda/Variable

当我调用函数时出现问题：

model.fit(x_train, y_train, batch_size=16, epochs=1000, shuffle=True, validation_data=[x_val, y_val])

我已经尝试使用以下方法初始化变量： tf.global_variables_initializer() 但问题仍然存在

我的 tensorflow-gpu 版本：'1.13.1'

我的python版本：'3.7.3'

显卡：Titan RTX

def U_Net(activation='relu', ft_root=64, batch_norm=True):

        inputs = Input((None, None, None, 1))
        x = inputs

        # Dictionary for long connections
        var_dict = {}
        # Down sampling

        for i in range(self.n_layers):
            out_channel= 2**i * ft_root
            # Convolutions
            conv1 = Conv3D(out_channel, kernel_size=3, padding='same')(x) 
            if batch_norm:
                conv1 = BatchNormalization()(conv1)
            act1 = Activation(activation)(conv1)

            conv2 = Conv3D(out_channel, kernel_size=3, padding='same')(act1) # deuxieme couche de convolution
            if batch_norm:
                conv2 = BatchNormalization()(conv2)
            act2 = Activation(activation)(conv2)
            # Max pooling
            if i < self.n_layers - 1:
                var_dict[str(i)] = act2
                x = MaxPooling3D(padding='same')(act2)
            else:
                x = act2

        # Upsampling

        for i in range(self.n_layers-2, -1, -1):
            out_channel = 2**(i)*ft_root
            # Convolution transposed
            upconv = Conv3DTranspose(out_channel, kernel_size=2, strides=2, use_bias=False)(x) # upsampling
            upconv_ = Lambda(CroppingLayer)([upconv, var_dict[str(i)]])
            uplong = Add()([upconv_, var_dict[str(i)]]) # skip connection

            # Convolutions
            conv1 = Conv3D(out_channel, kernel_size=3, padding='same')(uplong)
            if batch_norm:
                conv1 = BatchNormalization()(conv1)
            act1 = Activation(activation)(conv1)

            conv2 = Conv3D(out_channel, kernel_size=3, padding='same')(act1)
            if batch_norm:
                conv2 = BatchNormalization()(conv2)
            x = Activation(activation)(conv2)

        # Final convolution
        output = Conv3D(1, kernel_size=1, padding='same', activation='sigmoid')(x)
        return Model(inputs, output, name='U-Net')



def dice_loss(labels, logits):
        eps = 1e-5
        pred = K.flatten(logits) 
        lab = K.flatten(labels)
        intersection = K.sum(pred*lab)
        dice_0 = (2*intersection + eps) / (K.sum(pred) + K.sum(lab) + eps)
        return 1 - dice_0


tf.global_variables_initializer()
model.fit(x_train, y_train, batch_size=16, epochs=1000, shuffle=True, validation_data=[x_val, y_val])

我希望培训开始运行；但确切的输出是：

文件“/home/derf/anaconda3/lib/python3.7/site-packages/tensorflow/python/framework/errors_impl.py”，第 528 行，退出 c_api.TF_GetCode(self.status.status)) FailedPreconditionError: Attempting to使用未初始化的值 lambda/Variable [[{{node lambda/Variable/read}}]] [[{{node loss/mul}}]]

标签： pythontensorflowunity3d-unet