ValueError：模型的输出张量必须是TensorFlow`Layer`的输出

Question

I'm building a model in Keras using some tensorflow function (reduce_sum and l2_normalize) in the last layer while encountered this problem. I have searched for a solution but all of it related to "Keras tensor".

Here is my code:

import tensorflow as tf;
from tensorflow.python.keras import backend as K

vgg16_model = VGG16(weights = 'imagenet', include_top = False, input_shape = input_shape);

fire8 = extract_layer_from_model(vgg16_model, layer_name = 'block4_pool');

pool8 = MaxPooling2D((3,3), strides = (2,2), name = 'pool8')(fire8.output);

fc1 = Conv2D(64, (6,6), strides= (1, 1), padding = 'same', name = 'fc1')(pool8);

fc1 = Dropout(rate = 0.5)(fc1);

fc2 = Conv2D(3, (1, 1), strides = (1, 1), padding = 'same', name = 'fc2')(fc1);

fc2 = Activation('relu')(fc2);

fc2 = Conv2D(3, (15, 15), padding = 'valid', name = 'fc_pooling')(fc2);

fc2_norm = K.l2_normalize(fc2, axis = 3);

est = tf.reduce_sum(fc2_norm, axis = (1, 2));
est = K.l2_normalize(est);

FC_model = Model(inputs = vgg16_model.input, outputs = est);

and then the error:

ValueError: Output tensors to a Model must be the output of a TensorFlow Layer (thus holding past layer metadata). Found: Tensor("l2_normalize_3:0", shape=(?, 3), dtype=float32)

I noticed that without passing fc2 layer to these functions, the model works fine:

FC_model = Model(inputs = vgg16_model.input, outputs = fc2);

Can someone please explain to me this problem and some suggestion on how to fix it?

Answer 1

I have found a way to work around to solve the problem. For anyone who encounters the same issue, you can use the Lambda layer to wrap your tensorflow operations, this is what I did:

from tensorflow.python.keras.layers import Lambda;

def norm(fc2):

    fc2_norm = K.l2_normalize(fc2, axis = 3);
    illum_est = tf.reduce_sum(fc2_norm, axis = (1, 2));
    illum_est = K.l2_normalize(illum_est);

    return illum_est;

illum_est = Lambda(norm)(fc2);

Answer 2

I had this issue because I was adding 2 tensors as x1+x2 somewhere in my model instead of using Add()([x1,x2]) .

That solved the problem.