Keras Lambda layer runs twice even though called only once - tensorflow

Question

Why the following code runs twice gram layer?
import numpy as np
from keras.applications import vgg19
from keras import backend as K
from keras.layers import Input, Lambda
import tensorflow as tf
from keras.models import Model
def gram_layer(y):
print('Using Gram Layer')
# assert K.ndim(y) == 4
print(y.shape, 'y.shape')
# a = y.get_shape()[1].value
# b = y.get_shape()[2].value
# c = y.get_shape()[3].value
# print(a, b, c)
# x = K.squeeze(y, axis=0)
# features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))
# features_nomean = features - K.mean(features, axis=0, keepdims=True)
# gram = K.dot(features_nomean, K.transpose(features_nomean)) / (a * b * c)
print('exiting Gram')
# return x
return y
In = K.placeholder((1, 256, 256, 3))
model = vgg19.VGG19(input_tensor = In, weights='imagenet', include_top=False)
for layer in model.layers:
if layer.name == 'block1_conv1':
print(layer.name)
print(layer.output.shape)
outputs = (Lambda(gram_layer))(layer.output)
Debug info:
block1_conv1
(1, 256, 256, 64)
Using Gram Layer
(1, 256, 256, 64) y.shape
exiting Gram
Using Gram Layer
(?, ?, ?, 64) y.shape
exiting Gram
Debug information contains two "Using Gram Layer", it means this layer runs twice, and it fails at the second time, but strangely it's only called once.
Any idea what's wrong?
PS: I realize that the problem lies in the for loop part, if the last line
outputs = (Lambda(gram_layer))(layer.output)
is replaced as
outputs = (Lambda(gram_layer))(In)
the debug info goes as
block1_conv1
(1, 256, 256, 64)
Using Gram Layer
(1, 256, 256, 3) y.shape
exiting Gram
Using Gram Layer
(?, ?, ?, 3) y.shape
exiting Gram
If the last 5 lines are replaced as
outputs = (Lambda(gram_layer))(In)
then the debug info goes as
Using Gram Layer
(1, 256, 256, 3) y.shape
exiting Gram
Using Gram Layer
(1, 256, 256, 3) y.shape
exiting Gram
It still runs twice, but the shape inference is correct. Is this a bug? or should I report it at GitHub?

Answer 1

Not sure why your function is called twice, but it's not uncommon to see that. It's called during compilation first.
The problem there seems to te reshaping with "None" values. That's not supported.
You can reshape with "-1" instead of None, but you can have only one "-1" in a reshape.
Suggestion 1:
All your reshape code can be replaced with: x = K.squeeze(y,axis=0)
Warning:
But this is highly unusual in keras. The axis=0 dimension is the batch size. This code will only run fine with batch_size = 1. (Either your code and my suggestion).
Suggestion 2:
If you're going to use batch_flatten, why the reshape then?
Any reshape you do before batch_flatten() will be pointless, unless you really mean to flatten only the last two dimensions and have a (256,768) tensor.
Suggestion 3:
If you want the actual values of a, b, c for calculations, you need to get their tensor values instead of their config values:
shp = K.shape(y)
a = shp[1] #maybe you need shp[1:2], depending on whether you get an error in the division line
b = shp[2]
c = shp[3]
Suggestion 4:
It's quite strange to use a placeholder. That's not the keras way of doing it.
You should simply create the model and tell it the shape you want:
model = vgg19.VGG19(input_shape = (256,256,64), weights='imagenet', include_top=False)
If you do want to enforce a batch size of 1, then you can create an input tensor:
inputTensor = Input(batch_shape=(1,256,256,64)
output = model(inputTensor)
model = Model(inputTensor, output)