Hi All,

Does anyone understand the significance of identity_initializer?

def identity_initializer():
    def _initializer(shape, dtype=tf.float32, partition_info=None):
        array = np.zeros(shape, dtype=float)
        cx, cy = shape[0]//2, shape[1]//2
        for i in range(np.minimum(shape[2],shape[3])):
            array[cx, cy, i, i] = 1
        return tf.constant(array, dtype=dtype)
    return _initializer

Any intuition is much appreciated.!

you use Adaptive Normalization but I didn't notice you deal with moving_mean and moving_variance，Do I miss something?

Hi!
Is the dataset on which You trained Your network available publicly, or You can provide it?
And why is your single network input has unusual channels count(10 +3 instead of 3 ), what does additional 10 channels with default initialized const values do, or contribute to?
I’m looking forward to your reply.

When I tried to train the model, I got error as follows:

Traceback (most recent call last):
  File "demo.py", line 97, in <module>
    _,current=sess.run([opt,loss],feed_dict={input:input_images[id],output:output_images[id]})
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py", line 929, in
 run
    run_metadata_ptr)
  File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/client/session.py", line 1128, i
n _run
    str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1,) for Tensor 'Placeholder_1:0', which has shape '(?, ?,
 ?, 3)'

The version of tensorflow I used is 1.12, and the system is Ubuntu. Can you help me solve this problem.

I realized that the slim.con2d，if you use the normalizer_fn , it says the conv will have no bias.
but there comes the question since it uses adaptive BN ，as I can see in the code the initial parameter is w0=1 and w1=0, which means at first it doesn't use batch normalization, So the bias is used in futher forward-propagation.and even though w1 is not 0, as long as the w0 do not becoming to 0, I think the bias is still used in someway.So，Should I use bias in conv？I was doing this in pytorch.it doesn't have a function like slim.conv2d.

In your paper, the Fig 3 is very meaningful.
It's very easy to find the missing parts in the corresponding error map.
Can you provide the error map function?
Thanks a lot.