About the dimensions of inputs

[guvcolie]

def build_inpaint_net(self, x, mask, config=None, reuse=False, training=True, padding='SAME', name='inpaint_net'):
    xin = x
    offset_flow = None
    ones_x = tf.ones_like(x)[:, :, :, 0:1]
    x = tf.concat([x, ones_x, ones_x * mask], axis=3)

I find the input contains 'ones_x',is there anything meaningful?

[JiahuiYu]

Because the convolution automatically pad zeros as in 'SAME' mode. You can also have a look of issue #40.

[LeonCurry]

I am sorry to bother you. I want to ask you a question. Does the bbox2mask function generate a single-channel mask? @JiahuiYu

[JiahuiYu]

@LeonCurry Yes.

[LeonCurry]

Thanks for your reply! @JiahuiYu

[soar200]

what is yout data_list in inpatient.yml ?the example has bugger

[JiahuiYu]

@soar200 #15

[avisekiit]

Hello @soar200 and @JiahuiYu

Thanks for the informative thread. I had some specific doubts regarding the input which is fed to the network. I know ALMOST similar questions have been asked on this topic but, unfortunately, I could not get a full picture about the exact nature of the inputs which are fed to the network.

For example, if we consider: tf.concat([x, ones_x, ones_x*mask], axis=3). My doubts are:

1. I understand (x) is the INCOMPLETE normalized image in [-1, 1]. Do you make the masked pixels = 0 in (x) or you make masked pixels = 1 in (x)

2. Why do we need to input the (ones_x) tensor ? It is simply an all-white input? Is it explicitly used anywhere for inpainting purpose?

3. Is the (mask) tensor = 0 on masked pixels or is it = 1 on masked pixels ?

Thanks, Avisek