sjchoi86 / advanced-tensorflow Goto Github PK

Hi, writer. I am impressed with your code of srgan. But, I have some questions here.

1. At the function of setup_inputs, I can't understand why you need to crop the image data.
2. Why you use the queue format data image instead of np.array format.

I am looking forward to your reply.

In creating the 3 convolutional encoding layer you accidentally save the dropout of this layer to the wrong name

` _ce3 = tf.nn.sigmoid(tf.add(tf.nn.conv2d(_ce2, _W['ce3']
, strides=[1, 2, 2, 1], padding='SAME'), _b['be3']))

_ce1 = tf.nn.dropout(_ce3, _keepprob)`

When it should be

` _ce3 = tf.nn.sigmoid(tf.add(tf.nn.conv2d(_ce2, _W['ce3']
, strides=[1, 2, 2, 1], padding='SAME'), _b['be3']))

_ce3 = tf.nn.dropout(_ce3, _keepprob)`

Probably doesn't affect performance, as it just skips the drop out for one layer. But still, though it would be nice to point this out.