Inconsistency with paper in connection from vstack to hstack

[kkleidal]

Conditional-PixelCNN-decoder/models.py

Lines 30 to 34 in 9a5c9a3

	v_stack = GatedCNN([filter_size, filter_size, conf.f_map], v_stack_in, mask=mask, conditional=self.h).output()
	v_stack_in = v_stack
	with tf.variable_scope("v_stack_1"+i):
	v_stack_1 = GatedCNN([1, 1, conf.f_map], v_stack_in, gated=False, mask=mask).output()

This has the connection from the vstack to the hstack occuring after the gate, but in Figure 2 of the paper, it occurs after the masked convolution but before splitting the channels and gating. Any particular reason for this change?

[falconjhc]

Yeah I also found the issue. Is it the official code written by the original paper author?

[kkleidal]

Sidenote: Sorry for my previous comment (now deleted). I didn't read the email closely enough and thought it was in response to a similarly named repo I made.

I don't think this is the official implementation. I did try to correct the issue in #7 which has since been merged.

[yuffon]

In the original paper, the link from vstack to hstack is before gate.
This code is different.

[Raaaaaaat]

I'm new to this paper, and I found that the h_stack is different with the original paper.
the code is"[filter_size if full_horizontal else 1, filter_size, conf.f_map]"
and para "full_horizontal" is set as TRUE
Cuz the size of mask is changging from 7x7 to 3x3, and perhaps the size of mask should be 1x7/1x3?
I think the size of horizontal stack is changging from 1x3 to 1x7, but maybe I understand the wrong way, anyone can explain why the kernel changes from 7x7 to 3x3 and why it is 7x7 instead of 1x7?
I'm so confused, plzplzplz help me out, I really appreciate your help! THX!!!!