Possibly a mistake?

[alex-golts]

Shouldn't this line

pytorch-image-comp-rnn/train.py

Line 155 in 980f2ff

res = res - output

be res = res0 - output, where res0 = patches - 0.5?

Rather than res = res - output ?

The residual at each iteration should be with respect to the original image, no?

[1zb]

I think what you are talking about is one-shot reconstruction which is mentioned in section 2.2 of this paper.

[alex-golts]

@1zb I get it now, thanks!

[HeavvyRain]

@1zb I get it now, thanks!

hi alex,sorry to bother you,but i have been confused by a qustion for long time.The question is that if i trained this network with 300 epochs,and then i need to encode and decode with one checkpoint model, how do i choose the best checkpoint model from 300 checkpoint models?Is the 300th definitely best?