search

rosinality / vq-vae-2-pytorch

Implementation of Generating Diverse High-Fidelity Images with VQ-VAE-2 in PyTorch
Other
1.65k stars 275 forks source link

Results of Pixelsnail are not good #22

Open hasamkhalid opened 4 years ago

hasamkhalid commented 4 years ago

Hi,

I ran the code as you suggested and completed all 420 epochs for hier top and bottom but the generated results are not good, you can see below. Please suggest me what should I do to generate good images. Thanks

40out_check 49out_check

drtonyr commented 4 years ago

They look really good to me, better than what I've got (though I did run on 128*128 to match my GPU).

Here is one simple suggestion that worked for me.  Do you see how you don't have very good left/right symmetry?  (the eyes and the hair style don't match).  I think this is because pixelSNAIL (which is the bit that does all the hard structural work) is scanning vertically so the RNN has to store the left/right detail for a long time. If you run pixelSNAIL horizontally then the detail such as eyes only has to be stored for the width of the eye. It would be cleanest to change the code, but I just rotated the images prior to training.

Chauncy-Cai commented 3 years ago

It seems training pixelsnail (both top and bottom) take much longer than than train vqvae. 44min for every epoch. I don't know whether it is normal. Moreover, is there anyone have try other type of dataset, for instance imagenet? They seem like work for vq-vae-v2 in the paper.

rosinality commented 3 years ago

@Chauncy-Cai PixelSNAIL is quite heavy model. Actually, the default model setting in this respository is light version of the model specified in the paper.