vq-vae-2-pytorch icon indicating copy to clipboard operation
vq-vae-2-pytorch copied to clipboard

Unable to get good results

Open yellowwoods12 opened this issue 5 years ago • 3 comments
trafficstars

I was training the model with the default parameters set in the code over a small dataset (around 5k images) but cannot obtain any good results. This is the generated output after sampling :

vae-result

Is this because the hyper-parameters are not set properly or because the dataset is small? How can I get significant results using this model?

yellowwoods12 avatar May 20 '20 06:05 yellowwoods12

Hey @rosinality, I tried training my model according to your suggestions and was able to complete stage 1 and PixelSNAIL training top hier training for 420 epochs with a batch size of 10. Now when I try to train PixelSNAIL for even batch size 1, I get this error:

CUDA Error

I trained the entire model before once where I didn't get any such issue. Let me know what can be done here.

yellowwoods12 avatar May 23 '20 18:05 yellowwoods12

You may need to reduce model sizes, for example, decrease --n_res_block.

rosinality avatar May 23 '20 22:05 rosinality

@rosinality Thanks, it worked!

yellowwoods12 avatar May 25 '20 10:05 yellowwoods12