Closed gitlabspy closed 3 years ago
Hi, Thanks!
I've also found GPU utilization to be low for flow models. I haven't investigated this further, but if someone has any insight, I'd love to hear it!
Regarding reconstruction: If you are using only bijections, the input should be exactly reconstructable (up to numerical error). However, if you make use of surjective transformations (such as Slice, MaxPool2d, etc.), some information in lost in the x->z direction which is generated again in the z->x direction. The reconstructions might therefore vary.
@didriknielsen It's kinda rude for me to ask you for this but I wonder if you can upload the pretrained model forimagenet64
please?
Hi Didrik, Great work from you!! I was trying to run experiments from codes you've given. I run it on cifar10 and I found it has low GPU usage while it occupies a lot of GPU memory. I wonder what causes this phenomenon?
39'C, 35 % | 23651 / 32510 MB
35% is almost the highest usage rate.Another thing is, I found that image is barely reconstructable. (cuz from what I understand, flow is bijective (obviously it could be other cases according to this paper) so it could kinda always recover image from latent and the performance is good even when multi-scale architecture's on. I only try on 32x32 size image, will it be good (more reconstructable) on higher resolution images?
I was running this
python train.py --epochs 500 --batch_size 32 --optimizer adamax --lr 1e-3 --gamma 0.995 --eval_every 1 --check_every 10 --warmup 5000 --num_steps 12 --num_scales 2 --dequant flow --pooling max --dataset cifar10 --augmentation eta --name maxpool
line of command and it runs about 300 epochs.