search

dhgrs / chainer-VQ-VAE

A Chainer implementation of VQ-VAE.
82 stars 19 forks source link

out of memory to allocate #2

Closed jiqizaisikao closed 5 years ago

jiqizaisikao commented 6 years ago

Hi,I used the cpu version but is seems not to work,and i tried to run gpu version but end with err:cupy.cuda.memory.OutofMemoryError:out of memory to allocate 65536000 bytes(total 7991438336) My gpu is nvidia GTX1070

dhgrs commented 6 years ago

Hi, how do you set parameters in opt.py? It uses nearly 5.5GB(GPU) with parameters of newest commit( 7f8c94b ). GTX1070 has 8GB memories so may be able to run this code. Please check opt.py or other programs are running on same GPU.

jiqizaisikao commented 6 years ago

I download your source code yesterday on my ubuntu k=512 d=512 the n_channel1=256 the n_channel2=512 the n_channel3=256 your opt.py are much different from each other,maybe i should change it to be smaller

dhgrs commented 6 years ago

Yes, I'm searching the best parameters so opt.py are much different from each other. And the newest version that you try uses 5.5GB. Other parameters batchsize, length, n_loop, n_layer and n_filter change memory usage of GPU. If you changed these parameters, please try default parameters of newest commit. Althogh you can't run with these parameters, please tell me the version of chainer. I'm using chainer v3.2.0.