RuntimeError: CUDA error: out of memory

[damtharvey]

How much memory do I need for Capitals64? I have 11 GB.

[azadis]

That should be enough. Do you have a batch size larger than 150?

[damtharvey]

Sorry for the slow reply. I will try again when I get my GPU back after NIPS.

[sesebuckin]

I have 11GB too. I tried to train Capitals64 with batch size=150, but CUDA out of memory.
There's no problem with batch size=64. (batch size=120 will fails too. I haven't try size between 64 and 120)

[damtharvey]

I'm back. Trying it again, it seems like it doesn't use my GPU anymore.

I found the previous base_options.py and reverted to it. Tried it again with batch sizes 64 and 1 and still get CUDA error: out of memory.

[azadis]

the issue of not using GPU is fixed now!
I am not sure about your cuda out of memory error. Can yoy provide more details?

[leitianqi]

@azadis
How can I solve this problem?

Total number of parameters: 291649

model [cGANModel] was created
create web directory ./checkpoints/GlyphNet_pretrain/web...
Traceback (most recent call last):
File "train.py", line 32, in
model.optimize_parameters()
File "/home/ltq/FontTransfer/MC-GAN/models/cGAN_model.py", line 242, in optimize_parameters
self.backward_G()
File "/home/ltq/FontTransfer/MC-GAN/models/cGAN_model.py", line 224, in backward_G
self.loss_G.backward()
File "/usr/local/lib/python2.7/dist-packages/torch/tensor.py", line 93, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/usr/local/lib/python2.7/dist-packages/torch/autograd/init.py", line 90, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: CUDA error: out of memory