Bad results for train_cifar.lua
simopal6 opened this issue · 0 comments
simopal6 commented
Running scripts/train_cifar.lua
with default parameters generates the attached images at different epochs (I changed the checkpoint-saving part to append the epoch number). What may be wrong?
Here is the code I used for generation:
l = torch.load("adversarial-" .. e .. ".net")
i = torch.CudaTensor(42, 100):uniform(-1,1)
l.G:cuda()
l.G:evaluate()
o = l.G:forward(i)
img = image.toDisplayTensor(o)
image.save("gen-" .. e .. ".png", img)