when i trained the cifar folder,i go t'conditional_adversarial.net' ,but when i load it with 'torch.load("logs/conditional_adversarial.net")',error occurs. Could you give me some clues? Thanks

Hi,

For this demo code
"disp.image(...)" doesn't work for qlua/th
"image.display(...)" works for qlua

You do not seem to be using a nonlinearity in the large generative model?

https://github.com/facebook/eyescream/blob/master/lsun/model.lua#L52

In the small model and in the model generator you use ReLU? I assume that __R__ in the
model description is ReLU so I guess its just missing?

I was expecting the SpatialConvolutionUpsample class to do the expected "upsampling" but it seems like this class is doing something else. Here is one example:

conv = nn.SpatialConvolutionUpsample(1,1,1,1,3)
w, dw = conv:parameters()
w[1]:fill(1)
w[2]:zero()

This creates an upsampling module that upsamples the input image by a factor of 3, the convolution is 1x1 with weight 1 and bias 0 so it just copies the input.

I tried this on a 1x1x2x2 input tensor:

x = torch.range(1,4):resize(1,1,2,2)
y = conv:forward(x)

and here is the result:

th> x
(1,1,.,.) = 
  1  2
  3  4
[torch.DoubleTensor of size 1x1x2x2]

th> y
(1,1,.,.) = 
  1  2  3  4  1  2
  3  4  1  2  3  4
  1  2  3  4  1  2
  3  4  1  2  3  4
  1  2  3  4  1  2
  3  4  1  2  3  4
[torch.DoubleTensor of size 1x1x6x6]

However I was actually expecting y to be like this (which I think is the more standard "upsampling"):

1 1 1 2 2 2
1 1 1 2 2 2
1 1 1 2 2 2
3 3 3 4 4 4
3 3 3 4 4 4
3 3 3 4 4 4

The problem is, in the current SpatialConvolutionUpsample class, the new views created after computing the results do not play very well with element ordering. I wonder if this is the intended behaviour?

Hi,

How should we generate images using our trained model.
I am trying to generate images from a checkpoint state and I'm trying to use https://github.com/soumith/dcgan.torch/blob/master/generate.lua to generate images.
When I tried to convert the cuda tensor to a cpu-readable tensor using https://github.com/karpathy/neuraltalk2/blob/master/convert_checkpoint_gpu_to_cpu.lua , I get the following error:

/share/apps/torch/20160623/gnu/bin/luajit: convert_checkpoint_gpu_to_cpu.lua:79: bad argument #1 to 'pairs' (table expected, got nil)

Is there an easy way to generate images from the checkpointed model?

Is there some clues about how to replace ‘Linear‘ net with my designed ’SpatialConvolution ‘ net in the file ‘eyescream/cifar/scripts/train_cifar_classcond.lua’ ？

Does this code have a tensorflow version?
I am more familiar with tensorflow.

I think there is a small error in the L2/L1 regularization in the cifar models.

In the generative opfunc you regularize the discriminative parameters.

https://github.com/facebook/eyescream/blob/master/cifar/train/adversarial.lua#L87

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)

10

20

30

40

50

60

70

80

90

100

110