eifuentes / swae-pytorch Goto Github PK

Hi, thanks so much for the pytorch implementation of SWAE.

By the way, do you have any results such as FID score?

Thank you for creating a PyTorch version of the novel SWAE project.

Encountered an exception during runtime in line 48 of _sliced_wasserstein_distance

encoded_projections = encoded_samples.matmul(projections.transpose(0, 1))

with device being set to "cuda" in main()

Expected object of type torch.cuda.FloatTensor but found type torch.FloatTensor for arguement #2 'mat2'.

encoded_samples is a <Tensor, len() = 500> with is_cuda: True whereas
projections is a <Tensor, len() = 50> with is_cuda: False

System:
PyTorch 0.4.0
CUDA 9.0
Python 3.6.3 Anaconda/Intel
VS Code 1.25.1
Ubuntu Linux 18.04 x64

Regards.

After I ran python3 setup.py install, I modified the code to run again, but I found that it was still running code that was not modified before. Why? If you can answer it, I would be grateful.

Hi,

Thank you for sharing the code.

In the original implementation, I found that the function generateZ in MNIST_SlicedWassersteinAutoEncoder_Circle.ipynb does not generate sample points in a circle uniformly.
The generated sample points are more dense at the center of a circle. This can affect the resulting latent space.

I made an issue in the original implementation repository.

For this PyTorch implementation, It can be easily fixed by applying sqrt on random radius samples from a uniform distribution in rand_cirlce2d.

def rand_cirlce2d(batch_size):
    """ This function generates 2D samples from a filled-circle distribution in a 2-dimensional space.
        Args:
            batch_size (int): number of batch samples
        Return:
            torch.Tensor: tensor of size (batch_size, 2)
    """
    # r = np.random.uniform(size=(batch_size)) # before
    r = np.sqrt(np.random.uniform(size=(batch_size))) # after
    theta = 2 * np.pi * np.random.uniform(size=(batch_size))
    x = r * np.cos(theta)
    y = r * np.sin(theta)
    z = np.array([x, y]).T
    return torch.from_numpy(z).type(torch.FloatTensor)

Here is an related article: Generate a random point within a circle (uniformly)

Best,
Oh-Hyun