sagelywizard/pytorch-mdn

Mixture Density Networks for PyTorch

pytorch-mdn's Issues

installing issue

I had an issue installing this repo by setup.py:
pip install git+https://github.com/sagelywizard/pytorch-mdn
the error is:
error: error in 'egg_base' option: 'src' does not exist or is not a directory

changing src to mdn in this line and this line solved the problem. I guess the issue is solved for me but not sure it should be corrected in the source file or not

Bug: `gaussian_probability` - `data is not defined`

L75 in mdn.py returns data is not defined.
data needs to be replaced by target.

Speedup `sample` method

The sample method can be sped up ~10x by using torch.gather

def sample(self, pi, sigma, mu):
        """Draw samples from a MoG."""
        categorical = Categorical(pi)
        pis = categorical.sample().unsqueeze(1)
        sample = Variable(sigma.data.new(sigma.size(0), 1).normal_())
        # Gathering from the two Gaussian Distribution based on sampled indices
        sample = sample * sigma.gather(1, pis) + mu.gather(1, pis)
        return sample

Typo?

pytorch-mdn/mdn.py

Line 75 in 8bd3415

ret = ONEOVERSQRT2PI * torch.exp(-0.5 * ((data - mu) / sigma)**2) / sigma

Should data be target?

Numerical instability problem

First want to thank the creator of this repo, although simple it is a great and convenient implementation of MDN for the community.

During training, there is numerical instability problems arising from the multiplication of the probability that can be changed to addition.

typo

In def mdn_loss(pi, sigma, mu, data)

data should be target