e-hulten / made Goto Github PK

when I running train_gaussian.py by cuda in GPU, I think that there is better to have with torch.no_grad(): before for _, dim in enumerate(order): in utils.plot.sample_digits_gaussian. The above advice is to solve the error :runtimeerror: cuda out of memory. Noted that this issue is appearing when you running train_gaussian.py by cuda in GPU.

In made.py, I find an error about building masked matrix. The condition about the last hidden layer to output layer should be ">" not ">=". I think the following codes are more reasonable:

        if i == (len(self.masks) - 1 - 1):
            for j in range(len(m_next)):
                # Use broadcasting to compare m_next[j] to each element in m.
                M[j, :] = torch.from_numpy((m_next[j] > m).astype(int))
        else:
            for j in range(len(m_next)):
                # Use broadcasting to compare m_next[j] to each element in m.
                M[j, :] = torch.from_numpy((m_next[j] >= m).astype(int))