Luyu Wang

I tried to clip the weights and it will run a few epochs, but it seems this code simply does not work.

@prafullasd @dpkingma thanks for uploading the pretrained models and logs - super helpful! I notice in the log, for example, of the ablation experiments, there are a few parameters, including...

So I asked the authors at NeurIPS last year - using `sigmoid` here is to bound the gradients of the affine coupling layer. In the previous Real-NVP work a `tanh`...

I think the problem is not that the gradient is not strong enough. Actually quite the opposite you wanna bound it.

Note that `y = scale * x + shift` and `scale = tf.nn.relu(h[:, :, :, 1::2]), shift = h[:, :, :, ::2]`. I agree that `dy/dh` is bounded here due...