Real-NVP

PyTorch Implementation of the paper 'Density Estimation Using Real NVP' (ICLR 2017)

Requirements

• Python 3.6 >
• PyTorch 1.4 >

Details

• Training : 0.07s per epoch on TITAN Xp
• 4 affine coupling layers with 2 hidden layers MLP
• batch size 32, Adam with lr 0.0001

Results (num of affine coupling layers = 4)

After 1 epoch

After 2000 epochs

Results (num of affine coupling layers = 10)

After 500 epochs

Loss Curve

4 affine coupling layers

10 affine coupling layers

References

[1] Dinh, Laurent, Jascha Sohl-Dickstein, and Samy Bengio. "Density estimation using real nvp." arXiv preprint arXiv:1605.08803 (2016). (ICLR 2017)

[2] Dinh, Laurent, David Krueger, and Yoshua Bengio. "Nice: Non-linear independent components estimation." arXiv preprint arXiv:1410.8516 (2014). (ICLR 2015)