Reproducing Mocap Experiments and possible bug in code

[abdulfatir]

Hi!

Thanks a lot for releasing a clean implementation of your model.

I am trying to reproduce the results reported in the paper for the Mocap dataset. Can I confirm that the default arguments in this file correspond to the best results reported in the paper (most of the things seem to match except the number of epochs)? If not, could you please share the best config?

There also appears to be a bug in the following line:

scalable-inference-in-sdes/experiments/mocap/train_approx_ode.py

Line 446 in 3519d34

loss = recon_nll + encoder_kl*self.gamma + weight**model_kl + vae_init_loss

Should it be weight*model_kl instead of weight**model_kl?

I tried training the model both before and after fixing this potential bug and I got the following test MSE values:

weight**model_kl, decoder_dist: True, epochs: 1500 = 19.74 +/- 0.04
weight*model_kl, decoder_dist: True, epochs: 1500 = 17.32 +/- 0.03

[abdulfatir]

Hi! Any update on this?