Can you provide details on you configuration (theano version especially) about maf HOT 3 CLOSED

That's odd. I've never encountered this issue.

When I originally run the experiments, I used Theano v0.9.0.

I tried to reproduce the error with Theano v0.9.0 and Theano v1.0.2 (latest version), but I couldn't. The code seems to run fine as is.

My understanding is that you're issuing

python run_experiments.py mnist

and then the error happens when Conditional MAF is to be trained. When a model is being trained, the code displays training info on screen, something like:

Epoch = 1, train loss = 839.124948406, validation loss = 1755.41093458
Epoch = 2, train loss = 709.554258655, validation loss = 1751.73216701
...

Can you provide more details on when exactly you encounter the error? What does the code output up to the point the error happens? Are the other models (MADE, MADE MoG, RealNVP, and their conditional versions) trained successfully?

from maf.

@tdeboissiere Sorry to use this discuss on another question.

I want to train a mafs.ConditionalMaskedAutoregressiveFlow, but the model loss always go to negative.

It starts at like 58.7, but after few steps(about 800) it goes negative.
something like: Epoch 3 - Step 221102 - loss -426.319 - lr 1.67e-05 - 0.32 s/step

Does this normal for maf loss go to negative? because the determine Jacoian term?
If this is not normal, any suggestions I can debug this? lack or too much of params to learn? need more regularization?

I am using tensorflow code which adapt Theano to Tensorflow: https://github.com/spinaotey/maf_tf

from maf.

This is totally normal. Log-likelihood loss doesn't have to be positive, its lowest value can be negative.

from maf.