Closed disadone closed 3 months ago
It seems that jax.value_and_grad
in hmc_utils.py
gives the key contribution to this, though I still don't understand why.
Hi @disadone, could you post the question on forum https://forum.pyro.ai/ We have a couple of inference utilities that you might want to use for diagnosing the issues of your model/data.
Hi @disadone, could you post the question on forum forum.pyro.ai We have a couple of inference utilities that you might want to use for diagnosing the issues of your model/data.
Thank you, I move my post to the forum. It would be appreciate if there is any potential_energy
example.
Gibbs
part is basically erased in the following code because it is not related to the problem. I just don't know why thevv_state.z_grad
inhmc.py
is extremely large.https://github.com/pyro-ppl/numpyro/blob/0ba1306ea41c9865169360cb17c62efe7fc2bc94/numpyro/infer/hmc.py#L350
Here is the simplified version of the code. The acceptance rate keeps at 0. Is there any problem?
The outputs are: