pytorch about zhusuan HOT 7 CLOSED

Maybe not. I'm personally preferring Tensorflow for several reasons. I may list them here in case other people have interests.

1. Static graphs will be much easier to optimize. Now TF may be reported slow under some settings. But it will soon change. The XLA compiler is not merged yet. Once done one is able to select optimization options when executing graphs (tf.OptimizerOptions.ON_1).

2. Dynamic graph style primitives are not that hard to support. Actually Tensorflow has made attempts towards this. See the tf.contrib.eager module. It will be merged into core of TF in future. I'm expecting more such as defining a function with dynamic style of code, and adding a decorator to transform to a static graph. And TF eager is actually showing some examples like this.

3. Distributed execution is better supported in TF. Pytorch's distributed module is still only some abstractions over common protocols.

4. Many other reasons such as larger community, not good enough but complete documents, ...

from zhusuan.

Many thanks for reply.
Just wondering is there a model zoo for the–state–of–the–art DGMs developed in your group?
If there is, it would be much easier to follow the published papers.

from zhusuan.

When I run the bayesian_nn.py, the results show that "from examples import conf, ImportError: No module named 'examples'". However, I have installed the examples by "pip install ".[examples]"",. I wonder whether there is anything others need to be done before running it??

from zhusuan.

@renjinhao Since this is unrelated to this issue, can you raise your question in a new issue?

from zhusuan.

@ifangcheng In general any paper published with code should be released under the thu-ml organization. If you don't find them, that's probably due to it is still not ready for a release.

from zhusuan.

@thjashin I noted a DGM project (mmdcgm ) under the thu-ml organization, However, it seems that it is NOT developed based on zhusuan ?
Any zhusuan version about it? or will release in future?

from zhusuan.

@ifangcheng This is an early work before we have zhusuan, so unfortunately not :(

from zhusuan.