dibgerge opened a new issue #19722: URL: https://github.com/apache/incubator-mxnet/issues/19722
## Problem statement I am probably wrong about those, but I thought to ask here anyway. I was using `mxnet.gluon.probability.Categorical` from master, and two things came to my attention: 1. It requires a `num_events` argument, which feel unnecessary, and possibly could be eliminated? 2. [Possible bug] If providing `logit`, current `log_prob` implementation assumes we are providing `log softmax(x)`. Usually, when providing `logit`, it is the linear output of the net. This is the behavior also in torch/tensorflow, where internally, the logit is scaled by `logsumexp`. ## Proposed solutions 1. Remove `num_events` argument 2. If `logit` provided in constructor, shift it by `logsumexp` ## References - [Tensorflow implementation](https://github.com/tensorflow/probability/blob/v0.11.1/tensorflow_probability/python/distributions/categorical.py#L289) - [Torch implementation](https://github.com/pytorch/pytorch/blob/master/torch/distributions/categorical.py#L55) ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@mxnet.apache.org For additional commands, e-mail: issues-h...@mxnet.apache.org
