pytensor
pytensor copied to clipboard
pymc-devs
PyTensor allows you to define, optimize, and efficiently evaluate mathematical expressions involving multi-dimensional arrays.
### Before _No response_ ### After _No response_ ### Context for the issue: Currently, the Cholesky `Op` does not act in-place. This is a known problem since the Theano days...
### Description This is different than numpy ```python import numpy as np import pytensor.tensor as pt pt.random.normal([0], size=None).eval() # array([-0.57354733]) pt.random.normal([0], size=()).eval() # array([-2.39088062]) np.random.normal([0], size=None) # array([0.10438181]) np.random.normal([0], size=()).eval()...
### Before Currently the `val` argument in `fill_diagonal` [only accepts scalar values](https://github.com/pymc-devs/pytensor/blob/288a3f3481e7dfa6162e1b60a3909a8f67b29337/pytensor/tensor/extra_ops.py#L952-L982): fill_diagonal(a, val) For example, fill_diagonal(pt.ones(5), 3) will fill the diagonal element with 3. ### After This function should...
### Description That gets unrolled at some stage of the compilation pipeline. This allows us and users to easily rewrite gradient graphs before they get "messy". This covers a lot...
https://github.com/pymc-devs/pytensor/blob/9df55cc6c507ea5d78f88166664ab1325d1fd30d/tests/unittest_tools.py#L280-L282
### Describe the issue: The `ProdWithoutZeros` `Op` arises in the gradients of `pt.prod`. This currently cannot be compiled to gradient mode unless we specifically pass `no_zeros_in_input=True`. I guess we would...
### Description Working on https://github.com/pymc-devs/pymc/pull/6996 I added some functionality to rewrite inner graphs, but we want to move it into pytensor https://github.com/pymc-devs/pymc/pull/6996#discussion_r1411821274. It should be > `WalkingNestedGraphRewriter` which applies the...
### Description As discussed in https://github.com/pymc-devs/pytensor/issues/222 Function is quite bloated. Some things we could probably remove because nobody uses them: 1. Default values for function arguments 2. output_subset kwarg at...
### Motivation for these changes As requested in: - #449 ### Implementation details You can specify both a package positional argument and (one or more) `--file` / `-f` argument which...
I took the branch from #407 and added a `pytensor.xtensor.spaces` module that defines types to distinguish between "unordered spaces" (`BaseSpace`) and "ordered spaces" (`OrderedSpace`). `BaseSpace` and `OrderedSpace` are similar to...
