PyTensor is a Python library that allows one to define, optimize, and efficiently evaluate mathematical expressions involving multi-dimensional arrays. It provides the computational backend for PyMC.
- A hackable, pure-Python codebase
- Extensible graph framework suitable for rapid development of custom operators and symbolic optimizations
- Implements an extensible graph transpilation framework that currently provides compilation via C, JAX, and Numba
- Contrary to PyTorch and TensorFlow, PyTensor maintains a static graph which can be modified in-place to allow for advanced optimizations
import pytensor
from pytensor import tensor as pt
# Declare two symbolic floating-point scalars
a = pt.dscalar("a")
b = pt.dscalar("b")
# Create a simple example expression
c = a + b
# Convert the expression into a callable object that takes `(a, b)`
# values as input and computes the value of `c`.
f_c = pytensor.function([a, b], c)
assert f_c(1.5, 2.5) == 4.0
# Compute the gradient of the example expression with respect to `a`
dc = pytensor.grad(c, a)
f_dc = pytensor.function([a, b], dc)
assert f_dc(1.5, 2.5) == 1.0
# Compiling functions with `pytensor.function` also optimizes
# expression graphs by removing unnecessary operations and
# replacing computations with more efficient ones.
v = pt.vector("v")
M = pt.matrix("M")
d = a/a + (M + a).dot(v)
pytensor.dprint(d)
# Add [id A]
# ββ ExpandDims{axis=0} [id B]
# β ββ True_div [id C]
# β ββ a [id D]
# β ββ a [id D]
# ββ dot [id E]
# ββ Add [id F]
# β ββ M [id G]
# β ββ ExpandDims{axes=[0, 1]} [id H]
# β ββ a [id D]
# ββ v [id I]
f_d = pytensor.function([a, v, M], d)
# `a/a` -> `1` and the dot product is replaced with a BLAS function
# (i.e. CGemv)
pytensor.dprint(f_d)
# Add [id A] 5
# ββ [1.] [id B]
# ββ CGemv{inplace} [id C] 4
# ββ AllocEmpty{dtype='float64'} [id D] 3
# β ββ Shape_i{0} [id E] 2
# β ββ M [id F]
# ββ 1.0 [id G]
# ββ Add [id H] 1
# β ββ M [id F]
# β ββ ExpandDims{axes=[0, 1]} [id I] 0
# β ββ a [id J]
# ββ v [id K]
# ββ 0.0 [id L]See the PyTensor documentation for in-depth tutorials.
The latest release of PyTensor can be installed from PyPI using pip:
pip install pytensor
Or via conda-forge:
conda install -c conda-forge pytensor
The current development branch of PyTensor can be installed from GitHub, also using pip:
pip install git+https://github.com/pymc-devs/pytensor
PyTensor is a fork of Aesara, which is a fork of Theano.
We welcome bug reports and fixes and improvements to the documentation.
For more information on contributing, please see the contributing guide.
A good place to start contributing is by looking through the issues here.
