pystencils-autodiff¶
This is the documentation of pystencils-autodiff.
This document assumes that you are already familiar with pystencils. If not, here is a good tutorial to start.
Installation of this Auto-Diff Extension¶
Install via pip :
pip install pystencils-autodiff
or if you downloaded this repository using:
pip install -e .
Then, you can access the submodule pystencils.autodiff.
import pystencils.autodiff
Usage¶
Create a pystencils.AssignmentCollection with pystencils:
import sympy
import pystencils
z, y, x = pystencils.fields("z, y, x: [20,30]")
forward_assignments = pystencils.AssignmentCollection({
z[0, 0]: x[0, 0] * sympy.log(x[0, 0] * y[0, 0])
})
print(forward_assignments)
Subexpressions:
Main Assignments:
z[0,0] ← x_C*log(x_C*y_C)
You can then obtain the corresponding backward assignments:
from pystencils.autodiff import AutoDiffOp, create_backward_assignments
backward_assignments = create_backward_assignments(forward_assignments)
# Sorting for reproducible outputs
backward_assignments.main_assignments = sorted(backward_assignments.main_assignments, key=lambda a: str(a))
print(backward_assignments)
You can see the derivatives with respective to the two inputs multiplied by the gradient diffz_C of the output z_C.
Subexpressions:
Main Assignments:
\hat{x}[0,0] ← diffz_C*(log(x_C*y_C) + 1)
\hat{y}[0,0] ← diffz_C*x_C/y_C
You can also use the class pystencils_autodiff.AutoDiffOp to obtain both the assignments (if you are curious) and auto-differentiable operations for Tensorflow…
op = AutoDiffOp(forward_assignments)
backward_assignments = op.backward_assignments
tensorflow_op = op.create_tensorflow_op(backend='tensorflow_native', use_cuda=False)
… or Torch:
torch_op = op.create_tensorflow_op(backend='torch_native', use_cuda=False)
Contents¶
License¶
pystencils_autodiff – Adds automatic differentation and generation of Tensorflow/PyTorch operations to pystencils
Copyright © 2019 Stephan Seitz
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see <http://www.gnu.org/licenses/>.
Contributors¶
- Stephan Seitz <stephan.seitz@fau.de>
Changelog¶
0.2.2¶
- Add possibility to overwrite nvcc arch for Tensorflow compilation: PYSTENCILS_TENSORFLOW_NVCC_ARCH
- Add possibility to compile Tensorflow module without loading
0.2.1¶
- Bugfix: tensorflow_jit had erroneous code for writing to pystencils’ config file
0.2.0¶
- Compilation of Torch/Tensorflow/pybind11 modules with an AST (instead of pure jinja as before)