Exercise three - extending Pytorch

In this exercise, you will use setuptools to compile a Pytorch extension with PyTorch's cpp_extension module.

The task is to implement symmetric padding extension for PyTorch. Symmetric padding, for example, is the default in pywt (https://pywavelets.readthedocs.io). PyTorch's pad lacks this functionality. In this exercise, we will provide it.

Testing and Verification

Follow these steps:

1. Clone the repository: git clone https://github.com/your-username/cpp_pad.git.
2. Navigate to the project directory: cd sympad_pytorch .
3. Run the tests with nox -s test.

Implementing your module in c++.

Symmetric padding extends a signal by mirroring samples. This mode is also known as half-sample symmetric [pywt-docs ]. The idea is to repeat samples in reverse order at the boundary.

... x2 x1 | x1 x2 ... xn | xn xn-1 ...

Implement this boundary extension mode in the src/sympad.cpp file's _pad_symmetric_1d-function by resolving the TODO's in the code.

Using your Pytorch module

1. create a virtual environment python -m venv virtual_env or a conda environment.

2. activate the environment source virtual_env/bin/activate

3. Install your module

   • pip install torch numpy build
   • python -m build --no-isolation
   • pip install dist/sympad-0.0.1-cp312-cp312-linux_x86_64.whl

4. To run your module, start the Python interpreter with python.

5. import torch, sympad and import numpy as np.

6. Pad a tensor using your own method.

   • Run array = np.arange(9).reshape(3,3)
   • print(array)
   • Run sympad.pad_symmetric(torch.from_numpy(array), [(1,1), (1,1)])