для того, чтобы 让这个项目更加 in line with インターネット の Son Sürüm
the following content will be written by crappy machine translation.
This project calculates the gradient manually, further explains the underlying calculation process when using loss.bakcward() in pytorch
You can use above routines to see how to obtain the grad of weight in linear 、 conv2d manually.
To run the above routines, you need to make sure that the following libraries installed in your environment.
conda create -n manual_back python=3.5
conda activate manual_back
pip install graphviz
pip install torch torchvision torchaudio
Similar to ffmpeg, graphviz is also a separate software, the third-party library only provides an interface to use the software through python, you also need to go to the official website to download and install.
Or you can comment draw_forward() and the related plot code (you can also see all results on the command line)
For detail explanation, please visit AluminiumOxide@bilibili
enjoy