Experimental Art Projects with PyTorch
$ pip3 install -r ./requirements.txt
$ export PYTHONPATH="$PWD:$PYTHONPATH"
These images are visualized filters for VGG16 filters. Model is from torchvision's pretrained model.
You can check sample images in this directory.
Layer: model.features[3], Filter: 34(Left) 39(Center), 52(Right)
Layer: model.features[10], Filter: 1(Left) 163(Center) 237(Right)
Layer: model.features[29], Filter: 33(Left) 132(Center), 390(Right)
You can use layer and filters options to specify target number of layer and filters.
$ python3 torchart/filtervisualization/main.py
$ python3 torchart/filtervisualization/main.py layer=3 filters='[23, 34, 39, 52]'
$ python3 torchart/filtervisualization/main.py layer=10 filters='[1, 161, 163, 237, 241]'
$ python3 torchart/filtervisualization/main.py layer=29 filters='[5, 33, 132, 177, 241, 286, 312]'
If you want to change more hyper parametrs, Please check torchart/filtervisualization/config.yaml.
We are using hydra. And you can change hyper parameters as command line arguments.
Output directory structure is
outputs
└── yyyy-mm-dd
├── hh-mm-ss
│ ├── main.log
│ ├── vgg16_{layer_number}_{filter_number}.png
...
$ pytest torchart
$ flake8 torchart
$ mypy torchart
$ isort -c torchart