The TorchRay package implements several visualization methods for deep convolutional neural networks using PyTorch. In this release, TorchRay focuses on attribution, namely the problem of determining which part of the input, usually an image, is responsible for the value computed by a neural network.

TorchRay is research oriented: in addition to implementing well known techniques form the literature, it provides code for reproducing results that appear in several papers, in order to support reproducible research.

TorchRay was initially developed to support the paper:

• Understanding deep networks via extremal perturbations and smooth masks. Fong, Patrick, Vedaldi. Proceedings of the International Conference on Computer Vision (ICCV), 2019.

The package contains several usage examples in the examples subdirectory.

Here is a complete example for using GradCAM:

from torchray.attribution.grad_cam import grad_cam
from torchray.benchmark import get_example_data, plot_example

# Obtain example data.
model, x, category_id, _ = get_example_data()

# Grad-CAM backprop.
saliency = grad_cam(model, x, category_id, saliency_layer='features.29')

# Plots.
plot_example(x, saliency, 'grad-cam backprop', category_id)

TorchRay requires:

• Python 3.4 or greater
• pytorch 1.1.0 or greater
• matplotlib

For benchmarking, it also requires:

• torchvision 0.3.0 or greater
• pycocotools
• mongodb (suggested)
• pymongod (suggested)

On Linux/macOS, using conda you can install

while read requirement; do conda install \
-c defaults -c pytorch -c conda-forge --yes $requirement; done <<EOF
pytorch>=1.1.0
pycocotools
torchvision>=0.3.0
mongodb
pymongo
EOF

Using pip:

pip install torchray

From source:

python setup.py install

or

pip install .

The full documentation can be found here.

See the CHANGELOG.

• Website: https://github.com/facebookresearch/TorchRay

See the CONTRIBUTING file for how to help out.

TorchRay has been primarily developed by Ruth C. Fong and Andrea Vedaldi.

TorchRay is CC-BY-NC licensed, as found in the LICENSE file.