https://github.com/sandialabs/pdonet

Implementation and code to test the parameterized differential operator layer for "Parameterized Pseudo-Differential Operators for Graph Convolutional Neural Networks" GSP-CV 2021.
https://github.com/sandialabs/pdonet

scr-2563 snl-data-analysis

Last synced: 4 months ago
JSON representation

Implementation and code to test the parameterized differential operator layer for "Parameterized Pseudo-Differential Operators for Graph Convolutional Neural Networks" GSP-CV 2021.

• Host: GitHub
• URL: https://github.com/sandialabs/pdonet
• Owner: sandialabs
• License: bsd-3-clause
• Created: 2022-10-21T20:18:41.000Z (almost 4 years ago)
• Default Branch: master
• Last Pushed: 2022-11-21T16:50:21.000Z (over 3 years ago)
• Last Synced: 2025-03-01T00:25:55.929Z (over 1 year ago)
• Topics: scr-2563, snl-data-analysis
• Homepage:
• Size: 14.6 KB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Changelog: CHANGELOG.md
  • Contributing: Contributing.md
  • License: LICENSE.md

Awesome Lists containing this project

README

          
# PDONet: a graph convolutional layer based on differential operators implemented in PyTorch

Implementation and code to test the parameterized differential operator layer

for "Parameterized Pseudo-Differential Operators for Graph Convolutional Neural Networks" GSP-CV 2021.

## Installation

Some modules are dependent on python3 and code is written to run on python3 (given the long install times of some of the modules, taking the time to make sure you're using python is worthwhile if installing in a virtual environment).

Torch cluster/scatter expect certain packages to be available during install.

    pip3 install -r prereqs.txt

Note: Builds for torch-scatter and torch-sparse can be extremely slow.

    pip3 install -r requirements.txt

## Running

Datasets should be download as needed for runs.  MNIST hierarchical, CIFAR10, and CIFAR100 will have an extra delay as superpixel datasets are built.  Unfortunately, using a pruned dataset will result in an additional download despite being the same raw data as the superpixel version (this can be avoided by manually copying the raw folder to the *-SP-Pruned directory).

Each will train and validate against the test set and report the test accuracy for the best training epoch.

Running the superpixel classifier with the default configuration options:

    python3 run_classifier.py

Running the FAUST node correspondence task with default configuration:

    python3 run_faust.py

To override config options and/or load in additional configuration files:

    python3 run_classifier.py with depth=3             # set the model to use 3 layers

    python3 run_classifier.py with configs/cifar10.yml # override settings from a yml file

To run any of the classifiers with the hierarchical datasets, run a session such as:

    python3 run_classifier.py with hierarchical=True                      # hierarchical with default dataset (MNIST)

    python3 run_classifier.py with configs/cifar100.yml hierarchical=True # hierarchical with CIFAR100

In addition to the text output, full runs are logged to

`output/file\_observer/.`

## Datasets

Additional override configuration files are provided for MNIST hierarchical, CIFAR10, and CIFAR100 under `configs/`.  All configuration files correspond to the optimal values used in the paper.

## Configuration options

##### note

Blank by default.

Use for experiment notes at run time.

##### dataset

MNIST by default.  Classifier only.

Other options are CIFAR10 and CIFAR100.

##### batch\_size

Default 128.

##### learning\_rate

Default 0.0002.

##### weight\_decay

Default 0.0001.

##### device

Default cuda.

Use to specify a particular GPU to use on multigpu systems.

##### epochs

Default 100.

##### prune\_edges

Default False.  Classifier only.

Whether to reduce edges by replacing edge connections via Delauney triangulation.

##### hierarchical

Default False.  Classifier only.

Whether to include hierarchical graphs as described in Knyazev et al. 2019 https://arxiv.org/pdf/1907.09000.pdf.  Note: generating a CIFAR hierarchical graph is an order of magnitude slower than its superpixel counterpart.

##### edge\_dropout

Default 0.45.

Probability of dropping edges on input graph.

Following options are nested under model\_options:

##### depth

Default 7.

Number of convolutional/downsampling layers in model.

##### dropout\_rate

Default 0.5.

Dropout rate prior to fully connected layers.

##### grid\_scale

Default [0, 28].

Describes the relative scale of the node positions (assumes

equal scale for x and y)

##### initial\_features

Default 128.

Number of channels to output after first layer.  Doubled every

subsequent layer up to...

##### max\_features

Default 512.

Highest number of channels allowed.

##### voxels\_at\_depth

Default 3.  Classifier only.

After the downsampling layers, voxel grid should have an n x n

grid.  Voxels are halved in size each layer.

## Citation

Potter, Kevin et al. _Parameterized Pseudo-Differential Operators for Graph Convolutional Neural Networks_ GSP-CV 2021, Montreal, Canada, October 2021.

## License

Revised BSD.  See the LICENSE.txt file.