IsoGCN code for ICLR2021

Last update: Nov 25, 2022

Overview

IsoGCN

The official implementation of IsoGCN, presented in the ICLR2021 paper Isometric Transformation Invariant and Equivariant Graph Convolutional Networks [arXiv].

Please cite us as:

@inproceedings{
horie2021isometric,
title={Isometric Transformation Invariant and Equivariant Graph Convolutional Networks},
author={Masanobu Horie and Naoki Morita and Toshiaki Hishinuma and Yu Ihara and Naoto Mitsume},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=FX0vR39SJ5q}
}

General notice

If some of the following steps not working, please modify User settings section in the Makefile to fit with your environment.

Installation

You can either locally install locally or use Docker image. However, to generate an anisotropic nonlinear heat equation dataset, we recommend using Docker.

Local install

We are using poetry, thus first install it following the instruction on https://python-poetry.org/docs/ . Then, update the Makefile to PYTHON ?= 'poetry run python3' or explicitly specify the PYTHON environment variable on the execution of the make command.

For GPU environment,

PYTHON='poetry run python3' make poetry
poetry install
PYTHON='poetry run python3' make install_pyg_gpu

For CPU environment,

PYTHON='poetry run python3' make poetry
poetry install
PYTHON='poetry run python3' make install_pyg_cpu

, and set GPU_ID = -1 in the Makefile.

Also, optionally, please install FrontISTR and gmsh to generate an anisotropic nonlinear heat equation dataset (which are installed in the Docker image).

Docker image

Please download the docker image via https://savanna.ritc.jp/~horiem/isogcn_iclr2021/images/isogcn.tar, then place the image in the images directory. After that, plsease make in to login the docker befor perfroming all the following processes.

Differential operator dataset

Data generation

make differential_data

Training IsoGCN

make scalar2grad  # Scalar to gradient task
make scalar2grad  ADJ=5  # Scalar to gradient task with # hops = 5
make scalar2hessian  # Scalar to Hessian task
make grad2laplacian  # Gradient to Laplacian task
make grad2hessian  # Gradient to Hessian task

Training baseline models

make scalar2grad_baseline BASELINE_NAME=gcn  # BASELINE_NAME=[cluster_gcn, gcn, gcnii, gin, sgcn]

Similarly, one can perform baseline model trainings for other tasks.

Anisotropic nonlinear heat equation dataset

Run whole process with small data to check the process (Optional)

It generates a small dataset to simulate the whole process of data generation, preprocessing, training, and inference. This process requires either FrontISTR installed locally or Docker image.

make small_heat_nl_tensor_pipeline

Dataset download

The dataset containing finite element analysis results is generated from the ABC dataset using gmsh for meshing and FrontISTR for analysis.

Please download the dataset you need. (Note: To perform only training, you need only 'preprocessed' data.) The dataset can be downloaded via:

Raw data (FrontISTR analysis files)
- Train dataset
  - https://savanna.ritc.jp/~horiem/isogcn_iclr2021/data/heat_nl_tensor/raw/train_50.tar.gz
- Validation dataset
  - https://savanna.ritc.jp/~horiem/isogcn_iclr2021/data/heat_nl_tensor/raw/validation_16.tar.gz
- Test dataset
  - https://savanna.ritc.jp/~horiem/isogcn_iclr2021/data/heat_nl_tensor/raw/test_16.tar.gz
Interim data (.npy and .npz files before standardization)
Preprocessed data (.npy and .npz files after standardization)

After download finished, please merge the split archives with:

cat train_50.tar.gz.parta* > train.tar.gz

, extract them with tar xvf *.tar.gz, then place them in the corresponding data/heat_nl_tensor/(raw|interim|preprocessed) directory.

Training IsoGCN

make heat_nl_tensor

Training baseline models

make heat_nl_tensor_baseline BASELINE_NAME=gcn  # BASELINE_NAME=[cluster_gcn, gcn, gcnii, gin, sgcn]

IsoGCN core implementation

The core implementation of the IsoGCN layer is separated in the library SiML and can be found here. Also, the code to generate IsoAMs is separated in the library Femio and can be found here.

License

Apache License 2.0.

IsoGCN code for ICLR2021

Related tags

Overview

IsoGCN

General notice

Installation

Local install

Docker image

Differential operator dataset

Data generation

Training IsoGCN

Training baseline models

Anisotropic nonlinear heat equation dataset

Run whole process with small data to check the process (Optional)

Dataset download

Training IsoGCN

Training baseline models

IsoGCN core implementation

License

Owner

horiem

4D Human Body Capture from Egocentric Video via 3D Scene Grounding

SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021)

Code for the head detector (HeadHunter) proposed in our CVPR 2021 paper Tracking Pedestrian Heads in Dense Crowd.

PyTorch implementation for SDEdit: Image Synthesis and Editing with Stochastic Differential Equations

Deepparse is a state-of-the-art library for parsing multinational street addresses using deep learning

WHENet: Real-time Fine-Grained Estimation for Wide Range Head Pose

Supervised Classification from Text (P)

PyTorch implementation of the ACL, 2021 paper Parameter-efficient Multi-task Fine-tuning for Transformers via Shared Hypernetworks.

Learning Dense Representations of Phrases at Scale (Lee et al., 2020)

Public repository created to store my custom-made tools for Just Dance (UbiArt Engine)

Google Brain - Ventilator Pressure Prediction

Framework for Spectral Clustering on the Sparse Coefficients of Learned Dictionaries

A PyTorch Toolbox for Face Recognition

Context Decoupling Augmentation for Weakly Supervised Semantic Segmentation

The Multi-Mission Maximum Likelihood framework (3ML)

Image-Stitching - Panorama composition using SIFT Features and a custom implementaion of RANSAC algorithm

PyTorch code for DriveGAN: Towards a Controllable High-Quality Neural Simulation

Unofficial Tensorflow 2 implementation of the paper Implicit Neural Representations with Periodic Activation Functions

This is an open source library implementing hyperbox-based machine learning algorithms

这是一个mobilenet-yolov4-lite的库，把yolov4主干网络修改成了mobilenet，修改了Panet的卷积组成，使参数量大幅度缩小。