https://github.com/DeepLearnPhysics/dynamic-gcnn

Dynamic Graph Convolutional Neural Network for 3D point cloud semantic segmentation
https://github.com/DeepLearnPhysics/dynamic-gcnn

convolutional-neural-networks custom-data dgcnn gcnn graph-convolutional-networks python tensorflow

Last synced: 18 days ago
JSON representation

Dynamic Graph Convolutional Neural Network for 3D point cloud semantic segmentation

• Host: GitHub
• URL: https://github.com/DeepLearnPhysics/dynamic-gcnn
• Owner: DeepLearnPhysics
• License: mit
• Created: 2018-10-05T03:27:52.000Z (about 6 years ago)
• Default Branch: develop
• Last Pushed: 2018-11-07T22:21:51.000Z (about 6 years ago)
• Last Synced: 2024-08-01T03:46:09.257Z (4 months ago)
• Topics: convolutional-neural-networks, custom-data, dgcnn, gcnn, graph-convolutional-networks, python, tensorflow
• Language: Python
• Homepage:
• Size: 114 KB
• Stars: 60
• Watchers: 6
• Forks: 11
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.md

Awesome Lists containing this project

README

        
# dgcnn

This is an implementation of 3D point cloud semantic segmentation for [Dynamic Graph Convolutional Neural Network](https://arxiv.org/abs/1801.07829). The number of edge convolution layers, fully connected layers, and number of filters per each layer are all configurable. The implementation includes a few variations such as residual unit (edge convolution with identity mapping), with or without fully connected layers, etc.. Experimental results on DeepLearnPhysics open data set will be made available.

### Requirements

* `tensorflow >= v1.3`

* `numpy >= 1.13` 

* Optional requirements for IO include `h5py`, `larcv`

### Help

An executable script can be found at `bin/dgcnn.py`. The script takes `train` or `inference` arguments. Try `--help` to list available arguments:

```

bin/dgcnn.py train --help

```

### How to run

Below is an example of how to train the network using `mydata.hdf5` data file with `hdf5` format, 4 GPUs with batch size 24 and mini-batch size of 6, store snapshot every 500 iterations, print out info (loss,accuracy,etc) every 10 iterations, and store tensorboard summary every 50 iterations.

```

bin/dgcnn.py train --gpus 0,1,2,3 -bs 24 -mbs 6 -chks 500 -rs 10 -ss 50 -if mydata.hdf5 -io h5

```

See `--help` to find more flags and a descipriton for arguments.