voxnet

3D/Volumetric Convolutional Neural Networks with Theano+Lasagne.

Installation

voxnet is based on Theano and Lasagne.

You will also need path.py and scikit-learn. Scikit-learn is used purely for evaluation of accuracy and is an easily removable dependency.

You can do something like

git clone git@github.com:dimatura/voxnet.git
cd voxnet
pip install --editable .

ModelNet10 Example

Get data

In this example we will use the ModelNet 10 dataset, from the excellent 3D ShapeNet project.

To make our life easier we will use the voxelized version, which is included in the source code distribution. Unfortunately, it comes in evil .mat files, so we will convert them to a more python-friendly data format first.

scripts/download_shapenet10.sh will try to download and convert the data for you. This may take a while.

# scripts/download_shapenet10.sh
wget http://3dshapenets.cs.princeton.edu/3DShapeNetsCode.zip 
unzip 3DShapeNetsCode
python convert_shapenet10.py 3DShapeNets

If you're curious, the data format is simply a tar file consisting of zlib-compressed .npy files. Simple and effective!

Train model

We will be messy and do everything in the scripts/ directory.

cd scripts/
python train.py config/shapenet10.py shapenet10_train.tar

config/shapenet10.py stores the model architecture and hyperparameters related to training as Python code. train.py loads this code dynamically, compiles the Theano model, and begins training with the data from shapenet10_train.tar. Note that compiling the Theano model might around a minute for the first execution. As soon as training begins, metrics will be printed to stdout and learned weights are periodically saved to weights.npz.

During training (which will take a few hours) you can monitor progress visually by by running scripts/train_reports.py. Note that this script has a few dependencies, including seaborn and pandas. The script uses training metrics stored in a file called metrics.jsonl. This is simply a format with one json record per line (inspired by JSON Lines).

Test model

python test.py config/shapenet10.py shapenet10_test.tar --out-fname out.npz

test.py uses the same model as train.py, but only for classifying instances from the test set. It performs simple evaluation and optionally saves the predictions in an .npz file.

If you don't want to train your own, you can use the --weights option with the an example result of running this script, shapenet10_weights.npz. This file was committed with Git LFS, so you can use that, or simply download the raw version from github.

Visualize

If you have CPU cycles to burn, try

python output_viz.py out.npz shapenet10_test.tar out.html

This will randomly select 10 instances from the test set, render them with a very very inefficient renderer, and create a small page called out.html with the renders, the ground truth and the predicted label (see example above). Requires gizeh.

Reference

@inproceedings{maturana_iros_2015,
    author = "Maturana, D. and Scherer, S.",
    title = "{VoxNet: A 3D Convolutional Neural Network for Real-Time Object Recognition}",
    booktitle = "{IROS}",
    year = "2015",
    pdf = "/extra/voxnet_maturana_scherer_iros15.pdf",
}

TODO

• Cleaning up
• Testing
• More options
• Better visualization
• Use new cudnn 3D capabilities