Setting up training:

1. Install the Anaconda package manager (https://repo.continuum.io/archive/). Code works with Python 2 & 3.
2. Install the following packages: PyTorch: conda install pytorch=0.4.0 cuda90 torchvision -c pytorch (>= 0.4.1 has some compile issues) Tensorflow (only the CPU version - for tensor board visualisation, pip install tensorflow), OpenCV (conda install opencv -c menpo) configargparse (conda install configargparse)
3. Compile the code with: sh make.sh (in the main se3nets-pytorch folder)
4. To train the se3 pose nets: python train_se3posenets.py -c <config_file> For an example config file, look at config/icra18final/simdata/se3pose/def_rmsprop.yaml You need to change the “data” path inside the config file to be your path to the dataset, specifically the <path-to-data> should be set to the directory where the dataset is located (see below for link to the dataset).
5. To train the baseline se3/flow networks: python train_flow_se3_nets.py -c <config_file> For the flow network, the config file is at config/icra18final/simdata/flow/def.yaml For the se3 net, the config file is at config/icra18final/simdata/se3/def.yaml
6. You can visualize the errors and predictions using tensorboard: tensorboard --logdir=<log_dir> --port=8000 where log_dir can be the path to the directory specified in the config file as save-dir.

Simulated Baxter Data:

You can download the simulated Baxter dataset (~53G) described in the paper here: https://drive.google.com/open?id=1OL08RzPXJA89KQnmbImdcZXdxkH3Ee5U

Use rclone to download from the drive link as there are a lot of small files in the dataset.

Once done, you should change <path-to-data> in the config files to the correct data directory.

Paper:

Byravan, Arunkumar, et al. "SE3-Pose-Nets: Structured Deep Dynamics Models for Visuomotor Planning and Control.", ICRA 2018.