1. Make sure you have a working version of nvidia-docker or docker installed on your computer.
2. Clone this gqcnn repository to a directory of your choice: $WORKING_DIR/gqcnn
3. Create a directory to store your results in, e.g. $RESULTS.
4. Download the trained model from the SharePoint and extract the zip file in $WORKING_DIR/gqcnn/..
5. Run ./build.sh from $WORKING_DIR/gqcnn/. Choose 1 if you have nvidia-docker and would like to use the GPU, 0 otherwise.
6. Set the variables DATA_PATH and EXPER_PATH in gqcnn/run_docker.sh to your $DATA_DIR and $RESULTS directories. $DATA_DIR should be pointing to the directory where your DexNet datasets are stored.
7. Run ./run_docker.sh from $WORKING_DIR/gqcnn/.

In order to evaluate a dataset, run

python3 tools/detailed_analysis.py GQCNN-2.0_benchmark $DATASET_DIR --output_dir $DATASET_OUTPUT_DIR

For example, if you want to evaluate a recreated PerfectPredictions dataset, run:

python3 tools/detailed_analysis.py GQCNN-2.0_benchmark Recreated_grasps/tensors/ --output_dir Recreated_grasps

The logfile and the images with prediction values are stored in $RESULTS/$DATASET_OUTPUT_DIR/.

The gqcnn Python package is for training and analysis of Grasp Quality Convolutional Neural Networks (GQ-CNNs). It is part of the ongoing Dexterity-Network (Dex-Net) project created and maintained by the AUTOLAB at UC Berkeley.

Please see the docs for installation and usage instructions.

If you use any part of this code in a publication, please cite the appropriate Dex-Net publication.