gpd-revised

Introductions

GPD-revised is a repository that originated from the gpd [1] repository by atenpas. We make some modifications to the gpd network in order to adapt the new dataset generalized by Haoshu Fang and Chenxi Wang.

Files

gpd-revised
├── pytorch
|   ├── train_generator.py
|   ├── test_generator.py
|   ├── network.py
|   ├── json_dataset.py
|   ├── train_net.py
|   └── continue_train_net.py
└── README.md    

The pytorch folder is the code after modifications that can adapt the new dataset. Here are some specific explanations.

• train_generator.py and test_generator.py are data generator program, including the selection of the dataset, that is, balance the positive and negative samples;
• json_dataset.py contains the dataset manager, which are vital to our codes. Notice that we have changed the data manager file from h5 to json in order to satisfy the needs of the new dataset;
• network.py is the main structure of the network, and some vital modifications are made in order to suit the dataset well;
• train_net.py and continue_train_net.py are the trainning codes, and the latter one supports reloading the current model and continuing the training process. Both of the codes are modified from the original code train_net3.py.

Requirements

In order to use our codes to train the model, you need the Minkowski Engine dataset generalized by Haoshu Fang and Chenxi Wang, and the file structure of the dataset should be as follows.

gpd_data
├── train1
|   ├── image
|   |   ├── 000000.jpg
|   |   ├── 000001.jpg
|   |   └── ...
|   └── labels_train1.npy
├── train2
|   └── ...
├── train3
|   └── ...
├── train4
|   └── ...
└── test_seen
    └── ...

You may need to do some simple modifications to the train_generator.py and test_generator.py to satisfy your own path requirements.

Other requirements include pytorch framework and some other dependencies, you can refer to the codes for details.

Usages

To use our codes to train the net, you may follow the steps listed here.

• Make some modifications to the train_generator.py and test_generator.py to satisfy your own path requirements;
• Run train_generator.py and test_generator.py;
• Run train_net.py, and if you want to reload the old model and continue the training progress, you can run continue_train_net.py instead of the previous one.
  • For train_net.py, you need to add two more arguments corresponding to the path to the training set and the testing set (the data should be processed by the train_generator.py and test_generator.py);
  • For continue_train_net.py, besides the two more arguments of the path of the datasets, you need to add one more argument in the end, corresponds to the path to the old model file *.pwf.

References

[1] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt. Grasp Pose Detection in Point Clouds. The International Journal of Robotics Research, Vol 36, Issue 13-14, pp. 1455-1473. October 2017.