- The dockerfile creates a conda environment with all dependancies required. Make sure to activate the environment with
conda activate dope
before you run test scripts that uses DOPE.
- The model expects the weights to be stored in /root/git/dope/weights directory inside the docker. The way to use it is, first download the weights from here to docker/scratchpad directory and inside the docker run,
cd /root/git/scratchpad
cp <weights file> /root/git/dope/weights
- Make sure to edit your config files as needed, you can see example of config file at robot_toolkit/config/dope_config_pose.yaml. This file configured to run test_dope.py test script\
- Download the weights from here
- Copy the weights file to docker/scratchpad
Troubleshooting:
- If the inference time for contactgraspnet is very high(in minutes) or you notice that the predicted grasps are way off, this means that your tensorflow version could be off as reported here
- To solve this,
- First, consider getting the right cuda version for your GPU, if you think the version is not matching with what is installed in the docker (see line 1 in Dockerfile)
- You can edit above mentioned line 1 with the image name that has the right CUDA version here
- Upgrade tensorlow by running (make sure the conda environment for contact graspnet is active)
pip3 install tensorflow --upgrade - The opensourced contact graspnet model is compiled with cuDNN 8.1.1 so we need to install that as well, following instructions in here
- recompile pointnet model as mentioned here
- This shold solve the above menntioned problem
- Note: We tested contactgraspnet with tensorflow 2.11, CUDA 11.1.1, CUDNN 8.1.1
- Megapose models are already downloaded for you with a python command in the workstation/Dockerfile.
- You would need meshes of the objects whose pose we want to detect. For tests we use YCB objects soupcan, drill and cheezit
- These meshes should be stored in robot_toolkit/robot_arm_algos/src/inference/objects