This project depends on
TreePartNet which should be
accessable in the PYTHONPATH when working with the data directory of
SorghumPartNet.
Note: A suitable version of CUDA-toolkit is needed on the host to build the CUDA extensions for the TreePartNet repository when it is being installed. However, only the dataset generation scripts depend TreePartNet currently, so it should be possible to run training and inference on existing preprocessed datasets without worrying about this dependency.
It is recommended to use a conda environment with python 3.7 or 3.9 for development and usage.
Currently, the directory for the inputs, and model checkpoints are hard-coded. This means that in each of the three scripts described below it will be necessary to modify the filepaths inline before attempting to run locally.
There are two training scripts; one for whole plant instance segmentation, and one for leaf semantic segmentation.
python3 ./train_and_inference/train_instance_segmentor.py
python3 ./train_and_inference/train_semantic_segmentor.py
Once the input and model checkpoint paths have been updated to be valid on your system it will be possible to run each command to train the respective models.
The inputs directory should contain three dataset files intended to be used for training.
The output will be in the form of model checkpoints in the specified model directory. A path to a particular checkpoint can be specified in the inference script.
Similar to the training scripts, to run inference locally it will be necessary
to modify the file paths which ar currently hard-coded in the file
train_and_inference/predict_and_visualize.py.
With the paths correctly specified, it will be possible to execute the command.
python3 ./train_and_inference/predict_and_visualize.py -i=23
Where the -i flag will specify to run inference on the file 23.ply in your
inputs directory.
The this will execute and produce 4 output files
23_instance_full.ply 23_instance.ply 23_semantic_full.ply 23_semantic.ply
Each result ".ply" files contains a 2d array where each row represents
(x,y,z,r,g,b) where points segmented together have the same r,g,b value.