Access to the data

For now, the data we used cannot be shared.

All the datas for the input contour and output knots to predict have been prepared from X-Ray scans and the methodology of Khazem et al.(2023) has been applied.

The data have been split into a training fold and a validation fold. These two folds are stored in directories TRAIN_PATH and VALID_PATH

Dependencies

You can install the required dependencies in a virtualenvironment using the provided requirements.txt file.

virtualenv -p python3 venv
source venv/bin/activate 
python -m pip install -r requirement.txt

Training a network

Training a ConvLSTM network

To train a ConvLSTM network on the data, you can use the train_sequence.py script :

python train_sequence.py --model_name ConvLSTM --train_path TRAIN_PATH --valid_path VALID_PATH

The assets of the runs are saved in the args.output_path/models and args.output_path/tensorboard directories.

Training a non recurrent UNet or SegNet

To train a UNet or SegNet network on the data, you can use the train_sequence.py script. The call is similar than for training a ConvLSTM :

For a UNet :

python train_sequence.py --model_name UNet --train_path TRAIN_PATH --valid_path VALID_PATH

For a SegNet :

python train_sequence.py --model_name SegNet --train_path TRAIN_PATH --valid_path VALID_PATH

The assets of the runs are saved in the args.output_path/models and args.output_path/tensorboard directories.

Running an inference

To run inference on trees, with the specific goal of computing metrics on the test sets, we use labeled data with input in INPUT_PATH and corresponding labels in MASK_PATH. For a given model with weights saved in WEIGHTS_CKPT, you should call :

python inference.py --input_path INPUT_PATH --mask_path MASK_PATH --weights WEIGHT_CKPT --model {UNet, SegNet, ConvLSTM} --descriptor descriptor.json

The inference script can be more verbose by providing --save_img

The optional species argument, which on our case is either Fir or Spruce, is used to tag the computed metrics by a tree specie.

The outputs of this inference are saved in the outputs directory (can be customized by --output provided to the script).

For inference, the descriptor.json file must contain the image width and height as well as the number of slices per volume, for example with the following content :

seq_size: 40
input_shape: [192, 192]