All of these enviroments assume that the device has a NVIDIA GPU. These enviroments have not been tested on a cpu only devices but should be able to run and compile.
The preffered method is to use the Dev Container as this will ensure that the installed CUDA version matches what is specified in the packages. If you want to use the Dev Container via Visual Studio Code, here is a tutorial how to do so.
If you have Mamba or Conda enviroment managers run from the project source directory:
mamba/conda env create -f ./environment.yml
This will create a new enviroment called uda, run mamba/conda activate uda to activate it.
If you want to use Python PIP package manager, create a Python 3.11 enviroment and run from the project source directory:
pip install -r ./requirements.txt
To start training run train.py. All training arguments are CLI params and can be seen at the top of the script. Before running this script go through these params and adjust them according to your needs. Pay attention to -data_path and -test_data_path.
The data is split into two disjoint sets document wise for train and validation purposes. This means that no document appears in both sets.
All losses, metrics and visualisations are available through Tensorboard. Run tensorboard --logidr ./{-project_dir} to see the logged values.
If -run_inference is set to true during training inference will be ran after -inference_interval number of epochs, inference results are saved in storage under ./{-project_dir}/{-run_name}_{TIMESTAMP}. -test_data_path can contain unprocessed .pdf,.jpg (also .jpeg) and .png images.
The model used is UNet 3+ with ResBlocks.
The same can be done using the alternative Jupyter Notebook train.ipynb, by default it has parameters that require less computational power and all data paths are relative to the project directory.
To perform inference run inference.py.
Make sure to set up tile_size, model_path, input_dir and output_dir. By default these are relative to the project directory and should not require changes. The model_path is set to the best model according to evaluation IoU, tile_size is set to 512 (512x512), input_dir includes both subdirectories of data/test_data and output_dir is set to inference_results.
However, be warned that running inference with default arguments will overwrite the existing inference_results directory.
An alternative Jupyter Notebook inference.ipynb is available too that is identical to inference.py.