To train the model, you need to set some esstential parameters in the config file. Config files should be put in the ./config/ folder. In addition, the training dataset is a folder named 'DENSE', and please set parameter --datafolder in command to set the root folder of the dataset 'DENSE'. In the config file You alse may pay attention to the save_dir setting that saves all checkpoints of the model and detailed information during training.
CUDA_VISIBLE_DEVICES=0,1 python train_parallel.py --config configs/train_s2d_spiketransformer.json --datafolder ... --multiprocessing_distributedTesting a network is done by using two scripts. First, the test_DENSE.py script is used to save the outputs of the network.
As a second step, the evaluation_DENSE.py script is used to calculate the metrics based on these outputs.
To run the test_DENSE.py script, '--path_to_model' is the model path that you want to evaluate, '--output_path' is the folder that saves all the testing results including files with .npy and .png formats ( these files are used in the evaluation_DENSE.py script), and you alse need to set the '--data_folder' to set the root folder of the testing dataset.
For example:
CUDA_VISIBLE_DEVICES=0 python test_DENSE.py --path_to_model .../model_best.pth.tar --output_path ... --data_folder .../DENSE/test/To run the evaluation_DENSE.py script, three parameters is important. Specifically, '--target_dataset' has to be specified as .../ground_truth/npy/depth_image/, '--predictions_dataset' has to be specified as .../npy/image/, and '--output_folder' can be set to save the final visualization results.
For example:
CUDA_VISIBLE_DEVICES=0 python evaluation_DENSE.py --target_dataset .../ground_truth/npy/depth_image/ --predictions_dataset .../npy/image/ --clip_distance 1000 --reg_factor 5.7To help successfully run the code, the esstential environment on Linux is included in the requirements.txt.
The structure of the codebase is borrowed from RAMNet and the base of the encoder backbone is borrowed from Video Swin Transformer