Sovereign 8s.is mirror of standardgalactic/WoodVIT-V1-training
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-01-29 17:50:42 +01:00
Checkpoint_examples Initial upload 2026-01-29 15:44:00 +01:00
torchaskivit Upload Modul 2026-01-29 15:53:25 +01:00
analysis_ASKIVIT_V1.5.py Initial upload 2026-01-29 15:44:00 +01:00
askivit_V1.5_inference.py Initial upload 2026-01-29 15:44:00 +01:00
LICENSE Initial commit 2026-01-29 15:42:48 +01:00
README.md Update README.md 2026-01-29 17:50:42 +01:00
V1_5_train.py Initial upload 2026-01-29 15:44:00 +01:00
V1_5_train_cfg.py Initial upload 2026-01-29 15:44:00 +01:00
WoodVIT_train.yaml Initial upload 2026-01-29 15:44:00 +01:00

ASKIVIT-Project-WoodVIT-V1-training

Training and inference scripts for CNN-based models on the WoodVIT V1 dataset.

This repository focuses on:

  • computing dataset statistics (channel means) for normalization,
  • training models (GPU),
  • evaluating trained checkpoints,
  • running inference on the test split.

Repository structure (overview)

  • Checkpoint_examples/
    Example checkpoints: one CNN checkpoint per modality.

  • V1_5_train.py
    Main training script.

  • V1_5_train_cfg.py
    Training configuration (paths, hyperparameters, modality selection, etc.).

  • WoodVIT_train.yaml
    Conda environment file (requirements).

  • analysis_ASKIVIT_V1.5.py
    Computes channel-wise statistics (mean per channel) for the multimodal data.
    Required before training.

  • askivit_V1.5_inference.py
    Runs inference on the test data using a trained checkpoint.

  • torchaskivit/
    PyTorch utilities: models, datasets/dataloaders, evaluation, and plotting helpers.


Setup

1) Create the environment

Using the provided environment file:

conda env create -f WoodVIT_train.yaml
conda activate askivit_env

Data

Download the WoodVIT datasets:


Run

Step 1 - Compute channel means (required)

  1. Set dataset paths in analysis_ASKIVIT_V1.5.py
  2. Run:
python analysis_ASKIVIT_V1.5.py

This script computes the per-channel mean values used for normalization during training.

Step 2 - Train (GPU)

  1. Set paths and training parameters in V1_5_train_cfg.py
  2. Run:
python V1_5_train.py

Training checkpoints and logs will be written to the output directory configured in V1_5_train_cfg.py.

Step 4 - Inference (test split)

The output is a .json file, which can be used later for further evaluation.

  1. Set the paths (dataset + checkpoint) in askivit_V1.5_inference.py
  2. Run:
python askivit_V1.5_inference.py

Step 3 - Evaluate

Further evaluation scripts are located in the torchaskivit/evalutils module.
The main input is the .json file generated in Step 4 (Inference).


Checkpoints

  • Checkpoint_examples/ contains example checkpoints (one per modality).

Citing

If you use this code or the WoodVIT datasets in academic work, please cite the paper and the dataset.

WoodVIT paper

BibTeX:

IN REVIEW

Dataset

WoodVIT V1 patch:

@misc{WoodVIT_V1,
    author = {Manuel Bihler and Lukas Roming and Dovilė Čibiraitė-Lukenskienė and Jochen Aderhold and Andreas Keil and Friedrich Schlüter and Robin Gruna and Michael Heizmann},
    doi = {10.35097/aj4ve1c03pkan0dr},
    howpublished = {RADAR4KIT},
    publisher = {Karlsruhe Institute of Technology},
    title = {WoodVIT V1},
    url = {https://doi.org/10.35097/aj4ve1c03pkan0dr},
    year = {2023}
}

Further information about the acquisition, registration, and processing can be found in the paper cited above.


License

This repository is published under the CC0-1.0 license. See LICENSE.