This repository was initially forked from MEDIAR.

Experiments

Our experiments can be found in the notebook MEDIAR Experiments.ipynb. For reproducibility, or to use the fine-tuned weights resulting from the experiments, please see the last section "Reproducibility of best results on a test set" in the notebook. A quick demonstration of the (fine-tuned) model is shown in Fine_tuned_MEDIAR_Tutorial.ipynb.

Model checkpoints

The fine-tuned weights are downloadable from google drive.

What can be found in this repository?

In this repository, we also include some configurations files in the config folder. The image_processing.ipynb notebook contains the pre-processing methods for the YeaZ dataset. However, note that the data is not yet publicly available. Finally, the train_tools/data_utils/utils.py file was modified to obtain correct image-label mapping for the YeaZ data; the original file can be found in the original MEDIAR repo.

Reproducibility

All requirements are the same as in the original publication (see requirements.txt). However, we train our model with a NVIDIA A100-SXM4-40GB GPU which requires an extra installation (see below).

pip install -r requirements.txt
pip install segmentation-models-pytorch==0.3.1
pip install wandb
# Run the next line if training on google colab A100 GPUs
pip install torch==1.11.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113