irisdum / cgan_sent2_sim

cGAN_sent2_sim

These software codes can be used for non-profit academic research only. They are distributed under the terms of the GNU general public license v3.

Code corresponding to I. Dumeur, Y. Chen, F. Sur, Z.-S. Zhou. Sentinel-2 RGB and NIR bands simulation after fire events using a multi-temporal conditional generative adversarial network. Research report HAL-03327421, 2021, https://hal.archives-ouvertes.fr/hal-03327421

Data used for this research are available in : https://drive.google.com/drive/folders/1PSob58o66WE5l-RDcImVcrc67C7HlDCR?usp=sharing

The environment :

Use the specific training conda environment : env/training_env.yaml :

conda env create -f env/training_env.yaml
conda activate training_env
python -m ipykernel install --user --name=training_env

Start a training

With Jupyter

Start jupyter notebook. If in a remote machine : add --ip=0.0.0.0 --no-browser

You can open the jupyter Notebook : notebooks/Trainings.ipynb

Modify the constant, defined at the beginning at the notebook and run the training.

As a batch job

The conditional GAN model used is defined in models/clean_gan.py In order to train a model two yaml should be modified, examples available in GAN_confs :

• model.yaml
• train.yaml Then run gan_train.sh path_to_model_yaml path_to_train_yaml

Supervise the training

Tensorboard : metric supervision

The training is also configured to be supervised using Tensorboard.

Open a new terminal window within training_env and run tensorboard.

Local command tensorboard --logdir <path>

If in remote machine : tensorboard --logdir <path> --host 0.0.0.0

Validation image visualization : notebook

Eventually you can also look at how the images look like during the training.

Open the notebook Figure_plot

Prediction

To run prediction of the model. You can run in python predict.py --model_path ${model_path} --tr_nber ${train_nber} --dataset ${dataset} --pref ${pref} --weights ${weight}

model_path : path where all the training directories of one model are contained tr_nber: number used to name the training directory dataset : path to the directory which contains the two subdirectories dataX and test pref : string used to name the directory where all the predicted tiles are saved weights: int, corresponds of the weights from the model and training number to use to run the prediction

You can also use the gan_predict.sh or look at the Makefile with make predict command.

Compute metrics

Open Notebooks Assessment_cGAN_CD