Residual Dense Generative Adversarial Network for Pansharpening with Geometrical Constraints, IEEE ICIP 2020

Implementation of Residual Dense Generative Adversarial Network for Pansharpening with Geometrical Constraints

Anaïs GASTINEAU (1,2), Jean-François AUJOL (1), Yannick BERTHOUMIEU (2) and Christian GERMAIN (2)

(1) Univ. Bordeaux, Bordeaux INP, CNRS, IMB, UMR 5251, F-33400 Talence, France

(2) Univ. Bordeaux, Bordeaux INP, CNRS, IMS, UMR 5218, F-33400 Talence, France

contact : [email protected]

paper : https://hal.archives-ouvertes.fr/hal-02859866/document

To run the code with GPU: tensorflow-gpu=1.2.0,

cuda=8

cuDNN=5.1

python 3.6

First step: Use the file tfrecord.py to create a file with the .tfrecords extension

Second step: Use file rdgan.py to train and test the network

Usage for training: python rdgan.py --mode=train --output_dir=path_output_train_folder

--mode and --output_dir options are required. Other options are optionals. If not indicate, default values will be used. Other options are:

--batch_size : number of images in the batch

--beta1 : weight for ADAM

--checkpoint : path to the checkpoint

--display_freq : frequency for saving images during training

--gan_weight : weight for cross entropy term in the loss function of the generator

--l1_weight : weight for l1 term in the loss function of the generator

--lr : learing rate initial pour ADAM

--max_epochs : maximal number of epochs

--max_steps : maximal number of itérations

--mode : train or test

--ndf : number of filters in the first layer of the generator

--output_dir : path for the output directory. The directory will be created if it doesn't exist.

--progress_freq : frequency to display the progression in the terminal

--save_freq : frequency of saving the model

--test_count : number of test images

--test_tfrecord : path for the .tfrecords file obtained with test images

--train_count : number of train images

--train_tfrecord : path for the .tfrecords file obtained with train images

Usage for testing: python residual_dense_pxs.py --mode=test --checkpoint=path_folder_with_saved_model --output_dir=path_output_test_folder