brajard/inpainting

Python

Inpainting

In construction: some parts of the README and the deposit are not up to date.

Prerequisites

Python modules to be installed. Advice : use conda env

conda create -n nn python=3.5
source activate nn
pip install netCDF4 h5py numpy scipy matplotlib xarray tensorflow keras scikit-learn scikit-gstat

Regarding tensorflow, you can have a more specific installation looking at the website: https://www.tensorflow.org/install/

You have to set keras config file (generally at ~/.keras/keras.json)

{
    "image_data_format": "channels_last", 
    "epsilon": 1e-07, 
    "floatx": "float32", 
    "backend": "tensorflow"
}

Download data

download data in ./data folder

for the complete set of data: link to base-chl.tgz
for a small set (for quick testing): link to base-chl-small.tgz

Uncompress the data with tar -xvzf base-chl.tgz or tar -xvzf base-chl-small.tgz

Create training set

In file make-trainset.py:

set the basename to medchl.nc or medchl-small.nc
set the trainingname denoted [TNAME]
check the options (e.g. king of masking)

run the file in a terminal: ./make-trainset.py
Check some randomly picked images in ./figures/examples/[TNAME] Note : you can always have a look at all the .nc files using the software ncview

Train a first network

Network based on paper: Pathak, D., Krähenbühl, P., Donahue, J., Darrell, T., & Efros, A. A. (n.d.). Context Encoders: Feature Learning by Inpainting. Retrieved from http://people.eecs.berkeley.edu/~pathak/papers/cvpr16.pdf

In file Inpainting.py:

set the trainingname to [TNAME]
set the name to the name the neural network has to be saved (denoted [NETNAME])

run the file in a terminal: ./Inpainting.py

Test the model

In file test_model.py:

set the name to [NETNAME]

run the file in a terminal : ./test_model.py
Check some randomly picked images in ./figures/exemples/[NETNAME]