Skip to content

Latest commit

 

History

History
95 lines (63 loc) · 4 KB

USAGE.md

File metadata and controls

95 lines (63 loc) · 4 KB
Raw

License CC BY-NC-SA 4.0 Python 3.6

Visualizing Climate Change - MUNIT

License

This repo contains the code adapted from MUNIT for the needs of the VICC project. It keeps the same CC BY-NC-SA 4.0 license

Copyright (C) 2018 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).

Dependency

pytorch, yaml, comet_ml to visualize the results

The code base was developed using Anaconda with the following packages.

conda install pytorch=0.4.1 torchvision cuda90 -c pytorch
conda install -y -c anaconda pip
conda install -y -c anaconda pyyaml
pip install -r ./requirements.txt

Testing the demo

First, download the pretrained models and put them in models folder.

Run the following command to translate a folder of non flooded image to it's flooded version:

   python test.py --config ../configs/config.yaml --checkpoint models/gen_00370000.pt --input ../input_folder/ --output_folder output_folder/ --style ../Style_Image/style_image.jpg

It is possible to control the style of output using an example style_image. The results are stored in outputs folder.

For Developers

Training

  1. Generate the dataset you want to use. The data we use is not currently available for public re-use. Please contact us if you want more informations.

  2. Setup the yaml file. Check out configs/config_HD.yaml. Currently we only support list-based dataset organization, check out the parameter: data_list. In the yaml config file, for each domain and for train and test, specify the path to the folder containing the images and a txt file listing the images in the folder to be considered for training/testing. E.g.
    Example trainA.txt.

    ./images_A/dyzWTle6XGPRvdZiFqwMLQ.jpg
./images_A/sGXE_tiBOjqtidWK_VgUOA.jpg
./images_A/gxv9RvgHZi_XF11EbJ4Opw.jpg
./images_A/QV_vTfZC2JSs69XzsbmqXA.jpg

  3. Start training: in 256x256

    python train.py --config configs/config_256.yaml
Continue training

Use --resume with the same output_path AND for now, for some reason, you need to use the same name for the config file.

How to use comet_ml ?

Create a file at MUNIT/scripts/.comet.config (which is ignored by git):

[comet]
api_key=YOUR-API-KEY
workspace=YOUR-WORKSPACE
project_name=THE-PROJECT

For more information, see docs

Experiments run

https://docs.google.com/spreadsheets/d/1Csdi2B-LJPChLwO1ng4i2sjPgQxrNRlempa05o3a7og/edit?usp=sharing

How to reproduce an experiment ?

  • Open the link to the comet-ML experiment
  • Go to Hyper Parameters to find the git_hash
  • Go to assets and download the config.yaml
  • Roll back to github version corresponding to the git_hash useful docs

If one wants to use the checkpoints, they can be found in the output path written in hyperparameters. For some reasons they might have been moved to a commonly used folder: ccai/checkpoints/archive_XX

Branches

  • Master is the main branch, code that is meant to be deployed on the https://climatechangeai.org
  • feature/cocoStuff_merged_logits is a branch where we merged several classes of cocostuff so that we have a semantic segmentation consistency that is able to detect water properly. See utils.py and the assignment_dir function.
  • feature/highres is an attempt of using pix2pix upsampling in our settings with some major modifications.
  • feature/SpadeResBlock replaces the ResBlocks with SpadeResBlocks. The AdaIN coefficients are not computed from a style image but from the semantic segmentation of the image to be translated.