jkarolczak / image-inpainting

Image inpainting using GAN

Report

A report and live demo are available here

Configuration and usage

To create environment using Conda and yaml files:

conda env create -f environment.yaml

To activate created environment:

conda activate image-inpainting

To parametrize network or training:

Edit yaml adequate file in the /cfg directory.

To integrate training with neptune:
For more details see: https://docs.neptune.ai/getting-started/hello-world

Create /src/neptune.yaml file containing project name and token using temaplate in the file /src/neptune_template.yaml

To run the training process:

python train.py

To run the training process in debugging mode:
Debugging stops logging to neptune and display intermediate results to standard output.

python train.py --debug

To automatically reproduce the entire training with a current selection of parameters in cfg folder:

dvc repro

This will automatically run the following DVC stages:

• generate_data - generation of partial dataset with masked areas
• train_model - trains GAN-based architechture

To perform inference using generator run:

python infer.py --statedict path_to_statedict --images 10 Where --path_to_statedict stands for file to a pickled generators state dict and --images stands for number of images to use. Specifying number of images may be omitted.