Awesome

Inpainting Codebase

This repo implements a simple PyTorch codebase for training inpainting models with powerful tools including Docker, PyTorchLightning, and Hydra.

Currently, only DFNet is supported. More methods as well as some additional useful utilities for image inpainting will be implemented.

Prerequirements

We use docker to run all experiemnts.

• nvidia-docker
• docker-compose

Features

• PytorchLightning
  • logging (tensorboard, csv)
  • checkpoint
  • DistributedDataParallel
  • mixed-precision
• Hydra
  • flexible configuration system
  • logging (stream to file, folder structure)
• Others
  • save sample results

Setup

Build the environment

Build the image for the first time:

python core.py env prepare

Explaination:

• When you first run this command, you will be asked to give three items:
  1. a project name,
  2. the root folder of your train log,
  3. the root folder of your datasets,
• then an image is built based on /env/Dockerfile,
• and at last, a container is launched based on docker-compose.yml

The defualt setting of docker-compose.yml is shown as below, you can modify this setting before building accordingly:

version: "3.9"
services:
    lab:
        container_name: ${PROJECT}
        runtime: nvidia
        build:
            context: env/
            dockerfile: Dockerfile
            args:
                - USER_ID=${UID}
                - GROUP_ID=${GID}
                - USER_NAME=${USER_NAME}
        image: pytorch181_local
        environment:
            - TZ=Asia/Shanghai
            - TORCH_HOME=/data/torch_model
        ipc: host
        hostname: docker
        working_dir: /code
        command: ['sleep', 'infinity']
        volumes:
            - ${CODE_ROOT}:/code
            - ${DATA_ROOT}:/data
            - ${LOG_ROOT}:/outputs

Get into the environment

Simply run:

python core.py env

The default user is the same as the host to avoid permission issues. And of course you can enter the container with root:

python core.py env --root

Modify the environment at anytime

Basiclly, the environment are determined by four items:

1. /env/Dockerfile defines the logic of building the local docker image. For example, installing packages defined in requirements.txt based on deepbase/pytorch:latest.
2. Base docker image. From /env/Dockerfile, you can find deepbase/pytorch is the base image. The original Dockerfile of the base image is at deepcodebase/docker. You can change the base image as whatever you like.
3. /env/requirements.txt defines the python packages you want to install in the local docker image.
4. /docker-compose.yml defines the setting of running the container. For example, the volumes, timezone, etc.

After changing the settings as you want at anytime, you can rebuild the local image by running:

python core.py env prepare --build

Training

Data Prepartion

1. Image data: any image data you like. e.g. Places2, ImageNet, etc. Place your dataset into your DATAROOT in your local machine (mapped to docker's /data). For example: DATAROOT/places2 is used for training by default.
2. Masks: you can download and use free-form-mask. Decompress the file and place mask under DATAROOT.
3. By default, inside the environment, you need to have places2 and mask under /data.
4. If you use other datasets, remember to modify the settings especially the data location under conf/dataset.

Running

After entering the environment, you can launch training. Example training commands:

python train.py
python train.py mode=run pl_trainer.gpus=\'3,4\' logging.wandb.notes="tune model"
python train.py +experiment=k80 mode=run logging.wandb.tags='[k80]'

This project use wandb for logging by default, it will prompt if you run training the first time:

wandb: (1) Create a W&B account
wandb: (2) Use an existing W&B account
wandb: (3) Don't visualize my results
wandb: Enter your choice:

just follow the steps and wandb is convenient and easy use. If you wan't to use tensorboard instead, just add the flag when running:

python train.py logging=tensorboard

Reading Suggestions

Reading the offical documents of Hydra and PyTorchLightning to know more about:

• Hydra: Very powerful and convenient configuration system and more.
• PyTorchLightning: You almost only need to write codes for models and data. Say goodbye to massive codes for pipelines, mixed precision, logging, etc.

Results

Results of DFNet

Training on Places2 with 20 epochs.

License

<a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc/4.0/88x31.png" /></a><br />This work is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/">Creative Commons Attribution-NonCommercial 4.0 International License</a>.