(QoL improvements for) DLAS - A configuration-driven trainer for generative models

Go to file

James Betker 57d6f6d366 Big rework of flat_diffusion Back to the drawing board, boys. Time to waste some resources catching bugs....		2022-05-22 08:09:33 -06:00
.idea	reformat x_transformers	2022-04-07 23:08:03 -07:00
codes	Big rework of flat_diffusion	2022-05-22 08:09:33 -06:00
recipes	Add tacotron2 recipe	2021-07-20 08:37:57 -06:00
.flake8	mmsr	2019-08-23 21:42:47 +08:00
.gitignore	Add script for computing attention for gpt_asr	2021-11-07 18:42:06 -07:00
.gitmodules	fixes	2022-03-17 10:53:39 -06:00
LICENSE	Initial commit	2019-08-23 21:04:30 +08:00
README.md	Add distributed training guide to docs	2021-06-06 16:56:40 -06:00
sandbox.py	Misc	2020-08-12 08:46:15 -06:00

README.md

Deep Learning Art School

Send your Pytorch model to art class!

This repository is both a framework and a set of tools for training deep neural networks that create images. It started as a branch of the open-mmlab project developed by Multimedia Laboratory, CUHK but has been almost completely re-written at every level.

Why do we need another training framework

These are a dime a dozen, no doubt. DL Art School (DLAS) differentiates itself by being configuration driven. You write the model code (specifically, a torch.nn.Module) and (possibly) some losses, then you cobble together a config file written in yaml that tells DLAS how to train it. Swapping model architectures and tuning hyper-parameters is simple and often requires no changes to actual code. You also don't need to remember complex command line incantations. This effectively enables you to run multiple concurrent experiments that use the same codebase, as well as retain backwards compatibility for past experiments.

Training effective generators often means juggling multiple loss functions. As a result, DLAS' configuration language is specifically designed to make it easy to support large number of losses and networks that interact with each other. As an example: some GANs I have trained in this framework consist of more than 15 losses and use 2 separate discriminators and require no bespoke code.

Generators are also notorious GPU memory hogs. I have spent substantial time streamlining the training framework to support gradient checkpointing and FP16. DLAS also supports "mega batching", where multiple forward passes contribute to a single backward pass. Most models can be trained on midrange GPUs with 8-11GB of memory.

The final value-added feature is interpretability. Tensorboard logging operates out of the box with no custom code. Intermediate images from within the training pipeline can be intermittently surfaced as normal PNG files so you can see what your network is up to. Validation passes are also cached as images so you can view how your network improves over time.

Modeling Capabilities

DLAS was built with extensibility in mind. One of the reasons I'm putting in the effort to better document this code is the incredible ease with which I have been able to train entirely new model types with no changes to the core training code.

I intend to fill out the sections below with sample configurations which can be used to train different architectures. You will need to bring your own data.

Super-resolution

GAN-based SR (ESRGAN)
SRFlow
GLEAN
Video SR (TecoGAN) (documentation TBC)

Style Transfer

Stylegan2 (documentation TBC)

Latent development

BYOL
iGPT (documentation TBC)

Dependencies and Installation

Python 3
PyTorch >= 1.6
NVIDIA GPU + CUDA
Python packages: pip install -r requirements.txt
Some video utilities require FFMPEG

User Guide

TBC

Development Environment

If you aren't already using Pycharm - now is the time to try it out. This project was built in Pycharm and comes with an IDEA project for you to get started with. I've done all of my development on this repo in this IDE and lean heavily on its incredible debugger. It's free. Try it out. You won't be sorry.

Dataset Preparation

DLAS comes with some Dataset instances that I have created for my own use. Unless you want to use one of the recipes above, you'll need to provide your own. Here is how to add your own Dataset:

Create a Dataset in codes/data/ which takes a single Python dict as a constructor and extracts options from that dict.
Register your Dataset in codes/data/init.py
Your Dataset should return a dict of tensors. The keys of the dict are injected directly into the training state, which you can interact within your configuration file.

Training and Testing

There are currently 3 base scripts for interacting with models. They all take a single parameter, -opt which specifies the configuration file which controls how they work. Configs (will be) documented above in the user guide.

train.py

Start (or continue) a training session: python train.py -opt <your_config.yml>

Start a distributed training session: python -m torch.distributed.launch --nproc_per_node=<gpus> --master_port=1234 train.py -o <opt> --launcher=pytorch

test.py

Runs a model against a validation or test set of data and reports metrics (for now, just PSNR and a custom perceptual metric) python test.py -opt <your_config.yml>

process_video.py

Breaks a video into individual frames and uses a network to do processing on it, then reassembles the output back into video form. python process_video -opt <your_config.yml>

Contributing

At this time I am not taking feature requests or bug reports, but I appreciate all contributions.

License

This project is released under the Apache 2.0 license.