StarGAN2 for practice

Overview

StarGAN2 for practice

This version of StarGAN2 (coined as 'Post-modern Style Transfer') is intended mostly for fellow artists, who rarely look at scientific metrics, but rather need a working creative tool. At least, this is what I use nearly daily myself.
Here are few pieces, made with it: Terminal Blink, Occurro, etc.
Tested on Pytorch 1.4-1.8. Sequence-to-video conversions require FFMPEG. For more explicit details refer to the original implementation.

Features

  • streamlined workflow, focused on practical tasks [TBA]
  • cleaned up and simplified code for better readability
  • stricter memory management to fit bigger batches on consumer GPUs
  • models mixing (SWA) for better stability

NB: In the meantime here's only training code and some basic inference (processing). More various methods & use cases may be added later.

Presumed file structure

stargan2 root
├  _in input data for processing
├  _out generation output (sequences & videos)
├  data datasets for training
│  └  afhq [example] some dataset
│     ├  cats [example] images for training
│     │  └  test [example] images for validation
│     ├  dogs [example] images for training
│     │  └  test [example] images for validation
│     └  ⋯
├  models trained models for inference/processing
│  └  afhq-256-5-100.pkl [example] trained model file
├  src source code
└  train training folders
   └  afhq.. [example] auto-created training folder

Training

  • Prepare your multi-domain dataset as shown above. Main directory should contain folders with images of different domains (e.g. cats, dogs, ..); every such folder must contain test subfolder with validation subset. Such structure allows easy data recombination for experiments. The images may be of any sizes (they'll be randomly cropped during training), but not smaller than img_size specified for training (default is 256).

  • Train StarGAN2 on the prepared dataset (e.g. afhq):

 python src/train.py --data_dir data/afhq --model_dir train/afhq --img_size 256 --batch 8

This will run training process, according to the settings in src/train.py (check and explore those!). Models are saved under train/afhq and named as dataset-size-domaincount-kimgs, e.g. afhq-256-5-100.ckpt (required for resuming).

  • Resume training on the same dataset from the iteration 50 (thousands), presuming there's corresponding complete 3-models set (with nets and optims) in train/afhq:
 python src/train.py --data_dir data/afhq --model_dir train/afhq --img_size 256 --batch 8 --resume 50
  • Make an averaged model (only for generation) from the directory of those, e.g. train/select:
 python src/swa.py -i train/select 

Few personal findings

  1. Batch size is crucial for this network! Official settings are batch=8 for size 256, if you have large GPU RAM. One can fit batch 3 or 4 on 11gb GPU; those results are interesting, but less impressive. Batches of 2 or 1 are for the brave only.. Size is better kept as 256; the network has auto-scaling layer count, but I didn't manage to get comparable results for size 512 with batches up to 7 (max for 32gb).
  2. Model weights may seriously oscillate during training, especially for small batches (typical for Cycle- or Star- GANs), so it's better to save models frequently (there may be jewels). The best selected models can be mixed together with swa.py script for better stability. By default, Generator network is saved every 1000 iterations, and the full set - every 5000 iterations. 100k iterations (few days on a single GPU) may be enough; 200-250k would give pretty nice overfit.
  3. Lambda coefficients lambda_ds (diversity), lambda_cyc (reconstruction) and lambda_sty (style) may be increased for smaller batches, especially if the goal is stylization, rather than photo-realistic transformation. The videos above, for instance, were made with these lambdas equal 3. The reference-based generation is nearly lost with such settings, but latent-based one can make nice art.
  4. The order of domains in the training set matters a lot! I usually put some photos first (as it will be the main source imagery), and the closest to photoreal as second; but other approaches may go well too (and your mileage may vary).
  5. I particularly love this network for its' failures. Even the flawed results (when the batches are small, the lambdas are wrong, etc.) are usually highly expressive and "inventive", just the kind of "AI own art", which is so spoken about. Experimenting with such aesthetics is a great fun.

Generation

  • Transform image test.jpg with AFHQ model (can be downloaded here):
python src/test.py --source test.jpg --model models/100000_nets_ema.ckpt

This will produce 3 images (one per trained domain in the model) in the _out directory.
If source is a directory, every image in it will be processed accordingly.

  • Generate output for the domain(s), referenced by number(s):
python src/test.py --source test.jpg --model models/100000_nets_ema.ckpt --ref 2
  • Generate output with reference image for domain 1 (ref filename must start with that number):
python src/test.py --source test.jpg --model models/100000_nets_ema.ckpt --ref 1-ref.jpg

To be continued..

Credits

StarGAN2
Copyright © 2020, NAVER Corp. All rights reserved.
Made available under Creative Commons BY-NC 4.0 license.
Original paper: https://arxiv.org/abs/1912.01865

Owner
vadim epstein
vadim epstein
CMP 414/765 course repository for Spring 2022 semester

CMP414/765: Artificial Intelligence Spring2021 This is the GitHub repository for course CMP 414/765: Artificial Intelligence taught at The City Univer

ch00226855 4 May 16, 2022
Implementation for Curriculum DeepSDF

Curriculum-DeepSDF This repository is an implementation for Curriculum DeepSDF. Full paper is available here. Preparation Please follow original setti

Haidong Zhu 69 Dec 29, 2022
MoCoGAN: Decomposing Motion and Content for Video Generation

MoCoGAN: Decomposing Motion and Content for Video Generation This repository contains an implementation and further details of MoCoGAN: Decomposing Mo

Sergey Tulyakov 514 Dec 18, 2022
Re-implementation of the vector capsule with dynamic routing

VectorCapsule Re-implementation of the vector capsule with dynamic routing We implement the vector capsule and dynamic routing via graph neural networ

ZhenchaoTang 10 Feb 10, 2022
VQMIVC - Vector Quantization and Mutual Information-Based Unsupervised Speech Representation Disentanglement for One-shot Voice Conversion

VQMIVC: Vector Quantization and Mutual Information-Based Unsupervised Speech Representation Disentanglement for One-shot Voice Conversion (Interspeech

Disong Wang 262 Dec 31, 2022
This Artificial Intelligence program can take a black and white/grayscale image and generate a realistic or plausible colorized version of the same picture.

Colorizer The point of this project is to write a program capable of taking a black and white / grayscale image, and generating a realistic or plausib

Maitri Shah 1 Jan 06, 2022
An example of semantic segmentation using tensorflow in eager execution.

Semantic segmentation using Tensorflow eager execution Requirement Python 2.7+ Tensorflow-gpu OpenCv H5py Scikit-learn Numpy Imgaug Train with eager e

Iñigo Alonso Ruiz 25 Sep 29, 2022
Worktory is a python library created with the single purpose of simplifying the inventory management of network automation scripts.

Worktory is a python library created with the single purpose of simplifying the inventory management of network automation scripts.

Renato Almeida de Oliveira 18 Aug 31, 2022
Solving Zero-Shot Learning in Named Entity Recognition with Common Sense Knowledge

Zero-Shot Learning in Named Entity Recognition with Common Sense Knowledge Associated code for the paper Zero-Shot Learning in Named Entity Recognitio

Søren Hougaard Mulvad 13 Dec 25, 2022
HuSpaCy: industrial-strength Hungarian natural language processing

HuSpaCy: Industrial-strength Hungarian NLP HuSpaCy is a spaCy model and a library providing industrial-strength Hungarian language processing faciliti

HuSpaCy 120 Dec 14, 2022
Reinforcement learning framework and algorithms implemented in PyTorch.

Reinforcement learning framework and algorithms implemented in PyTorch.

Robotic AI & Learning Lab Berkeley 2.1k Jan 04, 2023
Catalyst.Detection

Accelerated DL R&D PyTorch framework for Deep Learning research and development. It was developed with a focus on reproducibility, fast experimentatio

Catalyst-Team 12 Oct 25, 2021
Official PyTorch implementation for paper Context Matters: Graph-based Self-supervised Representation Learning for Medical Images

Context Matters: Graph-based Self-supervised Representation Learning for Medical Images Official PyTorch implementation for paper Context Matters: Gra

49 Nov 23, 2022
Process JSON files for neural recording sessions using Medtronic's BrainSense Percept PC neurostimulator

percept_processing This code processes JSON files for streamed neural data using Medtronic's Percept PC neurostimulator with BrainSense Technology for

Maria Olaru 3 Jun 06, 2022
Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding (CVPR2022)

Incremental Transformer Structure Enhanced Image Inpainting with Masking Positional Encoding by Qiaole Dong*, Chenjie Cao*, Yanwei Fu Paper and Supple

Qiaole Dong 190 Dec 27, 2022
PyTorch implementation of D2C: Diffuison-Decoding Models for Few-shot Conditional Generation.

D2C: Diffuison-Decoding Models for Few-shot Conditional Generation Project | Paper PyTorch implementation of D2C: Diffuison-Decoding Models for Few-sh

Jiaming Song 90 Dec 27, 2022
Medical Insurance Cost Prediction using Machine earning

Medical-Insurance-Cost-Prediction-using-Machine-learning - Here in this project, I will use regression analysis to predict medical insurance cost for people in different regions, and based on several

1 Dec 27, 2021
TensorFlow implementation of "A Simple Baseline for Bayesian Uncertainty in Deep Learning"

TensorFlow implementation of "A Simple Baseline for Bayesian Uncertainty in Deep Learning"

YeongHyeon Park 7 Aug 28, 2022
A new codebase for Group Activity Recognition. It contains codes for ICCV 2021 paper: Spatio-Temporal Dynamic Inference Network for Group Activity Recognition and some other methods.

Spatio-Temporal Dynamic Inference Network for Group Activity Recognition The source codes for ICCV2021 Paper: Spatio-Temporal Dynamic Inference Networ

40 Dec 12, 2022
Decorator for PyMC3

sampled Decorator for reusable models in PyMC3 Provides syntactic sugar for reusable models with PyMC3. This lets you separate creating a generative m

Colin 50 Oct 08, 2021