Tensorflow implementation for "Improved Transformer for High-Resolution GANs" (NeurIPS 2021).

Related tags

Deep Learningtensorflowgenerative-adversarial-networkvision-transformer
Overview

HiT-GAN Official TensorFlow Implementation

HiT-GAN presents a Transformer-based generator that is trained based on Generative Adversarial Networks (GANs). It achieves state-of-the-art performance for high-resolution image synthesis. Please check our NeurIPS 2021 paper "Improved Transformer for High-Resolution GANs" for more details.

This implementation is based on TensorFlow 2.x. We use tf.keras layers for building the model and use tf.data for our input pipeline. The model is trained using a custom training loop with tf.distribute on multiple TPUs/GPUs.

Environment setup

It is recommended to run distributed training to train our model with TPUs and evaluate it with GPUs. The code is compatible with TensorFlow 2.x. See requirements.txt for all prerequisites, and you can also install them using the following command.

pip install -r requirements.txt

ImageNet

At the first time, download ImageNet following tensorflow_datasets instruction from the official guide.

Train on ImageNet

To pretrain the model on ImageNet with Cloud TPUs, first check out the Google Cloud TPU tutorial for basic information on how to use Google Cloud TPUs.

Once you have created virtual machine with Cloud TPUs, and pre-downloaded the ImageNet data for tensorflow_datasets, please set the following enviroment variables:

TPU_NAME=<tpu-name>
STORAGE_BUCKET=gs://<storage-bucket>
DATA_DIR=$STORAGE_BUCKET/<path-to-tensorflow-dataset>
MODEL_DIR=$STORAGE_BUCKET/<path-to-store-checkpoints>

The following command can be used to train a model on ImageNet (which reflects the default hyperparameters in our paper) on TPUv2 4x4:

python run.py --mode=train --dataset=imagenet2012 \
  --train_batch_size=256 --train_steps=1000000 \
  --image_crop_size=128 --image_crop_proportion=0.875 \
  --save_every_n_steps=2000 \
  --latent_dim=256 --generator_lr=0.0001 \
  --discriminator_lr=0.0001 --channel_multiplier=1 \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=True --master=$TPU_NAME

To train the model on ImageNet with multiple GPUs, try the following command:

python run.py --mode=train --dataset=imagenet2012 \
  --train_batch_size=256 --train_steps=1000000 \
  --image_crop_size=128 --image_crop_proportion=0.875 \
  --save_every_n_steps=2000 \
  --latent_dim=256 --generator_lr=0.0001 \
  --discriminator_lr=0.0001 --channel_multiplier=1 \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=False --use_ema_model=False

Please set train_batch_size according to the number of GPUs for training. Note that storing Exponential Moving Average (EMA) models is not supported with GPUs currently (--use_ema_model=False), so training with GPUs will lead to slight performance drop.

Evaluate on ImageNet

Run the following command to evaluate the model on GPUs:

python run.py --mode=eval --dataset=imagenet2012 \
  --eval_batch_size=128 --train_steps=1000000 \
  --image_crop_size=128 --image_crop_proportion=0.875 \
  --latent_dim=256 --channel_multiplier=1 \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=False --use_ema_model=True

This command runs models with 8 P100 GPUs. Please set eval_batch_size according to the number of GPUs for evaluation. Please also note that train_steps and use_ema_model should be set according to the values used for training.

CelebA-HQ

At the first time, download CelebA-HQ following tensorflow_datasets instruction from the official guide.

Train on CelebA-HQ

The following command can be used to train a model on CelebA-HQ (which reflects the default hyperparameters used for the resolution of 256 in our paper) on TPUv2 4x4:

python run.py --mode=train --dataset=celeb_a_hq/256 \
  --train_batch_size=256 --train_steps=250000 \
  --image_crop_size=256 --image_crop_proportion=1.0 \
  --save_every_n_steps=1000 \
  --latent_dim=512 --generator_lr=0.00005 \
  --discriminator_lr=0.00005 --channel_multiplier=2 \
  --use_consistency_regularization=True \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=True --master=$TPU_NAME

Evaluate on CelebA-HQ

Run the following command to evaluate the model on 8 P100 GPUs:

python run.py --mode=eval --dataset=celeb_a_hq/256 \
  --eval_batch_size=128 --train_steps=250000 \
  --image_crop_size=256 --image_crop_proportion=1.0 \
  --latent_dim=512 --channel_multiplier=2 \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=False --use_ema_model=True

FFHQ

At the first time, download the tfrecords of FFHQ from the official site and put them into $DATA_DIR.

Train on FFHQ

The following command can be used to train a model on FFHQ (which reflects the default hyperparameters used for the resolution of 256 in our paper) on TPUv2 4x4:

python run.py --mode=train --dataset=ffhq/256 \
  --train_batch_size=256 --train_steps=500000 \
  --image_crop_size=256 --image_crop_proportion=1.0 \
  --save_every_n_steps=1000 \
  --latent_dim=512 --generator_lr=0.00005 \
  --discriminator_lr=0.00005 --channel_multiplier=2 \
  --use_consistency_regularization=True \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=True --master=$TPU_NAME

Evaluate on FFHQ

Run the following command to evaluate the model on 8 P100 GPUs:

python run.py --mode=eval --dataset=ffhq/256 \
  --eval_batch_size=128 --train_steps=500000 \
  --image_crop_size=256 --image_crop_proportion=1.0 \
  --latent_dim=512 --channel_multiplier=2 \
  --data_dir=$DATA_DIR --model_dir=$MODEL_DIR \
  --use_tpu=False --use_ema_model=True

Cite

@inproceedings{zhao2021improved,
  title = {Improved Transformer for High-Resolution {GANs}},
  author = {Long Zhao and Zizhao Zhang and Ting Chen and Dimitris Metaxas abd Han Zhang},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2021}
}

Disclaimer

This is not an officially supported Google product.

Owner
Google Research
Google Research
GitHub Repository https://arxiv.org/abs/2106.07631
Towards Open-World Feature Extrapolation: An Inductive Graph Learning Approach

This repository holds the implementation for paper Towards Open-World Feature Extrapolation: An Inductive Graph Learning Approach Download our preproc

Qitian Wu 42 Dec 27, 2022
African language Speech Recognition - Speech-to-Text

Swahili-Speech-To-Text Table of Contents Swahili-Speech-To-Text Overview Scenario Approach Project Structure data: models: notebooks: scripts tests: l

2 Jan 05, 2023
Official Repo of my work for SREC Nandyal Machine Learning Bootcamp

About the Bootcamp A 3-day Machine Learning Bootcamp organised by Department of Electronics and Communication Engineering, Santhiram Engineering Colle

MS 1 Nov 29, 2021
Differential Privacy for Heterogeneous Federated Learning : Utility & Privacy tradeoffs

Differential Privacy for Heterogeneous Federated Learning : Utility & Privacy tradeoffs In this work, we propose an algorithm DP-SCAFFOLD(-warm), whic

19 Nov 10, 2022
TSDF++: A Multi-Object Formulation for Dynamic Object Tracking and Reconstruction

TSDF++: A Multi-Object Formulation for Dynamic Object Tracking and Reconstruction TSDF++ is a novel multi-object TSDF formulation that can encode mult

ETHZ ASL 130 Dec 29, 2022
Repository for the semantic WMI loss

Installation: pip install -e . Installing DL2: First clone DL2 in a separate directory and install it using the following commands: git clone https:/

Nick Hoernle 4 Sep 15, 2022
Spatial color quantization in Rust

rscolorq Rust port of Derrick Coetzee's scolorq, based on the 1998 paper "On spatial quantization of color images" by Jan Puzicha, Markus Held, Jens K

Collyn O'Kane 37 Dec 22, 2022
Official implementation of Self-supervised Image-to-text and Text-to-image Synthesis

Self-supervised Image-to-text and Text-to-image Synthesis This is the official implementation of Self-supervised Image-to-text and Text-to-image Synth

6 Jul 31, 2022
Code for our ALiBi method for transformer language models.

Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation This repository contains the code and models for our paper Tra

Ofir Press 211 Dec 31, 2022
The "breathing k-means" algorithm with datasets and example notebooks

The Breathing K-Means Algorithm (with examples) The Breathing K-Means is an approximation algorithm for the k-means problem that (on average) is bette

Bernd Fritzke 75 Nov 17, 2022
“Robust Lightweight Facial Expression Recognition Network with Label Distribution Training”, AAAI 2021.

EfficientFace Zengqun Zhao, Qingshan Liu, Feng Zhou. "Robust Lightweight Facial Expression Recognition Network with Label Distribution Training". AAAI

Zengqun Zhao 119 Jan 08, 2023
Simple, efficient and flexible vision toolbox for mxnet framework.

MXbox: Simple, efficient and flexible vision toolbox for mxnet framework. MXbox is a toolbox aiming to provide a general and simple interface for visi

Ligeng Zhu 31 Oct 19, 2019
VolumeGAN - 3D-aware Image Synthesis via Learning Structural and Textural Representations

VolumeGAN - 3D-aware Image Synthesis via Learning Structural and Textural Representations 3D-aware Image Synthesis via Learning Structural and Textura

GenForce: May Generative Force Be with You 116 Dec 26, 2022
Official repository for "PAIR: Planning and Iterative Refinement in Pre-trained Transformers for Long Text Generation"

pair-emnlp2020 Official repository for the paper: Xinyu Hua and Lu Wang: PAIR: Planning and Iterative Refinement in Pre-trained Transformers for Long

Xinyu Hua 31 Oct 13, 2022
Code for Multimodal Neural SLAM for Interactive Instruction Following

Code for Multimodal Neural SLAM for Interactive Instruction Following Code structure The code is adapted from E.T. and most training as well as data p

7 Dec 07, 2022
official implemntation for "Contrastive Learning with Stronger Augmentations"

CLSA CLSA is a self-supervised learning methods which focused on the pattern learning from strong augmentations. Copyright (C) 2020 Xiao Wang, Guo-Jun

Lab for MAchine Perception and LEarning (MAPLE) 47 Nov 29, 2022
Code for ACM MM2021 paper "Complementary Trilateral Decoder for Fast and Accurate Salient Object Detection"

CTDNet The PyTorch code for ACM MM2021 paper "Complementary Trilateral Decoder for Fast and Accurate Salient Object Detection" Requirements Python 3.6

CVTEAM 28 Oct 20, 2022
Deep learning for Engineers - Physics Informed Deep Learning

SciANN: Neural Networks for Scientific Computations SciANN is a Keras wrapper for scientific computations and physics-informed deep learning. New to S

SciANN 195 Jan 03, 2023
Generating Fractals on Starknet with Cairo

StarknetFractals Generating the mandelbrot set on Starknet Current Implementation generates 1 pixel of the fractal per call(). It takes a few minutes

Orland0x 10 Jul 16, 2022
Light-weight network, depth estimation, knowledge distillation, real-time depth estimation, auxiliary data.

light-weight-depth-estimation Boosting Light-Weight Depth Estimation Via Knowledge Distillation, https://arxiv.org/abs/2105.06143 Junjie Hu, Chenyou F

Junjie Hu 13 Dec 10, 2022