Awesome

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 and Han Zhang},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2021}
}

Disclaimer

This is not an officially supported Google product.