GSQ-F8-D8-V16k

Scaling Image Tokenizers with Grouped Spherical Quantization

---

Paper link | GITHUB REPO HF Checkpoints

In GSQ, we show the optimized training hyper-parameters and configs for quantization based image tokenizer. We also show how to scale the latent, vocab size etc. appropriately to achieve better reconstruction performance.

We also show how to scaling the latent (and group) appropriately when pursuing high down-sample ratio in compression.

The group scaling experiment of GSQ:

---

---

Use Pre-trained GSQ-Tokenizer

from flex_gen import autoencoders
from timm import create_model

# ============= From HF's repo
model=create_model('flexTokenizer', pretrained=True,
                   repo_id='HelmholtzAI-FZJ/GSQ-F8-D8-V64k',)
									 
# ============= From Local Checkpoint
model=create_model('flexTokenizer', pretrained=True,
                   path='PATH/your_checkpoint.pt', )

---

Training your tokenizer

Set-up Python Virtual Environment

sh gen_env/setup.sh

source ./gen_env/activate.sh

#! This will run pip install to download all required lib
sh ./gen_env/install_requirements.sh 

Run Training

# Single GPU
python -W ignore ./scripts/train_autoencoder.py 

# Multi GPU
torchrun --nnodes=1 --nproc_per_node=4 ./scripts/train_autoencoder.py --config-file=PATH/config_name.yaml \
--output_dir=./logs_test/test opts train.num_train_steps=100 train_batch_size=16

Run Evaluation

Add the checkpoint path that your want to test in evaluation/run_tokenizer_eval.sh

# For example
...
configs_of_training_lists=()
configs_of_training_lists=("logs_test/test/")
...

And run sh evaluation/run_tokenizer_eval.sh it will automatically scan folder/model/eval_xxx.pth for tokenizer evaluation

---

Citation

@misc{GSQ,
      title={Scaling Image Tokenizers with Grouped Spherical Quantization}, 
      author={Jiangtao Wang and Zhen Qin and Yifan Zhang and Vincent Tao Hu and Björn Ommer and Rania Briq and Stefan Kesselheim},
      year={2024},
      eprint={2412.02632},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.02632}, 
}