File size: 5,459 Bytes
36179f8 66ded8d 36179f8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 |
# **Scaling Image Tokenizers with Grouped Spherical Quantization**
---
[Paper link](https://arxiv.org/abs/2412.02632) | [GITHUB REPO](https://github.com/HelmholtzAI-FZJ/flex_gen) [HF Checkpoints](https://huggingface.co/collections/HelmholtzAI-FZJ/grouped-spherical-quantization-674d6f9f548e472d0eaf179e)
In [GSQ](https://arxiv.org/abs/2412.02632), 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:
---
| **Models** | \( G $\times$ d \) | **rFID ↓** | **IS ↑** | **LPIPS ↓** | **PSNR ↑** | **SSIM ↑** | **Usage ↑** | **PPL ↑** |
|--------------------------------------|---------------------|------------|----------|-------------|------------|------------|-------------|-------------|
| **GSQ F8-D64** \( V=8K \) | \( 1 $\times$ 64 \) | 0.63 | 205 | 0.08 | 22.95 | 0.67 | 99.87% | 8,055 |
| | \( 2 $\times$ 32 \) | 0.32 | 220 | 0.05 | 25.42 | 0.76 | 100% | 8,157 |
| | \( 4 $\times$ 16 \) | 0.18 | 226 | 0.03 | 28.02 | 0.08 | 100% | 8,143 |
| | \( 16 $\times$ 4 \) | **0.03** | **233** | **0.004** | **34.61** | **0.91** | **99.98%** | **6,775** |
| **GSQ F16-D16** \( V=256K \) | \( 1 $\times$ 16 \) | 1.63 | 179 | 0.13 | 20.70 | 0.56 | 100% | 254,044 |
| | \( 2 $\times$ 8 \) | 0.82 | 199 | 0.09 | 22.20 | 0.63 | 100% | 257,273 |
| | \( 4 $\times$ 4 \) | 0.74 | 202 | 0.08 | 22.75 | 0.63 | 62.46% | 43,767 |
| | \( 8 $\times$ 2 \) | 0.50 | 211 | 0.06 | 23.62 | 0.66 | 46.83% | 22,181 |
| | \( 16 $\times$ 1 \) | 0.52 | 210 | 0.06 | 23.54 | 0.66 | 50.81% | 181 |
| | \( 16 $\times$ 1^* \) | 0.51 | 210 | 0.06 | 23.52 | 0.66 | 52.64% | 748 |
| **GSQ F32-D32** \( V=256K \) | \( 1 $\times$ 32 \) | 6.84 | 95 | 0.24 | 17.83 | 0.40 | 100% | 245,715 |
| | \( 2 $\times$ 16 \) | 3.31 | 139 | 0.18 | 19.01 | 0.47 | 100% | 253,369 |
| | \( 4 $\times$ 8 \) | 1.77 | 173 | 0.13 | 20.60 | 0.53 | 100% | 253,199 |
| | \( 8 $\times$ 4 \) | 1.67 | 176 | 0.12 | 20.88 | 0.54 | 59% | 40,307 |
| | \( 16 $\times$ 2 \) | 1.13 | 190 | 0.10 | 21.73 | 0.57 | 46% | 30,302 |
| | \( 32 $\times$ 1 \) | 1.21 | 187 | 0.10 | 21.64 | 0.57 | 54% | 247 |
---
## Use Pre-trained GSQ-Tokenizer
```python
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
```python
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
```python
# 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`
```bash
# 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**
```bash
@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},
}
``` |