license: mit
pipeline_tag: audio-to-audio
SimVQ: Addressing Representation Collapse in Vector Quantized Models with One Linear Layer
Introduction
Vector Quantization (VQ) is essential for discretizing continuous representations in unsupervised learning but suffers from representation collapse, causing low codebook utilization and limiting scalability. Existing solutions often rely on complex optimizations or reduce latent dimensionality, which compromises model capacity and fails to fully solve the problem.
This paper introduces SimVQ, a novel approach that reparameterizes code vectors through a learnable linear transformation layer over a latent basis, optimizing the entire linear space rather than nearest individual code vectors. Although the multiplication of two linear matrices is equivalent to applying a single linear layer, this simple approach effectively prevents collapse. Extensive experiments on image and audio tasks demonstrate that SimVQ improves codebook usage, is easy to implement, and generalizes well across modalities and architectures.
For more details on the method and implementation, please refer to the official resources:
- Paper: Addressing Representation Collapse in Vector Quantized Models with One Linear Layer
- Code: https://github.com/youngsheen/SimVQ
Algorithm for SimVQ
The core idea of SimVQ is to reparameterize code vectors through a learnable linear transformation layer over a latent basis. You can find the core code here: https://github.com/youngsheen/SimVQ/blob/main/taming/modules/vqvae/quantize.py#L28-L33
Quantitative Comparison
Table 1. Reconstruction performance of different tokenizers on $128 \times 128$ ImageNet 50k validation set.
| Method | Codebook Size | Codebook Utilization | rFID | LPIPS | PSNR | SSIM | Checkpoint |
|---|---|---|---|---|---|---|---|
| VQGAN | 65,536 | 1.4% | 3.74 | 0.17 | 22.20 | 70.6 | - |
| VQGAN | 65,536 | 4.5% | 3.23 | 0.15 | 22.89 | 72.3 | - |
| VQGAN-FC | 65,536 | 100.0% | 2.63 | 0.13 | 23.79 | 77.5 | - |
| FSQ | 64,000 | 100.0% | 2.80 | 0.13 | 23.63 | 75.8 | - |
| LFQ | 65,536 | 100.0% | 2.88 | 0.13 | 23.60 | 77.2 | - |
| VQGAN-LC | 65,536 | 100.0% | 2.40 | 0.13 | 23.98 | 77.3 | - |
| SimVQ (ours) | 1024 | 100.0% | 3.67 | 0.16 | 22.34 | 70.8 | huggingface |
| SimVQ (ours) | 8192 | 100.0% | 2.98 | 0.14 | 23.23 | 74.7 | huggingface |
| SimVQ (ours) | 65,536 | 100.0% | 2.24 | 0.12 | 24.15 | 78.4 | huggingface |
| SimVQ (ours) | 262,144 | 100.0% | 1.99 | 0.11 | 24.68 | 80.3 | huggingface |
Table 2. Reconstruction performance of different tokenizers on LibriTTS test clean/other set.
| Method | Bandwidth | Codebook Utilization | UTMOS | PESQ | STOI | V/UV F1 | Checkpoint |
|---|---|---|---|---|---|---|---|
| Encodec | 3.0kbps | -/-% | 2.31/2.09 | 2.05/2.05 | 0.90/0.88 | 0.92/0.89 | - |
| Vocos | 3.0kbps | -/-% | 3.53/3.06 | 2.40/2.19 | 0.92/0.90 | 0.94/0.91 | - |
| SpeechTokenizer | 3.0kbps | -/-% | 3.56/3.02 | 1.93/1.74 | 0.88/0.84 | 0.93/0.89 | - |
| WavTokenizer | 0.9kbps | 100/100% | 3.74/3.43 | 2.01/2.26 | 0.89/0.89 | 0.92/0.92 | - |
| WavTokenizer | 1.05kbps | 27/-% | 4.00/- | 2.36/- | 0.81/- | 0.94/- | - |
| SimVQ (ours) | 0.9kbps | 100.0/100.0% | 4.00/3.51 | 2.33/2.08 | 0.91/0.88 | 0.94/0.91 | huggingface |
| SimVQ (ours) | 0.975kbps | 99.4/99.4% | 4.03/3.52 | 2.42/2.15 | 0.92/0.88 | 0.94/0.92 | huggingface |
| SimVQ (ours) | 1.2kbps | 99.4/99.0% | 4.03/3.52 | 2.54/2.26 | 0.93/0.90 | 0.94/0.92 | huggingface |
| SimVQ (ours) | 1.35kbps | 95.6/94.7% | 4.03/3.53 | 2.61/2.31 | 0.93/0.90 | 0.95/0.93 | huggingface |
Implementations
For detailed instructions on installation, training, and evaluation scripts, please refer to the GitHub repository.
Installation
- Dependencies:
pip install -r requirements.txt - Extra dependencies for audio evaluation:
pip install -r requirements_audio.txt
Training Scripts
Example scripts are provided for:
- Image Tokenizer Training (
configs/imagenet_simvq_128_B.yaml) - Audio Tokenizer Training (
configs/libritts_24khz.yaml)
Note: Some users have reported encountering NaN issues when training SimVQ on audio data. This appears to be a random occurrence, but we have found that using learning rate warmup can help mitigate the problem.
Evaluation Scripts
Example scripts are provided for:
- Image Tokenizer Evaluation
- Audio Tokenizer Evaluation
Reconstruction Visualization
Figure 2. Visualization of the Open-MAGVIT2 tokenizer trained at $128 \times 128$ resolution (imagenet_simvq_128_Base version). (a) indicates the original images while (b) specifies the reconstruction images.
Figure 3. Visualization of the Open-MAGVIT2 tokenizer trained at LibriTTS (libritts_24khz version). (a) indicates the original audio spectrograms while (b) specifies the reconstruction spectrograms.
Acknowledgement
The codebase of SimVQ is adapted from Open-MAGVIT2 and WavTokenizer. Thanks for their wonderful work.
Citation
If you find our work helpful or inspiring, please feel free to cite it.
@misc{zhu2024addressing,
title={Addressing Representation Collapse in Vector Quantized Models with One Linear Layer},
author={Yongxin Zhu and Bocheng Li and Hang Zhang and Xin Li and Linli Xu and Lidong Bing},
year={2024},
eprint={2411.02038},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2411.02038},
}