Pytorch Vector Quantization

A pytorch library for vector quantization methods. Vector quantization has been successfully used by high-quality image and audio generation, e.g., VQVAE, VQGAN.

Implemented methods:

• Vector Quantization
• Vector Quantization based on momentum moving average
• Vector Quantization based on gumbel-softmax trick
• Product Quantization
• Residual Quantization

Usage

import torch
from vector_quantize import VectorQuantizer

vq = VectorQuantizer(
    n_e = 1024,          # codebook vocalbulary size
    e_dim = 256,         # codebook vocalbulary dimension
    beta = 1.0,          # the weight on the commitment loss
)

x = torch.randn(1, 256, 16, 16)          # size of NCHW
quantized, commit_loss, indices = vq(x)          # shape of (1, 256, 16, 16), (1), (1, 16, 16)