Jumping Through Local Minima: Quantization in the Jagged Loss Landscape of Vision Transformers (Evol-Q)
Official PyTorch implementation of the ICCV 2023 paper
Jumping Through Local Minima: Quantization in the Jagged Loss Landscape of Vision Transformers
Natalia Frumkin, Dibakar Gope, and Diana Marculescu
https://arxiv.org/pdf/2308.10814
Abstract: Quantization scale and bit-width are the most important parameters when considering how to quantize a neural network. Prior work focuses on optimizing quantization scales in a global manner through gradient methods (gradient descent & Hessian analysis). Yet, when applying perturbations to quantization scales, we observe a very jagged, highly non-smooth test loss landscape. In fact, small perturbations in quantization scale can greatly affect accuracy, yielding a 0.5−0.8% accuracy boost in 4-bit quantized vision transformers (ViTs). In this regime, gradient methods break down, since they cannot reliably reach local minima. In our work, dubbed Evol-Q, we use evolutionary search to effectively traverse the non-smooth landscape. Additionally, we propose using an infoNCE loss, which not only helps combat overfitting on the small calibration dataset (1,000 images) but also makes traversing such a highly non-smooth surface easier. Evol-Q improves the top-1 accuracy of a fully quantized ViT-Base by 10.30%, 0.78%, and 0.15% for 3-bit, 4-bit, and 8-bit weight quantization levels. Extensive experiments on a variety of CNN and ViT architectures further demonstrate its robustness in extreme quantization scenarios.
- One high-end GPU for inference such as an RTX A5000, V100, or A100.
- Python 3.8 and PyTorch 1.9.1 (or later). We have provided environment.yml which has our library dependencies. To create a conda environment to reproduce our setup, see below:
conda env create -f environment.yml -n evol-q
conda activate evol-q
To run an initial experiment with 8-bit quantized DeiT-Tiny:
sh scripts/run.sh
This script contains details for how to quantize 3,4,8-bit models as well as model architectures.
Our repository is adapted from FQ-ViT and Meta's LeVit implementation. We have created our novel block-wise evolutionary algorithm in joint_evol_opt.py which can quantize the following model flavors out of the box:
- ViT-Base
- DeiT-Tiny, DeiT-Small, DeiT-Base
- LeViT-128S, LeViT-128, LeViT-192, LeViT-246, LeViT-384
Below, we show the accuracy for each 4-bit quantized model, averaged across seeds {0 ,1, 2}:
Model | Top-1 Accuracy |
---|---|
ViT-Base | |
DeiT-Tiny | |
DeiT-Small | |
DeiT-Base | |
LeViT-128S | |
LeViT-128 | |
LeViT-192 | |
LeViT-246 | |
LeViT-384 |