PTQ4ViT

Post-Training Quantization Framework for Vision Transformers. We use the twin uniform quantization method to reduce the quantization error on these activation values. And we use a Hessian guided metric to evaluate different scaling factors, which improves the accuracy of calibration with a small cost. The quantized vision transformers (ViT, DeiT, and Swin) achieve near-lossless prediction accuracy (less than 0.5% drop at 8-bit quantization) on the ImageNet classification task. Please read the paper for details.

Updates

19/07/2022 Add discussion on Base PTQ, and provide more ablation study results.

Number of Calibration Images

Model	W8A8 #ims=32	W6A6 #ims=32	W8A8 #ims=128	W6A6 #ims=128
ViT-S/224/32	75.58	71.91	75.54	72.29
ViT-S/224	81.00	78.63	80.99	78.44
ViT-B/224	84.25	81.65	84.27	81.84
ViT-B/384	85.83	83.35	85.81	83.84
DeiT-S/224	79.47	76.28	79.41	76.51
DeiT-B/224	81.48	80.25	81.54	80.30
DeiT-B/384	82.97	81.55	83.01	81.67
Swin-T/224	81.25	80.47	81.27	80.30
Swin-S/224	83.11	82.38	83.15	82.38
Swin-B/224	85.15	84.01	85.17	84.15
Swin-B/384	86.39	85.39	86.36	85.45

Model	Time #ims=32	Time #ims=128
ViT-S/224/32	2 min	5 min
ViT-S/224	3 min	7 min
ViT-B/224	4 min	13 min
ViT-B/384	12 min	43 min
DeiT-S/224	3 min	7 min
DeiT-B/224	4 min	16 min
DeiT-B/384	14 min	52 min
Swin-T/224	3 min	9 min
Swin-S/224	8 min	17 min
Swin-B/224	10 min	23 min
Swin-B/384	25 min	69 min

One of the targets of PTQ4ViT is to quickly quantize a vision transformer. We have proposed to pre-compute the output and gradient of each layer and compute the influence of scaling factor candidates in batches to reduce the quantization time. As demonstrated in the second table, PTQ4ViT can quantize most vision transformers in several minutes using 32 calibration images. Using 128 calibration images significantly increases the quantization time.
We observe the Top-1 accuracy varies slightly in the first table, demonstrating PTQ4ViT is not very sensitive to the number of calibration images.

Base PTQ

Base PTQ is a simple quantization strategy and serves as a benchmark for our experiments. Like PTQ4ViT, we quantize all weights and inputs for fully-connect layers (including the first projection layer and the last prediction layer), as well as all input matrices of matrix multiplication operations. For fully-connected layers, we use layerwise scaling factors $\Delta_W$ for weight quantization and $\Delta_X$ for input quantization; while for matrix multiplication operations, we use $\Delta_A$ and $\Delta_B$ for A's quantization and B's quantization respectively.

To get the best scaling factors, we apply a linear grid search on the search space. The same as EasyQuantand Liu et al., we take hyper-parameters $\alpha=0.5$, $\beta = 1.2$, one search round and use cosine distance as the metric. Note that in PTQ4ViT, we change the hyper-parameters to $\alpha=0$, $\beta = 1.2$ and three search rounds, which slightly improves the performance.

It should be noticed that Base PTQ adopts a parallel quantization paradigm, which makes it essentially different from sequential quantization paradigms such as EasyQuant. In sequential quantization, the input data of the current quantizing layer is generated with all previous layers quantizing weights and activations. While in parallel quantization, the input data of the current quantizing layer is simply the raw output of the previous layer.

In practice, we found sequential quantization on vision transformers suffers from significant accuracy degradation on small calibration datasets. While parallel quantization shows robustness on small calibration datasets. Therefore, we choose parallel quantization for both Base PTQ and PTQ4ViT.

More Ablation Study

We supply more ablation studies for the hyper-parameters. It is enough to set the number of quantization intervals $\ge$ 20 (accuracy change $< 0.3%$). It is enough to set the upper bound of m $\ge$ 15 (no accuracy change). The best settings of alpha and beta vary from different layers. It is appropriate to set $\alpha=0$ and $\beta=1/2^{k-1}$, which has little impact on search efficiency. We observe that search rounds has little impact on the prediction accuracy (accuracy change $<$ 0.05% when search rounds $>1$).

We randomly take 32 calibration images to quantize different models 20 times and we observe the fluctuation is not significant. The mean/std of accuracies are: ViT-S/32 $75.55%/0.055%$ , ViT-S $80.96%/0.046%$, ViT-B $84.12%/0.068%$, DeiT-S $79.45%/0.094%$ , and Swin-S $83.11%/0.035%$.

15/01/2022 Add saved quantized models with PTQ4ViT.

model	link
ViT-S/224/32	Google
ViT-S/224	Google
ViT-B/224	Google
ViT-B/384	Google
DeiT-S/224	Google
DeiT-B/224	Google
DeiT-B/384	Google
Swin-T/224	Google
Swin-S/224	Google
Swin-B/224	Google
Swin-B/384	Google

10/12/2021 Add utils/integer.py, you can now:

convert calibrated fp32 model into int8
register pre-forward hook in the model, and fetch activation in int8. (We use uint8 to store results of twin quantization, please refer to the paper to see the bits' layout).

Install

Requirement

python>=3.5
pytorch>=1.5
matplotlib
pandas
timm

Datasets

To run example testing, you should put your ImageNet2012 dataset in path /datasets/imagenet.

We use ViTImageNetLoaderGenerator in utils/datasets.py to initialize our DataLoader. If your Imagenet datasets are stored elsewhere, you'll need to manually pass its root as an argument when instantiating a ViTImageNetLoaderGenerator.

Usage

1. Run example quantization

To test on all models with BasePTQ/PTQ4ViT, run

python example/test_all.py

To run ablation testing, run

python example/test_ablation.py

You can run the testing scripts with multiple GPUs. For example, calling

python example/test_all.py --multigpu --n_gpu 6

will use 6 gpus to run the test.

2. Download quantized model checkpoints

(Coming soon)

Results

Results of BasePTQ

model	original	w8a8	w6a6
ViT-S/224/32	75.99	73.61	60.144
ViT-S/224	81.39	80.468	70.244
ViT-B/224	84.54	83.896	75.668
ViT-B/384	86.00	85.352	46.886
DeiT-S/224	79.80	77.654	72.268
DeiT-B/224	81.80	80.946	78.786
DeiT-B/384	83.11	82.33	68.442
Swin-T/224	81.39	80.962	78.456
Swin-S/224	83.23	82.758	81.742
Swin-B/224	85.27	84.792	83.354
Swin-B/384	86.44	86.168	85.226

Results of PTQ4ViT

model	original	w8a8	w6a6
ViT-S/224/32	75.99	75.582	71.908
ViT-S/224	81.39	81.002	78.63
ViT-B/224	84.54	84.25	81.65
ViT-B/384	86.00	85.828	83.348
DeiT-S/224	79.80	79.474	76.282
DeiT-B/224	81.80	81.482	80.25
DeiT-B/384	83.11	82.974	81.55
Swin-T/224	81.39	81.246	80.47
Swin-S/224	83.23	83.106	82.38
Swin-B/224	85.27	85.146	84.012
Swin-B/384	86.44	86.394	85.388

Results of Ablation

ViT-S/224 (original top-1 accuracy 81.39%)

Hessian Guided	Softmax Twin	GELU Twin	W8A8	W6A6
			80.47	70.24
✓			80.93	77.20
✓	✓		81.11	78.57
✓		✓	80.84	76.93
	✓	✓	79.25	74.07
✓	✓	✓	81.00	78.63

ViT-B/224 (original top-1 accuracy 84.54%)

Hessian Guided	Softmax Twin	GELU Twin	W8A8	W6A6
			83.90	75.67
✓			83.97	79.90
✓	✓		84.07	80.76
✓		✓	84.10	80.82
	✓	✓	83.40	78.86
✓	✓	✓	84.25	81.65

ViT-B/384 (original top-1 accuracy 86.00%)

Hessian Guided	Softmax Twin	GELU Twin	W8A8	W6A6
			85.35	46.89
✓			85.42	79.99
✓	✓		85.67	82.01
✓		✓	85.60	82.21
	✓	✓	84.35	80.86
✓	✓	✓	85.89	83.19

Citation

@article{PTQ4ViT_arixv2022,
    title={PTQ4ViT: Post-Training Quantization Framework for Vision Transformers},
    author={Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, Guangyu Sun},
    journal={arXiv preprint arXiv:2111.12293},
    year={2022},
}

hahnyuan / PTQ4ViT