This is the source code of model evalution/inference for MPCViT: Searching for Accurate and Efficient MPC-friendly Vision Transformer with Heterogeneous Attention, which has been accpeted by ICCV 2023.
Our trained MPCViT models which are reported in the paper are avaliable here.
During inference, some of the attentions are computed using ReLU Softmax Attention while others are computed using Scaling Attention. Hence, in this implementation, attention heads are first split and then computed based on the architecture parameter alpha.
Secure multi-party computation (MPC) enables computation directly on encrypted data and protects both data and model privacy in deep learning inference. However, existing neural network architectures, including Vision Transformers (ViTs), are not designed or optimized for MPC and incur significant latency overhead. We observe Softmax accounts for the major latency bottleneck due to a high communication complexity, but can be selectively replaced or linearized without compromising the model accuracy. Hence, in this paper, we propose an MPC-friendly ViT, dubbed MPCViT, to enable accurate yet efficient ViT inference in MPC. Based on a systematic latency and accuracy evaluation of the Softmax attention and other attention variants, we propose a heterogeneous attention optimization space. We also develop a simple yet effective MPC-aware neural architecture search algorithm for fast Pareto optimization. To further boost the inference efficiency, we propose MPCViT+, to jointly optimize the Softmax attention and other network components, including GeLU, matrix multiplication, etc. With extensive experiments, we demonstrate that MPCViT achieves 1.9%, 1.3% and 3.6% higher accuracy with 6.2×, 2.9× and 1.9× latency reduction compared with baseline ViT, MPCFormer and THE-X on the Tiny-ImageNet dataset, respectively. MPCViT+ further achieves a better Pareto front compared with MPCViT.
Command examples:
Below is an example to evaluate MPCViT with $\mu=0.5$ w/o knowledge distillation (KD) on CIFAR-10.
python inference.py --config configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar10 --model-checkpoint mpcvit_checkpoints/cifar-10/mpcvit_cifar10-0.5.pth.tar
Below is an example to evaluate MPCViT with $\mu=0.5$ w/ knowledge distillation (KD) on CIFAR-10.
python inference.py --config configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar10 --model-checkpoint mpcvit_checkpoints/cifar-10/mpcvit_cifar10-0.5-kd.pth.tar
Datasets:
We provide model checkpoints on three widely used datasets, i.e., CIFAR-10, CIFAR-100 and Tiny-ImageNet.
Usage: just simply modify the command above including config, model, data_dir, model-checkpoint
.
config:
configs/datasets/cifar10.yml
configs/datasets/cifar100.yml
configs/datasets/tiny_imagenet.yml
model:
vit_7_4_32 for CIFAR-10/100
vit_9_12_64 for Tiny-ImageNet
data_dir:
set as the corresponding datasets.
set as the corresponding checkpoints.
Checkpoints:
Our trained MPCViT models which are reported in the paper are avaliable here. Your can freely download the checkpoints that you need to evaluate the model performance.
Numerical results:
Note that we missed a part of the model checkpoints during completing this paper, so we re-run our experiments again to obtain the missing checkpoints. Therefore, there is a slight fluctuation (less than 0.1%) between the model results and the results reported in the paper, which is considered to be acceptable.
Besides MPCViT, we also propose MPCViT+ to accelerate MLP blocks including GeLU and linear operators. You can evaluate MPCViT+ models on CIFAR-10/100 with the following command:
python inference.py --config configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar10 --model-checkpoint mpcvit_checkpoints/cifar-10/mpcvit_plus_cifar10.pth.tar --linear-gelu
Note that some data points need post-added ReLU after GeLU linearization.
We provide the training code for MPCViT.
To switch inference to training, the following code in src/utils/transformers.py
should be replaced with the below one:
# modified: split heads and then concat them
for i, h in enumerate(self.alpha.squeeze()):
# print('i:', i, 'h:', h)
attn_head = attn[:, i, :, :]
# print(attn_head.shape)
if h == 1: # relusoftmax
attn_head = self.relu(attn_head) / (torch.sum(self.relu(attn_head), dim=-1, keepdim=True) + self.eps)
elif h == 0: # scalattn
attn_head = attn_head / attn_head.size(-1)
attn[:, i, :, :] = attn_head
scalattn = attn / attn.size(3) # scaling attention
attn = self.relu(attn)
attn = attn / (torch.sum(attn, dim=-1, keepdim=True) + self.eps) # ReLUSoftmax attention
attn = self.alpha * attn + (1 - self.alpha) * scalattn # weighted-sum for arch searching
Below we give CIFAR-10 as an example.
Search
python train.py -c configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar-10/ --search-mode --epochs 300
Retrain
python train.py -c configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar-10/ --retrain-mode --search-ckpt /path/to/ckpt --epochs 600 --rs-ratio 0.7
- Note that you can directly use our searched alpha (see here) to retrain the model like below:
python train.py -c configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar-10/ --retrain-mode --search-ckpt ./mpcvit_cifar10-0.7.pth.tar --epochs 600 --rs-ratio 0.7
Train with knowledge distillation
python train.py -c configs/datasets/cifar10.yml --model vit_7_4_32 /path/to/cifar-10/ --retrain-mode --search-ckpt /path/to/ckpt --epochs 600 --rs-ratio 0.7 --use-token-kd --teacher-ckpt /path/to/ckpt
@inproceedings{zeng2023mpcvit,
title={Mpcvit: Searching for accurate and efficient mpc-friendly vision transformer with heterogeneous attention},
author={Zeng, Wenxuan and Li, Meng and Xiong, Wenjie and Tong, Tong and Lu, Wen-jie and Tan, Jin and Wang, Runsheng and Huang, Ru},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={5052--5063},
year={2023}
}