Robust and Efficient Quantization-aware Training via Coreset Selection

Authors: Xijie Huang, Zechun Liu, Shih-Yang Liu, Kwang-Ting Cheng

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on various networks (Res Net-18, Mobile Net V2, Retina Net), datasets(CIFAR-10, CIFAR-100, Image Net-1K, COCO), and under different quantization settings. Specifically, our method can achieve an accuracy of 68.39% of 4-bit quantized Res Net-18 on the Image Net-1K dataset with only a 10% subset, which has an absolute gain of 4.24% compared to the baseline. Our method can also improve the robustness of QAT by removing noisy samples in the training set.
Researcher Affiliation Collaboration 1Hong Kong University of Science and Technology (HKUST), 2Meta Reality Lab EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode Yes The pseudo-code for our ACS algorithm is shown in Alg. 1.
Open Source Code No The paper does not explicitly state that source code for the methodology is openly available, nor does it provide a link to a code repository.
Open Datasets Yes The efficiency experiments are conducted on CIFAR-100 (Krizhevsky et al., 2009) and Image Net-1K dataset (Deng et al., 2009). We evaluate Mobile Net V2 (Howard et al., 2017) on CIFAR-100 and evaluate Res Net-18 (He et al., 2016) on the Image Net-1K dataset. We further provide experimental results of quantized Retina Net Lin et al. (2017) on MS COCO object detection benchmark (Lin et al., 2014).
Dataset Splits Yes For 2-bit weights-only quantization of Mobile Net V2 on the CIFAR-100 dataset, QAT based on our ACS can achieve a mean accuracy of 67.19% with only 50% training data used for training per epoch. For 4-bit quantization of Res Net-18 on the Image Net-1K dataset, our ACS can achieve top-1 accuracy of 68.39% compared to the 64.15% of random selection when only 10% training data is selected for the training of every epoch. We investigate two coreset fractions: 10% and 50% and use R=10. When full data are selected (S = 100%), the mean accuracy and standard deviation is 68.1 0.9%. When full data are selected (S = 100%), the accuracy is 72.46%.
Hardware Specification Yes All the experiments were carried out on 2 NVIDIA RTX 3090 GPUs.
Software Dependencies No The paper mentions using SGD and Adam optimizers, and the LSQ+ quantization method, but it does not specify version numbers for any software libraries or dependencies (e.g., PyTorch, TensorFlow, CUDA).
Experiment Setup Yes For Mobile Net V2, we train the network for 200 epochs using a learning rate of 0.01, weight decay of 5e-4, batch size of 512, R = 20, and SGD optimizer. For Res Net-18 on Image Net-1K, we train the network for 120 epochs using a learning rate of 1.25e-3, no weight decay, batch size of 512, R = 10, and Adam optimizer.