ReX: An Efficient Approach to Reducing Memory Cost in Image Classification
Authors: Xuwei Qian, Renlong Hang, Qingshan Liu2099-2107
AAAI 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two benchmark datasets, i.e., Visual Wake Words, Image Net-1k, demonstrate that our method consistently reduces the peak RAM and average latency of a wide variety of adaptive models on low-power devices. |
| Researcher Affiliation | Academia | Xuwei Qian, Renlong Hang*, Qingshan Liu Nanjing University of Information Science and Technology, Nanjing, China EMAIL, renlong EMAIL, EMAIL |
| Pseudocode | No | The paper describes the inference process of CBEE using text and equations (e.g., in the 'Inference' section and Figure 2), but it does not include a formally labeled 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | No | The paper does not contain any statements about making its source code publicly available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Our experiments are based on two widely-used visual datasets: (1) Visual Wake Words is a binary classification dataset proposed by (Chowdhery et al. 2019)...; (2) Image Net (Deng et al. 2009) is a 1000-class dataset from ILSVRC2012, with 1.2 million images for training and 50000 images for validation. |
| Dataset Splits | Yes | Visual Wake Words is a binary classification dataset proposed by (Chowdhery et al. 2019). The dataset contains a total of 115K training images and 8K validation images. ... Image Net (Deng et al. 2009) is a 1000-class dataset from ILSVRC2012, with 1.2 million images for training and 50000 images for validation. |
| Hardware Specification | Yes | Experiments on an ARM processor and an i Phone. Since the proposed Re X is designed for edge devices, we investigate the practical inference speed of our method on an ARM processor1 and an i Phone 12 (with Apple A14 Bionic) using Pytorch Mobile2. The single-thread mode with batch size 1 is used following (Howard et al. 2019). 1Quad-Core ARM Cortex-A57 MPCore combined with Dual Core NVIDIA Denver 2 64-Bit CPU. |
| Software Dependencies | No | The paper mentions 'Pytorch Mobile' in the context of conducting experiments on an i Phone, but it does not specify a version number for PyTorch Mobile or any other software dependencies like Python, CUDA, or specific libraries with their versions. |
| Experiment Setup | Yes | Ablation: Re X. We first consider the changes in accuracy, peak RAM, and FLOPs on different hyperparameters of Re X like patch size, hidden dimensions, and stride. ... Table 1: Re X-Mobile Net V2 (6 exits) : Re X Layer with patchsize 6 6 and hidden sizes h1 =16, h2 =8 is used. ... Impact of Predefined Count k. As illustrated in Figures 4 and 7, varying count k can lead to different speed-up ratios and performance. |