Learning a Wavelet-Like Auto-Encoder to Accelerate Deep Neural Networks
Authors: Tianshui Chen, Liang Lin, Wangmeng Zuo, Xiaonan Luo, Lei Zhang
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we first apply our method to the widely used VGG16-Net and conduct extensive evaluations on two large-scale datasets, i.e., the Image Net dataset for image classification and the CACD dataset for face identification. Our method achieves an acceleration rate of 3.13 with merely 0.22% top-5 accuracy drop on Image Net (see Tabel 1). |
| Researcher Affiliation | Academia | 1School of Data and Computer Science, Sun Yat-sen University, Guangzhou, China 2School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China 3School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin, China 4Department of Computing, The Hong Kong Polytechnic University, Hong Kong |
| Pseudocode | No | The paper includes architectural diagrams (Figure 1, Figure 2) but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | In the experiments, we first apply our method to the widely used VGG16-Net and conduct extensive evaluations on two large-scale datasets, i.e., the Image Net dataset for image classification and the CACD dataset for face identification. Image Net dataset (Russakovsky et al. 2015). CACD is a large-scale and challenging dataset for face identification. It contains 163,446 images of 2,000 identities collected from the Internet that vary in age, pose and illumination. A subset of 56,138 images that cover 500 identities are manually annotated (Lin et al. 2017b). |
| Dataset Splits | Yes | The dataset covers 1,000 classes and comprises a training set of about 1.28 million images and a validation set of 50,000 images. (ImageNet) and We randomly select 44,798 images as the training set and the rest as the test set. (CACD) |
| Hardware Specification | Yes | The execution time is computed with a C++ implementation on Intel i7 CPU (3.50GHz) and Nvidia Ge Force GTX TITAN-X GPU. |
| Software Dependencies | No | The paper mentions "C++ implementation" but does not specify any software dependencies (libraries, frameworks) with version numbers. |
| Experiment Setup | Yes | The parameters are initialized with the Xavier algorithm (Glorot and Bengio 2010) and the WAE is trained using SGD algorithm with a mini-batch of 4, momentum of 0.9 and weight decay of 0.0005. We set the initial learning rate as 0.000001, and divide it by 10 after 10 epochs. The network is also trained using SGD algorithm with the same momentum and weight decay as Stage 1. The mini-batch is set as 256, and the learning rate is initialized as 0.01 (0.1 if using Res Net-50 as the baseline), which is divided by 10 when the error plateaus. The network is fine tuned using SGD with the mini-batch size, momentum, and weight decay the same as Stage 2. We utilize a small learning rate of 0.0001 and train the network until the error plateaus. |