CNNpack: Packing Convolutional Neural Networks in the Frequency Domain
Authors: Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, Chao Xu
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The compression and speed-up ratios of the proposed algorithm are thoroughly analyzed and evaluated on benchmark image datasets to demonstrate its superiority over state-of-the-art methods. |
| Researcher Affiliation | Academia | Yunhe Wang1,3, Chang Xu2, Shan You1,3, Dacheng Tao2, Chao Xu1,3 1Key Laboratory of Machine Perception (MOE), School of EECS, Peking University 2Centre for Quantum Computation and Intelligent Systems, School of Software, University of Technology Sydney 3Cooperative Medianet Innovation Center, Peking University |
| Pseudocode | No | The procedures of the proposed algorithm are detailed in the supplementary materials. |
| Open Source Code | No | The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described in this paper. |
| Open Datasets | Yes | The evaluation was conducted using the MNIST and ILSVRC2012 datasets. Compression Alex Net and VGGNet on Image Net. We next employed CNNpack for CNN compression on the Image Net ILSVRC-2012 dataset [18], which contains over 1.2M training images and 50k validation images. We first tested their impact on the network accuracy by conducting an experiment using MNIST [21]. |
| Dataset Splits | Yes | We next employed CNNpack for CNN compression on the Image Net ILSVRC-2012 dataset [18], which contains over 1.2M training images and 50k validation images. |
| Hardware Specification | Yes | All methods were implemented using Mat Conv Net [21] and run on NVIDIA Titan X graphics cards. |
| Software Dependencies | No | The paper mentions the use of 'Mat Conv Net [21]' but does not provide specific version numbers for this or any other software dependencies. |
| Experiment Setup | Yes | The proposed compression method has several important parameters: λ, d, K, b, and Ω. We first tested their impact on the network accuracy by conducting an experiment using MNIST [21], where the network has two convolutional layers and two fullyconnected layers of size 5 5 1 20, 5 5 20 50, 4 4 50 500, and 1 1 500 10, respectively. The model accuracy was 99.06%. The compression results of different λ and d after fine-tuning 20 epochs are shown in Fig. 2 in which k was set as 16, b was equal to + since it did not make an obvious contribution to the compression ratio even when set at a relatively smaller value (e.g., b = 0.05) but caused the accuracy reduction. Ωwas set to 500... Based on the above analysis, we kept λ = 0.04 and K = 16 for this network (an accuracy of 99.14%). |