Towards the Limit of Network Quantization

Authors: Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, using the simple uniform quantization followed by Huffman coding, we show from our experiments that the compression ratios of 51.25, 22.17 and 40.65 are achievable for Le Net, 32-layer Res Net and Alex Net, respectively.
Researcher Affiliation Industry Yoojin Choi, Mostafa El-Khamy, and Jungwon Lee Samsung US R&D Center, San Diego, CA 92121, USA {yoojin.c,mostafa.e,jungwon2.lee}@samsung.com
Pseudocode Yes Algorithm 1 Iterative solution for entropy-constrained network quantization
Open Source Code No The paper does not provide any links to source code repositories, nor does it state that its code is publicly available or in supplementary materials.
Open Datasets Yes Le Net (Le Cun et al., 1998) for the MNIST data set, (b) Res Net (He et al., 2015) for the CIFAR-10 data set, and (c) Alex Net (Krizhevsky et al., 2012) for the Image Net ILSVRC-2012 data set (Russakovsky et al., 2015).
Dataset Splits Yes For the data set X used to compute Hessian in (8), we can either reuse a training data set or use some other data set, e.g., validation data set.
Hardware Specification No The paper does not provide any specific hardware details such as GPU or CPU models, memory, or cloud computing instances used for the experiments.
Software Dependencies No The paper mentions using a 'pre-trained Alex Net Caffe model' and refers to the 'Adam SGD optimizer (Kingma & Ba, 2014)' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes For Hessian computation, 50,000 samples of the training set are reused. We also evaluate the performance when Hessian is computed with 1,000 samples only.