Wavelet Feature Maps Compression for Image-to-Image CNNs

Abstract

Convolutional Neural Networks (CNNs) are known for requiring extensive computational resources, and quantization is among the best and most common methods for compressing them. While aggressive quantization (i.e., less than 4-bits) performs well for classification, it may cause severe performance degradation in image-to-image tasks such as semantic segmentation and depth estimation. In this paper, we propose Wavelet Compressed Convolution (WCC)-a novel approach for high-resolution activation maps compression integrated with point-wise convolutions, which are the main computational cost of modern architectures. To this end, we use an efficient and hardware-friendly Haar-wavelet transform, known for its effectiveness in image compression, and define the convolution on the compressed activation map. We experiment on various tasks, that benefit from high-resolution input, and by combining WCC with light quantization, we achieve compression rates equivalent to 1-4bit activation quantization with relatively small and much more graceful degradation in performance. * Contributed equally.Preprint. Under review.

ASC: Adaptive Scale Feature Map Compression for Deep Neural Network

“…To address this task, we selected the popular EDSR network [13] with its basic configuration for 2× scaling, trained on the DIV2K dataset [14]. Following a similar approach to [5], we focus our compression efforts solely on the body of the network. The model has a baseline PSNR of 35.0581.…”

“…These methods can be executed in the time domain, as seen with Least-Squares Fitting Compression (LSFC) [3] or revised texture compression [4]. Alternatively, they can be applied in the frequency domain with different transforms [5], [6]. However, LSFC offers a limited compression rate.…”

“…The study in [4] modified S3TC [7] to have 32-point interpolation, but also leads to a low compression rate. The approach in [5] only supports 1x1 convolutions. Moreover, the transforms presented in [5], [6] come with a significant computational cost.…”

“…The approach in [5] only supports 1x1 convolutions. Moreover, the transforms presented in [5], [6] come with a significant computational cost.…”

Compressing the Activation Maps in Deep Convolutional Neural Networks and the Regularization Effect of Compression

“…Huffman coding [39] and arithmetic coding [40] are two other methods proposed in the research community for accomplishing lossless weight compression. Finder et al [41] proposed Wavelet Compressed Convolution (WCC), a compression method based on the Wavelet Transform (WT) for high-resolution feature maps. The WCC was integrated with point-wise convolutions and light quantization (1-4 bits).…”