Resolution Adaptive Networks for Efficient Inference
Abstract: In this paper the authors (Yang et al.) propose a Resolution Adaptive CNN network (RANet) to efficiently classify images at different resolutions depending based on the classification difficulty (prediction confidence).The main idea is that some images are easy to classify, only requiring low resolution and simple features (low computational cost) while other images require higher resolution and larger networks (large computational cost). RANet attempts to classify a low-resolution representation of the image and if the classification confidence is below some threshold, will attempt will try to classify the image with higher and higher resolution with subsequent subnetworks. The authors make two main assertions that we (The Einsteins) hope to verify in this reproducibility study. Firstly, does the RANet architecture offer significant reductions in computational cost, while maintaining accuracy to comparable models. Secondly, in an arbitrary dataset there’s such a distribution of ‘easy’ and ‘hard’ classifications tasks that would warrant the use of and adaptive networks. We tested the first assumption by creating our implementation of RANet in Tensorflow Keras and found it to have comparable accuracy withVGG16 and Resnet50, while having a reduction in computational time by up to 64%. We tested the second assumption by creating our own “spalling” classifier using the authors’ provided code and kept track of how often each subnetwork returned the prediction. We found that the assertion that some images are ‘easier’ or more ‘difficult’ to classify is valid, with 74% of spalling images being considered ‘hard’ to classify and the rest of the spalling images being ‘easy’ to classify. While, we have not reproduced the authors’ results directly, we show that the proposed method shows promise and may have application for real-time and/or AI at the edge applications.
Paper Url: https://openreview.net/forum?id=Fhc2m204jFms
3 Replies
Loading