Shape adaptor: a learnable resizing module
File(s)
Accepted version
OA Location
Author(s)
Type
Conference Paper
Abstract
We present a novel resizing module for neural networks: shape adaptor, a drop-in enhancement built on top of traditional resizing layers, such as pooling, bilinear sampling, and strided convolution. Whilst traditional resizing layers have fixed and deterministic reshaping factors, our module allows for a learnable reshaping factor. Our implementation enables shape adaptors to be trained end-to-end without any additional supervision, through which network architectures can be optimised for each individual task, in a fully automated way. We performed experiments across seven image classification datasets, and results show that by simply using a set of our shape adaptors instead of the original resizing layers, performance increases consistently over human-designed networks, across all datasets. Additionally, we show the effectiveness of shape adaptors on two other applications: network compression and transfer learning.
Date Issued
2020-10-07
Date Acceptance
2020-07-02
Citation
Lecture Notes in Computer Science, 2020, 12357, pp.661-677
ISBN
978-3-030-58579-2
ISSN
0302-9743
Publisher
Springer Verlag
Start Page
661
End Page
677
Journal / Book Title
Lecture Notes in Computer Science
Volume
12357
Copyright Statement
© Springer Nature Switzerland AG 2020. The final publication is available at Springer via https://link.springer.com/chapter/10.1007/978-3-030-58610-2_39
Source
European Conference on Computer Vision 2020
Subjects
Artificial Intelligence & Image Processing
Publication Status
Published
Start Date
2020-08-23
Finish Date
2020-08-28
Coverage Spatial
Glasgow, UK (Virtual)
Date Publish Online
2020-10-07