Exploring the sequence length bottleneck in the Transformer for Image Captioning
About
Most recent state of the art architectures rely on combinations and variations of three approaches: convolutional, recurrent and self-attentive methods. Our work attempts in laying the basis for a new research direction for sequence modeling based upon the idea of modifying the sequence length. In order to do that, we propose a new method called "Expansion Mechanism" which transforms either dynamically or statically the input sequence into a new one featuring a different sequence length. Furthermore, we introduce a novel architecture that exploits such method and achieves competitive performances on the MS-COCO 2014 data set, yielding 134.6 and 131.4 CIDEr-D on the Karpathy test split in the ensemble and single model configuration respectively and 130 CIDEr-D in the official online evaluation server, despite being neither recurrent nor fully attentive. At the same time we address the efficiency aspect in our design and introduce a convenient training strategy suitable for most computational resources in contrast to the standard one. Source code is available at https://github.com/jchenghu/exploring
Related benchmarks
No related benchmarks have been indexed for this paper yet.