In this project, we try to reproduce the results from the paper titled Residual Attention Network for Image Classification [4]. The network in question is generated by stacking series of Attention Modules. These modules generate attention-aware features, which change adaptively as a function of the depth in the network. A brief introduction into the concept of attention for learning and residual networks is developed and motivated for the image classification task. Some relevant literature which inspired the work of the authors is also mentioned. The structure and functioning of each component of the Residual Attention Network is described, and the authors’ implementations are listed for comparison. The Residual Attention Network is tested on benchmark datasets, namely CIFAR-10, CIFAR-100 and ImageNet. The code for the attention module is reproduced and the performance our own implementation is compared with the results of the paper, and the shortcomings are documented comprehensively.
vm2656 / residual-attention-network-for-image-classification Goto Github PK
View Code? Open in Web Editor NEWIn this project, we try to reproduce the results from the paper titled Residual Attention Network for Image Classification [4]. The network in question is generated by stacking series of Attention Modules. These modules generate attention-aware features, which change adaptively as a function of the depth in the network. A brief introduction into the concept of attention for learning and residual networks is developed and motivated for the image classification task. Some relevant literature which inspired the work of the authors is also mentioned. The structure and functioning of each component of the Residual Attention Network is described, and the authors’ implementations are listed for comparison. The Residual Attention Network is tested on benchmark datasets, namely CIFAR-10, CIFAR-100 and ImageNet. The code for the attention module is reproduced and the performance our own implementation is compared with the results of the paper, and the shortcomings are documented comprehensively.