Repo of visual attention mechanisms from various papers, including their implementations, testing, and studies of them in various model architectures.
Implementations are done in PyTorch (v0.3 for now).
Models are implemented as extending nn.Module, and done in a way that lets
them be imported and used in any PyTorch Module.
Mechanism tests and studies are done in Jupyter notebooks, which can be viewed (but not run) without having PyTorch installed.
- Recurrent Models of Visual Attention (Mnih, Graves, Kavukcuoglu. 2014): recurrent [visual] attention model (RAM)
- Multiple Object Recognition with Visual Attention (Ba, Mnih, Kavukcuoglu. 2014): improvements to the RAM architecture, plus an algorithm for learning multiple sequential targets