Hierarchical-ConvLSTM

Hierarchical ConvLSTM has a multi-layer architecture and output of each layer is used for classification.

How to use

The Hierarchical ConvLSTM module derives from nn.Module so it can be used as any other PyTorch module.

Example usage:

model = HConvLSTM(input_channels=1024,
                  idden_channels=[512,256,128],
                  kernel_size=3,
                  num_classes=101)

TODO (in progress...)

• Add spatial attention
• Replace the average pooling by weighted pooling with spatial attention as weights
• Add example useage on the task of video-based action recognition
• ...

Disclaimer

This is still a work in progress and is far from being perfect: if you find any bug please don't hesitate to open an issue.