LSTM with convolution instead of matrix multiplication (ConvLSTM)

Hello Community

I have already used Julia for many tasks but not for machine learning purposes. Now, I would need a LSTM network, which uses internal convolutional operations (not convolved matrices as input). Such a net is called ConvLSTM which is already implemented in Python packages. I would use ConvLSTMs to model short range forecasts for precipitation fields. Details can be found in publication (equation 3)

Up to now, I did not found hints for be available within the Flux tool. Do somebody know, if such a ConvLSTM is under development in Julia and when it will be available?

Thanks for any comments,
Dieter

3 Likes

It is not, to my knowledge. If you or anyone else implements one, Iā€™d be more than happy to help guide them with upstreaming that into Flux itself.