jdiaz4302 / convtimelstm Goto Github PK

ConvTimeLSTM performs poorly on MissingMNIST, however it performs great on MovingMNIST. Perhaps saving the weights from the MovingMNIST-trained-model will provide a good starting point for MissingMNIST because they differ only in non-uniform time changes

Results are much worse than ConvLSTM paper; however, we have much less "data cases". Getting the same number of data cases would reasonably requiring stronger/parallel GPU because a minibatch size of 8 maxes out the current GPU.

It would be worth adding an visualization/evaluation of what happens in the 06_0X notebooks when the trained ConvTimeLSTM is provided the same sequence of images and time steps however the final time step is altered to 1, 2, 3, etc...

The current implementation doesn't have any out-of-sample data. Needs done, will be easy enough