According to my data handler the data needs to be of dimension (#frames, 7_7_1024)  so you should resize all the (7,7,1024) 3D matrices to vectors of length 7_7_1024 and then store these in a 2D matrix (#frames, 7_7_1024).

From: https://github.com/kracwarlock/action-recognition-visual-attention/issues/1