LSTM seq2seq input and output with different number of time steps

Question

I am new to this field and currently working on a video action prediction project using keras. The input data takes 10% frames of each video and convert all same successive