Input sequence padding about indrnn_pytorch

[Sunnydreamrain]

Hi,

Yes. The sequences are filled with zero to be 300. That is why we used another data array 'len' to indicate the real length of the data.

[Anirudh257]

@Sunnydreamrain, but in your approach, you are randomly sampling 20 frames from each sequence. If a sequence is padded with many 0's ( like the sequences having the length 154, 155 as specified in the above example), there is a chance that we pick only the 0's consisting frames during training causing a loss in training. Isn't it better to create variable length samples in the .npy files?

[Sunnydreamrain]

@Anirudh257 The "len" arrary indicates the real length of the data, thus you can sample 20 frames within the real data.

Yes. It is better to create variable length samples in the npy files if it can be loaded to the memory and does not increase reading time.

[Anirudh257]

@Sunnydreamrain, thanks for the clarification.