search

stuarteiffert / RNN-for-Human-Activity-Recognition-using-2D-Pose-Input

Activity Recognition from 2D pose using an LSTM RNN
291 stars 77 forks source link

Question: n_steps #7

Open SlavaKeshkov opened 6 years ago

SlavaKeshkov commented 6 years ago

Hi Stuart!

Can I ask you how did you choose: n_steps = 32? Does it have to do with the average length of clips in the dataset? Were they all equal?

Trying to reproduce the code with another data (less observations, two classes only) and wondering how to set the parameters.

stuarteiffert commented 6 years ago

It was a parameter found experimentally, I tried values from about 16-64 to compare. This doesn't actually need to remain constant. It needs to be the same within a batch for training, but can be variable between batches, and for inference, I just didn't implement that.

The clips weren't all of equal length, so this just helped to get a constant number of frames for use within a batch. Total frames per clip was about 100 on average, so this split the clips without wasting too many frames. I would recommend splitting the clips you use into varied lengths and using all for training (making sure your testing set is completely separate, taken out prior).