C2W4 Video: Time Series

Course Q&A Machine Learning Engineering for Production(MLOps) Machine Learning Data Lifecycle in Production
1

Dear team,

Thank you for the wonderful course. I have a question regarding the video “Time Series” to solve the problem with variable FPS/frequency/framerate of inputs. I am working on a sign language recognition project.

Say I have the application where the frequency of the input is known but can vary widely across sequences. For example, they may be audio recordings acquired at different frequency, or videos for my work that come from mobile cameras whose framerate can vary from 60fps down to as low as 20fps.

The straightworard thing to do would be to either

  • Time series data augmentation
  • Resample inputs to a constant frequency
  • Ignore input frequency and a RNN will figure it all out
  • Other methods

None sound very appealing. Is there a better way to handle variable input frequency in RNNs?

2

The only restriction RNN has is that a single batch should be rectangular. To put it another way, each sample within a batch should have the same dimension of (num timesteps, num features per timestep). The way you accomplish this is be specifying input_shape=[None, num_features_per_timestep]. Hope this helps.