Skip to main content

Real-time Application (Online Use)

Once you have detected sequences, you can use RT-Sort for real-time applications. Here's how to do it:

Reset the RTSort Object​

Before online use, reset the RTSort object:

rt_sort.reset()

Note: For optimal performance, call rt_sort.reset() if more than 50ms have passed since the last rt_sort.running_sort(obs) method call.

Continuous Sorting​

During online use, continuously call the following method to sort a stream of data:

sequence_detections = rt_sort.running_sort(obs)

Parameters​

  • obs: NumPy array with shape (num_frames, num_electrodes). This should contain only the most recent frames of data.

Returns​

A list where each element is a tuple of length 2 containing the data for a sorted spike in the obs recording chunk:

  • 0th element: ID number of the sequence that the spike was assigned to
  • 1st element: Time the spike occurred (in milliseconds)

Notes​

  • The first 50ms of data passed as obs to running_sort is needed to initialize the RTSort object after the last rt_sort.reset(), so no spikes will be sorted during this time.
  • num_frames must be at least 1 and can be different for each running_sort call.
  • To minimize sorting latency, num_frames should ideally change and be equal to the amount of time that has passed since the last running_sort call.
  • To align with the RT-Sort Methods, use 100 frames for 20kHz MEAs and 150 frames for 30kHz Neuropixels.

Example with Maxwell MEA​

Here's an example of how to use RT-Sort with a Maxwell MEA:

from braindance.core.maxwell_env import MaxwellEnv

env = MaxwellEnv(**params)
done = False
while not done:
    obs, done = env.step()
    sequence_detections = rt_sort.running_sort(obs)
    # Process sequence_detections

Next Steps​

If you want to customize RT-Sort further, you might be interested in training your own models.