Training Your Own Detection Models

RT-Sort allows you to train your own detection models using leave-one-out cross-validation. Here's how to do it:

Training Detection Models

from braindance.core.spikedetector.train import train_detection_model

model = train_detection_model(recordings)

recordings: A list containing the recordings to use for training the detection model (at least two).

kilosort_path: Folder where Kilosort2 is installed (if not set in environment variable "KILOSORT_PATH").
input_scale: Multiplier for recording traces before input into the detection model (default: 0.01).
learning_rate: Learning rate for training (default: 7.76e-4).
validation_recording: Recording to use as validation (default: None).

If you prefer to sort recordings with Kilosort2 separately before training detection models, you can use:

from braindance.core.spikesorter.kilosort2 import run_kilosort2

run_kilosort2(recording_files)

kilosort_path: Folder where Kilosort2 is installed.
results_folders: List of folders to store sorted results (default: None).
compile_to_npz: Whether to save results as "sorted.npz" (default: True).
compile_to_mat: Whether to save results as "sorted.mat" (default: False).
intermediate_folders: List of folders to store intermediate results (default: None).
delete_inter: Whether to delete intermediate folders after sorting (default: True).

None (results are saved to files)

The training process is divided into two main steps: spike sorting with Kilosort2 and training the detection model.
Adjust the learning rate based on the loss curve created at the end of training.
A straight loss curve might indicate a learning rate that is too low, while a curve that decreases initially but then strongly increases might indicate a rate that is too high.

For more details on parameters and advanced usage, refer to the API Reference.