speechbrain.utils.kmeans module

Utilities for training kmeans model.

Author

• Pooneh Mousavi 2023

Summary

Functions:

accumulate_and_extract_features	Extract features (output of SSL model) and acculamte them on cpu to be used for clustering.
fetch_kmeans_model	Return a k-means clustering model with specified parameters.
process_chunks	Process data in chunks of a specified size.
save_model	Save a Kmeans model .
train	Train a Kmeans model .

Reference

speechbrain.utils.kmeans.accumulate_and_extract_features(batch, features_list, ssl_model, ssl_layer_num, device)

Extract features (output of SSL model) and acculamte them on cpu to be used for clustering.

Parameters:

• batch (tensor) – Single batch of data.

• features_list (list) – accumulate features list.

• ssl_model (torch.nn.Module) – SSL-model used to extract features used for clustering.

• ssl_layer_num (int) – specify output of which layer of the ssl_model should be used.

• device (str) – cpu or cuda device.

speechbrain.utils.kmeans.fetch_kmeans_model(n_clusters, init, max_iter, batch_size, tol, max_no_improvement, n_init, reassignment_ratio, random_state, checkpoint_path)

Return a k-means clustering model with specified parameters.

Parameters:

• n_clusters (MiniBatchKMeans) – The number of clusters to form as well as the number of centroids to generate.

• init (int) – Method for initialization: {‘k-means++’’, ‘’random’’}

• max_iter (int) – Maximum number of iterations over the complete dataset before stopping independently of any early stopping criterion heuristics.

• batch_size (int) – Size of the mini batches.

• tol (float) – Control early stopping based on the relative center changes as measured by a smoothed, variance-normalized of the mean center squared position changes.

• max_no_improvement (int) – Control early stopping based on the consecutive number of mini batches that does not yield an improvement on the smoothed inertia.

• n_init (int) – Number of random initializations that are tried

• reassignment_ratio (float) – Control the fraction of the maximum number of counts for a center to be reassigned.

• random_state (int) – Determines random number generation for centroid initialization and random reassignment.

• checkpoint_path (str) – Path to saved model.

Returns:

a k-means clustering model with specified parameters.

Return type:

MiniBatchKMeans

speechbrain.utils.kmeans.process_chunks(data, chunk_size, model)

Process data in chunks of a specified size.

Parameters:

• data (list) – The list of integers to be processed.

• chunk_size (int) – The size of each chunk.

• model (MiniBatchKMeans) – The initial kmeans model for training.

speechbrain.utils.kmeans.train(model, train_set, ssl_model, save_path, ssl_layer_num, kmeans_batch_size=1000, device='cpu', checkpoint_interval=10)

Train a Kmeans model .

Parameters:

• model (MiniBatchKMeans) – The initial kmeans model for training.

• train_set (Dataloader) – Batches of tarining data.

• ssl_model (torch.nn.Module) – SSL-model used to extract features used for clustering.

• save_path (string) – Path to save intra-checkpoints and dataloader.

• ssl_layer_num (int) – Specify output of which layer of the ssl_model should be used.

• kmeans_batch_size (int) – Size of the mini batches.

• device (str) – cpu or cuda device.

• checkpoint_interval (int) – Determine at which iterations to save the checkpoints.

speechbrain.utils.kmeans.save_model(model, checkpoint_path)

Save a Kmeans model .

Parameters:

• model (MiniBatchKMeans) – The kmeans model to be saved.

• checkpoint_path (str) – Path to save the model.