xvector

Copyright 2014-2020 Yevhenii Prokopalo, Anthony Larcher

class nnet.xvector.Xtractor(speaker_number, model_archi=None)

Class that defines an x-vector extractor based on 5 convolutional layers and a mean standard deviation pooling

forward(x, is_eval=False)

Parameters

• x –

• is_eval –

Returns

sequence_network_weight_decay = None

Prepapre last part of the network (after pooling)

nnet.xvector.cross_validation(model, validation_loader, device)

Parameters

• model –

• validation_loader –

• device –

Returns

nnet.xvector.extract_idmap(args, segment_indices, fs_params, idmap_name, output_queue)

Function that takes a model and an idmap and extract all x-vectors based on this model and return a StatServer containing the x-vectors

Parameters

• args –

• segment_indices –

• fs_params –

• idmap_name –

• output_queue –

Returns

nnet.xvector.extract_parallel(args, fs_params)

Parameters

• args –

• fs_params –

Returns

nnet.xvector.get_lr(optimizer)

Parameters

optimizer –

Returns

nnet.xvector.save_checkpoint(state, is_best, filename='checkpoint.pth.tar', best_filename='model_best.pth.tar')

Parameters

• state –

• is_best –

• filename –

• best_filename –

Returns

nnet.xvector.train_epoch(model, epoch, training_loader, optimizer, log_interval, device, clipping=False)

Parameters

• model –

• epoch –

• training_loader –

• optimizer –

• log_interval –

• device –

• clipping –

Returns

nnet.xvector.xtrain(speaker_number, dataset_yaml, epochs=100, lr=0.01, model_yaml=None, model_name=None, tmp_model_name=None, best_model_name=None, multi_gpu=True, clipping=False, num_thread=1)

Parameters

• speaker_number –

• dataset_yaml –

• epochs –

• lr –

• model_yaml –

• model_name –

• tmp_model_name –

• best_model_name –

• multi_gpu –

• clipping –

• num_thread –

Returns