speechbrain.lobes.models.MetricGAN module
Generator and discriminator used in MetricGAN
Authors: * Szu-Wei Fu 2020
Summary
Classes:
Simple LSTM for enhancement with custom initialization. |
|
Implementation of a leanable sigmoid. |
|
Metric estimator for enhancement training. |
Functions:
Computes the shifted sigmoid. |
|
Create a layer with spectral norm, xavier uniform init and zero bias |
Reference
- speechbrain.lobes.models.MetricGAN.xavier_init_layer(in_size, out_size=None, spec_norm=True, layer_type=<class 'torch.nn.modules.linear.Linear'>, **kwargs)[source]
Create a layer with spectral norm, xavier uniform init and zero bias
- class speechbrain.lobes.models.MetricGAN.Learnable_sigmoid(in_features=257)[source]
Bases:
Module
Implementation of a leanable sigmoid.
- Parameters:
in_features (int) – Input dimensionality
- class speechbrain.lobes.models.MetricGAN.EnhancementGenerator(input_size=257, hidden_size=200, num_layers=2, dropout=0)[source]
Bases:
Module
Simple LSTM for enhancement with custom initialization.
- Parameters:
- blstm
Use orthogonal init for recurrent layers, xavier uniform for input layers Bias is 0
- class speechbrain.lobes.models.MetricGAN.MetricDiscriminator(kernel_size=(5, 5), base_channels=15, activation=<class 'torch.nn.modules.activation.LeakyReLU'>)[source]
Bases:
Module
Metric estimator for enhancement training.
- Consists of:
four 2d conv layers
channel averaging
three linear layers
- Parameters: