speechbrain.lobes.models.L2I module

This file implements the necessary classes and functions to implement Listen-to-Interpret (L2I) interpretation method from https://arxiv.org/abs/2202.11479v2

Authors * Cem Subakan 2022 * Francesco Paissan 2022

Summary

Classes:

NMFDecoderAudio	This class implements an NMF decoder
NMFEncoder	This class implements an NMF encoder with a convolutional network
Psi	Convolutional Layers to estimate NMF Activations from Classifier Representations
PsiOptimized	Convolutional Layers to estimate NMF Activations from Classifier Representations, optimized for log-spectra.
Theta	This class implements a linear classifier on top of NMF activations

Functions:

weights_init

Applies Xavier initialization to network weights.

Reference

class speechbrain.lobes.models.L2I.Psi(n_comp=100, T=431, in_emb_dims=[2048, 1024, 512])

Bases: Module

Convolutional Layers to estimate NMF Activations from Classifier Representations

Parameters:

• n_comp (int) – Number of NMF components (or equivalently number of neurons at the output per timestep)

• T (int) – The targeted length along the time dimension

• in_emb_dims (List with int elements) – A list with length 3 that contains the dimensionality of the input dimensions The list needs to match the number of channels in the input classifier representations The last entry should be the smallest entry

Example

>>> inp = [torch.ones(2, 150, 6, 2), torch.ones(2, 100, 6, 2), torch.ones(2, 50, 12, 5)]
>>> psi = Psi(n_comp=100, T=120, in_emb_dims=[150, 100, 50])
>>> h = psi(inp)
>>> print(h.shape)
torch.Size([2, 100, 120])

__init__(n_comp=100, T=431, in_emb_dims=[2048, 1024, 512])

Computes NMF activations given classifier hidden representations

forward(inp)

This forward function returns the NMF time activations given classifier activations :param inp: :type inp: A length 3 list of classifier input representions.

class speechbrain.lobes.models.L2I.NMFDecoderAudio(n_comp=100, n_freq=513, device='cuda')

Bases: Module

This class implements an NMF decoder

Parameters:

• n_comp (int) – Number of NMF components

• n_freq (int) – The number of frequency bins in the NMF dictionary

• device (str) – The device to run the model

• Example

• --------

• NMFDecoderAudio(20 (>>> NMF_dec =)

• 210

• device='cpu')

• torch.rand(1 (>>> H =)

• 20

• 150)

• NMF_dec.forward(H) (>>> Xhat =)

• print(Xhat.shape) (>>>)

• torch.Size([1

• 210

• 150])

forward(H)

The forward pass for NMF given the activations H

Arguments:

Htorch.Tensor

The activations Tensor with shape B x n_comp x T

where B = Batchsize

n_comp = number of NMF components T = number of timepoints

return_W()

This function returns the NMF dictionary

speechbrain.lobes.models.L2I.weights_init(m)

Applies Xavier initialization to network weights.

Parameters:

m (nn.Module) – Module to initialize.

class speechbrain.lobes.models.L2I.PsiOptimized(dim=128, K=100, numclasses=50, use_adapter=False, adapter_reduce_dim=True)

Bases: Module

Convolutional Layers to estimate NMF Activations from Classifier Representations, optimized for log-spectra.

Parameters:

• dim (int) – Dimension of the hidden representations (input to the classifier).

• K (int) – Number of NMF components (or equivalently number of neurons at the output per timestep)

• num_classes (int) – Number of possible classes.

• use_adapter (bool) – True if you wish to learn an adapter for the latent representations.

• adapter_reduce_dim (bool) – True if the adapter should compress the latent representations.

Example

>>> inp = torch.randn(1, 256, 26, 32)
>>> psi = PsiOptimized(dim=256, K=100, use_adapter=False, adapter_reduce_dim=False)
>>> h, inp_ad= psi(inp)
>>> print(h.shape, inp_ad.shape)
torch.Size([1, 1, 417, 100]) torch.Size([1, 256, 26, 32])

__init__(dim=128, K=100, numclasses=50, use_adapter=False, adapter_reduce_dim=True)

Computes NMF activations from hidden state.

forward(hs)

Computes forward step. :param hs: Latent representations (input to the classifier). Expected shape torch.Size([B, C, H, W]). :type hs: torch.Tensor

Returns:

NMF activations and adapted representations. Shape `torch.Size([B, 1, T, 100])`.

Return type:

torch.Tensor

class speechbrain.lobes.models.L2I.Theta(n_comp=100, T=431, num_classes=50)

Bases: Module

This class implements a linear classifier on top of NMF activations

Parameters:

• n_comp (int) – Number of NMF components

• T (int) – Number of Timepoints in the NMF activations

• num_classes (int) – Number of classes that the classifier works with

• Example

• --------

• Theta(30 (>>> theta =)

• 120

• 50)

• torch.rand(1 (>>> H =)

• 30

• 120)

• theta.forward(H) (>>> c_hat =)

• print(c_hat.shape) (>>>)

• torch.Size([1

• 50])

forward(H)

We first collapse the time axis, and then pass through the linear layer

Arguments:

Htorch.Tensor

The activations Tensor with shape B x n_comp x T

where B = Batchsize

n_comp = number of NMF components T = number of timepoints

class speechbrain.lobes.models.L2I.NMFEncoder(n_freq, n_comp)

Bases: Module

This class implements an NMF encoder with a convolutional network

Parameters:

• n_freq (int) – The number of frequency bins in the NMF dictionary

• n_comp (int) – Number of NMF components

• Example

• --------

• NMFEncoder(513 (>>> nmfencoder =)

• 100)

• torch.rand(1 (>>> X =)

• 513

• 240)

• nmfencoder(X) (>>> Hhat =)

• print(Hhat.shape) (>>>)

• torch.Size([1

• 100

• 240])

forward(X)

Arguments:

Xtorch.Tensor

The input spectrogram Tensor with shape B x n_freq x T

where B = Batchsize

n_freq = nfft for the input spectrogram T = number of timepoints