espnet2.uasr package

espnet2.uasr.__init__

espnet2.uasr.espnet_model

class espnet2.uasr.espnet_model.ESPnetUASRModel(frontend: Optional[espnet2.asr.frontend.abs_frontend.AbsFrontend], segmenter: Optional[espnet2.uasr.segmenter.abs_segmenter.AbsSegmenter], generator: espnet2.uasr.generator.abs_generator.AbsGenerator, discriminator: espnet2.uasr.discriminator.abs_discriminator.AbsDiscriminator, losses: Dict[str, espnet2.uasr.loss.abs_loss.AbsUASRLoss], kenlm_path: Optional[str], token_list: Optional[list], max_epoch: Optional[int], vocab_size: int, cfg: Optional[Dict] = None, pad: int = 1, sil_token: str = '<SIL>', sos_token: str = '<s>', eos_token: str = '</s>', skip_softmax: espnet2.utils.types.str2bool = False, use_gumbel: espnet2.utils.types.str2bool = False, use_hard_gumbel: espnet2.utils.types.str2bool = True, min_temperature: float = 0.1, max_temperature: float = 2.0, decay_temperature: float = 0.99995, use_collected_training_feats: espnet2.utils.types.str2bool = False)

Bases: espnet2.train.abs_espnet_model.AbsESPnetModel

Unsupervised ASR model.

The source code is from FAIRSEQ: https://github.com/facebookresearch/fairseq/tree/main/examples/wav2vec/unsupervised

collect_feats(speech: torch.Tensor, speech_lengths: torch.Tensor, text: Optional[torch.Tensor] = None, text_lengths: Optional[torch.Tensor] = None, **kwargs) → Dict[str, torch.Tensor]

encode(speech: torch.Tensor, speech_lengths: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor]

forward(speech: torch.Tensor, speech_lengths: torch.Tensor, text: Optional[torch.Tensor] = None, text_lengths: Optional[torch.Tensor] = None, pseudo_labels: Optional[torch.Tensor] = None, pseudo_labels_lengths: Optional[torch.Tensor] = None, do_validation: Optional[espnet2.utils.types.str2bool] = False, print_hyp: Optional[espnet2.utils.types.str2bool] = False, **kwargs) → Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]

Frontend + Segmenter + Generator + Discriminator + Calc Loss

Args:

get_optim_index()

inference(speech: torch.Tensor, speech_lengths: torch.Tensor)

is_discriminative_step()

property number_updates

espnet2.uasr.generator.__init__

espnet2.uasr.generator.abs_generator

class espnet2.uasr.generator.abs_generator.AbsGenerator(*args, **kwargs)

Bases: torch.nn.modules.module.Module, abc.ABC

Initializes internal Module state, shared by both nn.Module and ScriptModule.

abstract forward(xs_pad: torch.Tensor, ilens: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

abstract output_size() → int

espnet2.uasr.generator.conv_generator

class espnet2.uasr.generator.conv_generator.ConvGenerator(input_dim: int, output_dim: int, cfg: Optional[Dict] = None, conv_kernel: int = 3, conv_dilation: int = 1, conv_stride: int = 9, pad: int = -1, bias: espnet2.utils.types.str2bool = False, dropout: float = 0.0, batch_norm: espnet2.utils.types.str2bool = True, batch_norm_weight: float = 30.0, residual: espnet2.utils.types.str2bool = True)

Bases: espnet2.uasr.generator.abs_generator.AbsGenerator

convolutional generator for UASR.

bn_padded_data(feature: torch.Tensor, padding_mask: torch.Tensor)

forward(feats: torch.Tensor, text: Optional[torch.Tensor], feats_padding_mask: torch.Tensor)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

output_size()

class espnet2.uasr.generator.conv_generator.SamePad(kernel_size, causal=False)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class espnet2.uasr.generator.conv_generator.TransposeLast(deconstruct_idx=None)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

espnet2.uasr.segmenter.join_segmenter

class espnet2.uasr.segmenter.join_segmenter.JoinSegmenter(cfg: Optional[Dict] = None, subsample_rate: float = 0.25, mean_pool: espnet2.utils.types.str2bool = True, mean_join_pool: espnet2.utils.types.str2bool = False, remove_zeros: espnet2.utils.types.str2bool = False)

Bases: espnet2.uasr.segmenter.abs_segmenter.AbsSegmenter

logit_segment(logits: torch.Tensor, padding_mask: torch.Tensor) → torch.Tensor

pre_segment(xs_pad: torch.Tensor, padding_mask: torch.Tensor) → torch.Tensor

espnet2.uasr.segmenter.__init__

espnet2.uasr.segmenter.random_segmenter

class espnet2.uasr.segmenter.random_segmenter.RandomSegmenter(subsample_rate: float = 0.25, mean_pool: espnet2.utils.types.str2bool = True, mean_join_pool: espnet2.utils.types.str2bool = False, remove_zeros: espnet2.utils.types.str2bool = False)

Bases: espnet2.uasr.segmenter.abs_segmenter.AbsSegmenter

logit_segment(xs_pad: torch.Tensor, padding_mask: torch.Tensor) → torch.Tensor

pre_segment(xs_pad: torch.Tensor, padding_mask: torch.Tensor) → torch.Tensor

espnet2.uasr.segmenter.abs_segmenter

Segmenter definition for UASR task

Practially, the output of the generator (in frame-level) may predict the same phoneme for consecutive frames, which makes it too easy for the discriminator. So, the segmenter here is to merge frames with a similar prediction from the generator output.

class espnet2.uasr.segmenter.abs_segmenter.AbsSegmenter(*args, **kwargs)

Bases: torch.nn.modules.module.Module, abc.ABC

Initializes internal Module state, shared by both nn.Module and ScriptModule.

abstract logit_segment(xs_pad: torch.Tensor, ilens: torch.Tensor) → torch.Tensor

abstract pre_segment(xs_pad: torch.Tensor, ilens: torch.Tensor) → torch.Tensor

espnet2.uasr.discriminator.__init__

espnet2.uasr.discriminator.abs_discriminator

class espnet2.uasr.discriminator.abs_discriminator.AbsDiscriminator(*args, **kwargs)

Bases: torch.nn.modules.module.Module, abc.ABC

Initializes internal Module state, shared by both nn.Module and ScriptModule.

abstract forward(xs_pad: torch.Tensor, padding_mask: torch.Tensor) → torch.Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

espnet2.uasr.discriminator.conv_discriminator

class espnet2.uasr.discriminator.conv_discriminator.ConvDiscriminator(input_dim: int, cfg: Optional[Dict] = None, conv_channels: int = 384, conv_kernel: int = 8, conv_dilation: int = 1, conv_depth: int = 2, linear_emb: espnet2.utils.types.str2bool = False, causal: espnet2.utils.types.str2bool = True, max_pool: espnet2.utils.types.str2bool = False, act_after_linear: espnet2.utils.types.str2bool = False, dropout: float = 0.0, spectral_norm: espnet2.utils.types.str2bool = False, weight_norm: espnet2.utils.types.str2bool = False)

Bases: espnet2.uasr.discriminator.abs_discriminator.AbsDiscriminator

convolutional discriminator for UASR.

forward(x: torch.Tensor, padding_mask: Optional[torch.Tensor])

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class espnet2.uasr.discriminator.conv_discriminator.SamePad(kernel_size, causal=False)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

espnet2.uasr.loss.smoothness_penalty

class espnet2.uasr.loss.smoothness_penalty.UASRSmoothnessPenalty(weight: float = 1.0, reduction: str = 'none')

Bases: espnet2.uasr.loss.abs_loss.AbsUASRLoss

smoothness penalty for UASR.

forward(dense_logits: torch.Tensor, dense_padding_mask: torch.Tensor, sample_size: int, is_discriminative_step: bool)

Forward.

Parameters:

• dense_logits – output logits of generator

• dense_padding_mask – padding mask of logits

• sample_size – batch size

• is_discriminative_step – Whether is training discriminator

espnet2.uasr.loss.__init__

espnet2.uasr.loss.abs_loss

class espnet2.uasr.loss.abs_loss.AbsUASRLoss(*args, **kwargs)

Bases: torch.nn.modules.module.Module, abc.ABC

Base class for all Diarization loss modules.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

abstract forward() → torch.Tensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

property name

espnet2.uasr.loss.discriminator_loss

class espnet2.uasr.loss.discriminator_loss.UASRDiscriminatorLoss(weight: float = 1.0, smoothing: float = 0.0, smoothing_one_side: espnet2.utils.types.str2bool = False, reduction: str = 'sum')

Bases: espnet2.uasr.loss.abs_loss.AbsUASRLoss

discriminator loss for UASR.

forward(dense_y: torch.Tensor, token_y: torch.Tensor, is_discriminative_step: espnet2.utils.types.str2bool)

Forward.

Parameters:

• dense_y – predicted logits of generated samples

• token_y – predicted logits of real samples

espnet2.uasr.loss.pseudo_label_loss

class espnet2.uasr.loss.pseudo_label_loss.UASRPseudoLabelLoss(weight: float = 1.0, input_dim: int = 128, output_dim: int = 64, downsample_rate: int = 2, ignore_index: int = -1, reduction: str = 'none')

Bases: espnet2.uasr.loss.abs_loss.AbsUASRLoss

auxiliary pseudo label loss for UASR.

forward(inter_x: torch.Tensor, pseudo_labels: torch.Tensor, is_discriminative_step: espnet2.utils.types.str2bool)

Forward.

Args:

espnet2.uasr.loss.phoneme_diversity_loss

class espnet2.uasr.loss.phoneme_diversity_loss.UASRPhonemeDiversityLoss(weight: float = 1.0)

Bases: espnet2.uasr.loss.abs_loss.AbsUASRLoss

phoneme diversity loss for UASR.

forward(dense_x: torch.Tensor, sample_size: int, is_discriminative_step: espnet2.utils.types.str2bool)

Forward.

Parameters:

• dense_x – predicted logits of generated samples

• sample_size – batch size

• is_dicriminative_step – whether is training discriminator

espnet2.uasr.loss.gradient_penalty

class espnet2.uasr.loss.gradient_penalty.UASRGradientPenalty(discriminator: espnet2.uasr.discriminator.abs_discriminator.AbsDiscriminator, weight: float = 1.0, probabilistic_grad_penalty_slicing: espnet2.utils.types.str2bool = False, reduction: str = 'sum')

Bases: espnet2.uasr.loss.abs_loss.AbsUASRLoss

gradient penalty for UASR.

forward(fake_sample: torch.Tensor, real_sample: torch.Tensor, is_training: espnet2.utils.types.str2bool, is_discrimininative_step: espnet2.utils.types.str2bool)

Forward.

Parameters:

• fake_sample – generated sample from generator

• real_sample – real sample

• is_training – whether is at training step

• is_discriminative_step – whether is training discriminator