class NLUFinetuningConfig(AudioPretrainingConfig):
# Options for reporting WER metrics during validation. Only applicable to
# Seq2Seq models during fine-tuning
eval_wer: bool = field(
default=False, metadata={"help": "compute WER for Seq2Seq models"}
)
eval_wer_parse: bool = field(
default=False, metadata={"help": "compute WER for Seq2Seq models"}
)
eval_wer_config: GenerationConfig = field(
default_factory=lambda: GenerationConfig(),
metadata={"help": "beam search config for evaluating wer during training"},
)
eval_wer_tokenizer: Any = field(
default=None,
metadata={"help": "tokenizer config for evaluating wer during training"},
)
eval_wer_post_process: str = field(
default="letter",
metadata={
"help": "remove BPE tokens before scoring (can be sentencepiece, letter, and more)"
},
)
eval_bleu: bool = field(
default=False, metadata={"help": "evaluation with BLEU scores"}
)
eval_bleu_detok: Optional[str] = field(
default=None,
metadata={
"help": "detokenize before computing BLEU (e.g., 'moses'); "
"required if using --eval-bleu; use 'space' to disable "
"detokenization; see fairseq.data.encoders for other options"
},
)
eval_bleu_detok_args: str = field(
default="{}", metadata={"help": "args for building the tokenizer, if needed"}
)
eval_tokenized_bleu: bool = field(
default=False, metadata={"help": "compute tokenized BLEU instead of sacrebleu"}
)
eval_bleu_remove_bpe: Optional[str] = field(
default=None, metadata={"help": "remove BPE before computing BLEU"}
)
eval_bleu_args: str = field(
default="{}",
metadata={
"help": "generation args for BLUE scoring, e.g., "
'\'{"beam": 4, "lenpen": 0.6}\''
},
)
eval_bleu_print_samples: bool = field(
default=False, metadata={"help": "print sample generations during validation"}
)
autoregressive: bool = field(
default=False,
metadata={
"help": "required for autoregressive decoders (like seq2seq models); "
"adds 'prev_output_tokens' to input and appends eos to target"
},
)
class InferConfig(FairseqDataclass):
task: Any = None
decoding: DecodingConfig = DecodingConfig()
common: CommonConfig = CommonConfig()
common_eval: CommonEvalConfig = CommonEvalConfig()
checkpoint: CheckpointConfig = CheckpointConfig()
generation: GenerationConfig = GenerationConfig()
distributed_training: DistributedTrainingConfig = DistributedTrainingConfig(
)
dataset: DatasetConfig = DatasetConfig()
def build_generator(self, models, cfg: GenerationConfig):
if cfg.score_reference:
cfg.score_reference = False
logger.warning(
"--score-reference is not applicable to speech recognition, ignoring it."
)
from espresso.tools.generate_log_probs_for_decoding import GenerateLogProbsForDecoding
apply_log_softmax = getattr(cfg, "apply_log_softmax", False)
return GenerateLogProbsForDecoding(models,
apply_log_softmax=apply_log_softmax)
class AudioPretrainingConfig(FairseqDataclass):
data: str = field(default=MISSING, metadata={"help": "path to data directory"})
labels: Optional[str] = field(
default=None,
metadata={"help": "extension of the label file to load, used for fine-tuning"},
)
sample_rate: int = field(
default=16_000,
metadata={
"help": "target sample rate. audio files will be up/down sampled to this rate"
},
)
normalize: bool = field(
default=False,
metadata={"help": "if set, normalizes input to have 0 mean and unit variance"},
)
enable_padding: bool = field(
default=False, metadata={"help": "pad shorter samples instead of cropping"}
)
max_sample_size: Optional[int] = field(
default=None, metadata={"help": "max sample size to crop to for batching"}
)
min_sample_size: Optional[int] = field(
default=None, metadata={"help": "min sample size to skip small examples"}
)
# Options for reporting WER metrics during validation. Only applicable to
# Seq2Seq models during fine-tuning
eval_wer: bool = field(
default=False, metadata={"help": "compute WER for Seq2Seq models"}
)
eval_wer_config: GenerationConfig = field(
default_factory=lambda: GenerationConfig(),
metadata={"help": "beam search config for evaluating wer during training"},
)
eval_wer_tokenizer: Any = field(
default=None,
metadata={"help": "tokenizer config for evaluating wer during training"},
)
eval_wer_post_process: str = field(
default="letter",
metadata={
"help": "remove BPE tokens before scoring (can be sentencepiece, letter, and more)"
},
)
autoregressive: bool = field(
default=False,
metadata={
"help": "required for autoregressive decoders (like seq2seq models); "
"adds 'prev_output_tokens' to input and appends eos to target"
},
)
def add_generation_args(parser):
group = parser.add_argument_group("Generation")
add_common_eval_args(group)
gen_parser_from_dataclass(group, GenerationConfig())
return group
class AudioPretrainingConfig(FairseqDataclass):
data: str = field(default=MISSING, metadata={"help": "path to data directory"})
label_dir: Optional[str] = field(default=None, metadata={"help": "path to label directory"})
labels: Optional[str] = field(
default=None,
metadata={"help": "extension of the label file to load, used for fine-tuning"},
)
binarized_dataset: bool = field(
default=False,
metadata={
"help": "if true, loads binarized dataset (useful for very large datasets). "
"See examples/wav2vec/scripts/binarize_manifest.sh"
},
)
sample_rate: int = field(
default=16_000,
metadata={
"help": "target sample rate. audio files will be up/down sampled to this rate"
},
)
normalize: bool = field(
default=False,
metadata={"help": "if set, normalizes input to have 0 mean and unit variance"},
)
enable_padding: bool = field(
default=False, metadata={"help": "pad shorter samples instead of cropping"}
)
max_sample_size: Optional[int] = field(
default=None, metadata={"help": "max sample size to crop to for batching"}
)
min_sample_size: Optional[int] = field(
default=None, metadata={"help": "min sample size to skip small examples"}
)
# Options for reporting WER metrics during validation. Only applicable to
# Seq2Seq models during fine-tuning
eval_wer: bool = field(
default=False, metadata={"help": "compute WER for Seq2Seq models"}
)
eval_wer_config: GenerationConfig = field(
default_factory=lambda: GenerationConfig(),
metadata={"help": "beam search config for evaluating wer during training"},
)
eval_wer_tokenizer: Any = field(
default=None,
metadata={"help": "tokenizer config for evaluating wer during training"},
)
eval_wer_post_process: str = field(
default="letter",
metadata={
"help": "remove BPE tokens before scoring (can be sentencepiece, letter, and more)"
},
)
autoregressive: bool = field(
default=False,
metadata={
"help": "required for autoregressive decoders (like seq2seq models); "
"adds 'prev_output_tokens' to input and appends eos to target"
},
)
num_batch_buckets: int = field(
default=0,
metadata={"help": "number of buckets"},
)
precompute_mask_indices: bool = field(
default=False,
metadata={
"help": "flag to compute mask indices in data preparation.",
},
)
# The following are needed to precompute mask and mask channel indices
# before model's forward.
# mask_length: Optional[int] = II("model.mask_length")
# mask_prob: Optional[float] = II("model.mask_prob")
# mask_selection: Optional[str] = II("model.mask_selection")
# mask_other: Optional[float] = II("model.mask_other")
# no_mask_overlap: Optional[bool] = II("model.no_mask_overlap")
# mask_min_space: Optional[int] = II("model.mask_min_space")
# mask_channel_length: Optional[int] = II("model.mask_channel_length")
# mask_channel_prob: Optional[float] = II("model.mask_channel_prob")
# mask_channel_selection: Optional[str] = II("model.mask_channel_selection")
# mask_channel_other: Optional[float] = II("model.mask_channel_other")
# no_mask_channel_overlap: Optional[bool] = II("model.no_mask_channel_overlap")
# mask_channel_min_space: Optional[int] = II("model.mask_channel_min_space")
# conv_feature_layers: Optional[str] = II("model.conv_feature_layers")
# encoder_embed_dim: Optional[int] = II("model.encoder_embed_dim")
tpu: bool = II("common.tpu")
class AudioPretrainingConfig(FairseqDataclass):
data: str = field(default=MISSING,
metadata={"help": "path to data directory"})
labels: Optional[str] = field(
default=None,
metadata={
"help": "extension of the label file to load, used for fine-tuning"
},
)
binarized_dataset: bool = field(
default=False,
metadata={
"help":
"if true, loads binarized dataset (useful for very large datasets). "
"See examples/wav2vec/scripts/binarize_manifest.sh"
},
)
sample_rate: int = field(
default=16_000,
metadata={
"help":
"target sample rate. audio files will be up/down sampled to this rate"
},
)
normalize: bool = field(
default=False,
metadata={
"help": "if set, normalizes input to have 0 mean and unit variance"
},
)
enable_padding: bool = field(
default=False,
metadata={"help": "pad shorter samples instead of cropping"})
max_sample_size: Optional[int] = field(
default=None,
metadata={"help": "max sample size to crop to for batching"})
min_sample_size: Optional[int] = field(
default=None,
metadata={"help": "min sample size to skip small examples"})
dataset_sampling_alpha: Optional[float] = field(
default=0.5,
metadata={
"help": "smoothing alpha for sample rations across datasets"
})
# Options for reporting WER metrics during validation. Only applicable to
# Seq2Seq models during fine-tuning
eval_wer: bool = field(default=False,
metadata={"help": "compute WER for Seq2Seq models"})
eval_wer_config: GenerationConfig = field(
default_factory=lambda: GenerationConfig(),
metadata={
"help": "beam search config for evaluating wer during training"
},
)
eval_wer_tokenizer: Any = field(
default=None,
metadata={
"help": "tokenizer config for evaluating wer during training"
},
)
eval_wer_post_process: str = field(
default="letter",
metadata={
"help":
"remove BPE tokens before scoring (can be sentencepiece, letter, and more)"
},
)
autoregressive: bool = field(
default=False,
metadata={
"help":
"required for autoregressive decoders (like seq2seq models); "
"adds 'prev_output_tokens' to input and appends eos to target"
},
)
num_batch_buckets: int = field(
default=0,
metadata={"help": "number of buckets"},
)
precompute_mask_indices: bool = field(
default=False,
metadata={
"help": "flag to compute mask indices in data preparation.",
},
)
inferred_w2v_config: Optional[InferredW2vConfig] = field(
default=None,
metadata={
"help":
"wav2vec 2.0 masking arguments used to pre-compute masks (required for TPU)",
},
)
tpu: bool = II("common.tpu")