def class_or_method_args():
this_args = super(Seq2Seq, Seq2Seq).class_or_method_args()
this_args.extend([
# for creating data pipelines
ModuleFlag("src_data_pipeline", DataPipeline.REGISTRY_NAME,
help="The source side data pipeline."),
ModuleFlag("trg_data_pipeline", DataPipeline.REGISTRY_NAME,
help="The target side data pipeline."),
# for preprocessing data
Flag("max_src_len", dtype=Flag.TYPE.INTEGER, default=None,
help="The maximum source length of training data."),
Flag("max_trg_len", dtype=Flag.TYPE.INTEGER, default=None,
help="The maximum target length of training data."),
Flag("truncate_src", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to truncate source to max_src_len."),
Flag("truncate_trg", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to truncate target to max_trg_len."),
# for batching dataset
Flag("batch_by_tokens", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to batch the data by word tokens."),
Flag("target_begin_of_sentence", dtype=Flag.TYPE.STRING, default="bos",
choices=["bos", "eos"],
help="The begin of sentence symbol for target side. The choice 'eos' "
"is for compatibility with fairseq transformer.")
])
return this_args
def class_or_method_args():
return [
ModuleFlag(Metric.REGISTRY_NAME,
help="The evaluation metric for the generation results."),
ModuleFlag(SequenceSearch.REGISTRY_NAME, help="The search layer for sequence generation."),
Flag("output_file", dtype=Flag.TYPE.STRING, default=None,
help="The path to a file for generated outputs. If MultipleDataset is provided, "
"it should be a dict like {dataset_name0: data_path0, ...}"),
Flag("save_metric", dtype=Flag.TYPE.STRING, default=None,
help="The path to a file that metrics will be saved to, in json format."),
]
def class_or_method_args():
""" Returns a list of args for flag definition. """
this_args = super(SpeechToText, SpeechToText).class_or_method_args()
this_args.extend([
ModuleFlag("transcript_data_pipeline",
DataPipeline.REGISTRY_NAME,
default=TranscriptDataPipeline.__name__,
help="The data pipeline for ASR transcription."),
ModuleFlag("translation_data_pipeline",
DataPipeline.REGISTRY_NAME,
default=TranscriptDataPipeline.__name__,
help="The data pipeline for translation."),
])
return this_args
def class_or_method_args():
this_args = super(SpeechToText, SpeechToText).class_or_method_args()
this_args.extend([
ModuleFlag("transcript_data_pipeline",
DataPipeline.REGISTRY_NAME,
default=TranscriptDataPipeline.__name__,
help="The target side transcript data pipeline."),
Flag("audio_feature_dim",
dtype=Flag.TYPE.INTEGER,
default=80,
help="The dimension of audio features."),
Flag("audio_feature_channels",
dtype=Flag.TYPE.INTEGER,
default=1,
help="The number of channels of audio features."),
Flag("max_src_len",
dtype=Flag.TYPE.INTEGER,
default=None,
help=
"The maximum source length of training data (audio frames)."),
Flag(
"min_src_bucket_boundary",
dtype=Flag.TYPE.INTEGER,
default=128,
help=
"The minimum source length of the training bucket (audio frames)."
),
Flag("max_trg_len",
dtype=Flag.TYPE.INTEGER,
default=None,
help="The maximum target length of training data."),
Flag("truncate_src",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help="Whether to truncate source to max_src_len."),
Flag("truncate_trg",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help="Whether to truncate target to max_trg_len."),
Flag(
"experimental_frame_transcript_ratio",
dtype=Flag.TYPE.INTEGER,
default=None,
help=
"The ratio of the number of frames and its transcript for training batch bucket."
),
Flag(
"specaug",
dtype=Flag.TYPE.STRING,
default=None,
help=
"The arguments for spec augment, can be either predefined settings "
"like LB, LD, SM, SS... or a dict containing detailed arguments."
),
Flag("disable_batch_efficiency",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help="Whether to disable the batch efficiency.")
])
return this_args
def class_or_method_args():
this_args = super(LanguageModel, LanguageModel).class_or_method_args()
this_args.extend([
# for creating data pipelines
ModuleFlag(DataPipeline.REGISTRY_NAME, help="The data pipeline."),
# for preprocessing data
Flag("max_len",
dtype=Flag.TYPE.INTEGER,
default=None,
help="The maximum length of training data."),
Flag("truncate",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help="Whether to truncate data to max_len."),
# for batching dataset
Flag("batch_by_tokens",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help="Whether to batch the data by word tokens."),
Flag(
"begin_of_sentence",
dtype=Flag.TYPE.STRING,
default="bos",
choices=["bos", "eos"],
help=
"The begin of sentence symbol for target side. The choice 'eos' "
"is for compatibility with fairseq transformer."),
Flag("gpu_efficient_level",
dtype=Flag.TYPE.INTEGER,
default=GPU_EFFICIENT_LEVEL.LEVEL0,
choices=tuple(GPU_EFFICIENT_LEVEL),
help="The efficient level for training using XLA, from 0~5."),
])
return this_args
def class_or_method_args():
this_args = super(SeqGenerationValidator,
SeqGenerationValidator).class_or_method_args()
this_args.extend([
ModuleFlag("eval_metric",
Metric.REGISTRY_NAME,
help="The metric for evaluating generation results."),
ModuleFlag("eval_search_method",
SequenceSearch.REGISTRY_NAME,
help="The search layer for sequence generation."),
Flag(
"eval_estop_patience",
dtype=Flag.TYPE.INTEGER,
default=0,
help=
"The training process will automatically shut down until the program "
"fail to acquire a better metric score anymore if `early_stop_patience` greater than 0."
),
Flag(
"eval_best_checkpoint_path",
dtype=Flag.TYPE.STRING,
default=None,
help=
"The path for checkpoints with best metric scores if provided,"
"otherwise, default \"`model_dir`_best\" will be used."),
Flag(
"eval_auto_average_checkpoints",
dtype=Flag.TYPE.BOOLEAN,
default=True,
help=
"Whether to do checkpoint average on all model weights. An extra directory for averaged "
"weights will be created. It is only available when `eval_best_checkpoint_path` is provided."
),
Flag("eval_best_avg_checkpoint_path",
dtype=Flag.TYPE.STRING,
default=None,
help="The path to saving the averaged checkpoints."),
Flag(
"eval_top_checkpoints_to_keep",
dtype=Flag.TYPE.INTEGER,
default=10,
help=
"The maximum number of checkpoints to be saved (`max_to_keep` for checkpoint manager), "
"and the number of latest checkpoints to be averaged if `eval_auto_average_checkpoints` is True. "
"If <= 0, no more checkpoints will be saved."),
])
return this_args
def class_or_method_args():
return [
ModuleFlag(Criterion.REGISTRY_NAME, help="The criterion for training or evaluation."),
ModuleFlag(OPTIMIZER_REGISTRY_NAME, help="The optimizer for training."),
ModuleFlag(LR_SCHEDULE_REGISTRY_NAME, help="The learning schedule for training."),
ModuleFlag(Validator.REGISTRY_NAME, help="The validation process while training."),
ModuleFlag(PruningSchedule.REGISTRY_NAME, help="The schedule for weight weight_pruning.",
default=PolynomialDecay.__name__),
Flag("tb_log_dir", dtype=Flag.TYPE.STRING, default=None,
help="The path to store tensorboard summary, or `model_dir`/train by default."),
Flag("train_steps", dtype=Flag.TYPE.INTEGER, default=10000000,
help="The maximum steps for training loop."),
Flag("summary_steps", dtype=Flag.TYPE.INTEGER, default=200,
help="Doing summary(logging & tensorboard) this every steps."),
Flag("save_checkpoint_steps", dtype=Flag.TYPE.INTEGER, default=1000,
help="Saving checkpoints this every steps."),
Flag("checkpoints_max_to_keep", dtype=Flag.TYPE.INTEGER, default=8,
help="The maximum checkpoints to be kept."),
Flag("initial_global_step", dtype=Flag.TYPE.INTEGER, default=None,
help="The manually specified initial global step."),
Flag("pretrain_model", dtype=Flag.TYPE.STRING, default=None, multiple=True,
help="The path to a pretrained model directory(a seq2seq model, bert model, etc.). "
"(V2) Or a json/yaml-like dict string indicating pretrained models from "
"either neurst checkpoints or publicly available models converted "
"by neurst.utils.converters. Each entry has the elements: "
"path, model_name, from_prefix, to_prefix, name_pattern. "
"Multiple pretrain models are also available."),
Flag("pretrain_variable_pattern", dtype=Flag.TYPE.STRING, default=None, multiple=True,
help="One can restore specified variables in the `pretrain_model` by this regular expression."
"Multiple pattern are also available, but must match to `pretrain_model`."),
Flag("update_cycle", dtype=Flag.TYPE.INTEGER, default=1,
help="Training step with this many batches (Gradient Accumulation)."),
Flag("clip_value", dtype=Flag.TYPE.FLOAT, default=None, help="Gradient clipping by value."),
Flag("clip_norm", dtype=Flag.TYPE.FLOAT, default=None, help="Gradient clipping by norm."),
Flag("experimental_count_batch_num", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Pre-scan the dataset for training and count the number of batches."),
Flag("freeze_variables", dtype=Flag.TYPE.STRING, default=None,
help="Variables whose names are matched with this regex will be freezed."),
Flag("pruning_variable_pattern", dtype=Flag.TYPE.STRING, default=None,
help="The regular expression that indicates the variables will be pruned."),
Flag("nopruning_variable_pattern", dtype=Flag.TYPE.STRING, default=None,
help="The regular expression that indicates the variables will NOT be pruned "
"(will take effect if `pruning_variable_pattern`=None)."),
]
def class_or_method_args():
return [
Flag("input_tarball",
dtype=Flag.TYPE.STRING,
default=None,
help="The original tarball."),
ModuleFlag(FeatureExtractor.REGISTRY_NAME,
default=None,
help="The audio feature extractor.")
]
def class_or_method_args():
return [
Flag("path", dtype=Flag.TYPE.STRING,
help="The path to multilingual datasets. "
"The record files should be stored in sub directories, which are named by src2trg, "
"e.g. rootpath/en2de, rootpath/en2fr..."),
Flag("auto_switch_langs", dtype=Flag.TYPE.BOOLEAN,
help="Whether to switch source and target langauges (which will doubled the dataset)."),
ModuleFlag(DataSampler.REGISTRY_NAME, default=TemperatureSampler.__name__,
help="The sampler for unbalanced datasets.")
]
def class_or_method_args():
this_args = super(CriterionValidator,
CriterionValidator).class_or_method_args()
this_args.extend([
ModuleFlag("eval_criterion",
Criterion.REGISTRY_NAME,
help="The criterion for validation."),
ModuleFlag("eval_dataset",
Dataset.REGISTRY_NAME,
help="The dataset for validation."),
Flag("eval_batch_size",
dtype=Flag.TYPE.INTEGER,
default=32,
help="The batch size for validation process."),
Flag("eval_task_args",
dtype=Flag.TYPE.STRING,
default=None,
help="Other parameters for building validation dataset.")
])
return this_args
def class_or_method_args():
this_args = super(Seq2Seq, Seq2Seq).class_or_method_args()
this_args.extend([
# for creating data pipelines
ModuleFlag("src_data_pipeline", DataPipeline.REGISTRY_NAME,
help="The source side data pipeline."),
ModuleFlag("trg_data_pipeline", DataPipeline.REGISTRY_NAME,
help="The target side data pipeline."),
# for preprocessing data
Flag("max_src_len", dtype=Flag.TYPE.INTEGER, default=None,
help="The maximum source length of training data."),
Flag("max_trg_len", dtype=Flag.TYPE.INTEGER, default=None,
help="The maximum target length of training data."),
Flag("truncate_src", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to truncate source to max_src_len."),
Flag("truncate_trg", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to truncate target to max_trg_len."),
# for batching dataset
Flag("batch_by_tokens", dtype=Flag.TYPE.BOOLEAN, default=None,
help="Whether to batch the data by word tokens."),
])
return this_args
def class_or_method_args():
return [
ModuleFlag(Encoder.REGISTRY_NAME, default=None, help="The encoder."),
ModuleFlag(Decoder.REGISTRY_NAME, default=None, help="The decoder."),
Flag("modality.share_source_target_embedding", dtype=Flag.TYPE.BOOLEAN, default=False,
help="Whether to share source and target embedding table."),
Flag("modality.share_embedding_and_softmax_weights", dtype=Flag.TYPE.BOOLEAN, default=False,
help="Whether to share the target embedding table and softmax weights."),
Flag("modality.dim", dtype=Flag.TYPE.INTEGER, default=None,
help="The default embedding dimension for both source and target side."),
Flag("modality.source.dim", dtype=Flag.TYPE.INTEGER, default=None,
help="The source-side embedding dimension, or `modality.dim` if not provided."),
Flag("modality.target.dim", dtype=Flag.TYPE.INTEGER, default=None,
help="The target-side embedding dimension, or `modality.dim` if not provided."),
Flag("modality.timing", dtype=Flag.TYPE.STRING, default=None,
help="The arbitrary parameters for positional encoding of both source and target side."),
Flag("modality.source.timing", dtype=Flag.TYPE.STRING, default=None,
help="The arbitrary parameters for source-side positional encoding, "
"or `modality.timing` if not provided."),
Flag("modality.target.timing", dtype=Flag.TYPE.STRING, default=None,
help="The arbitrary parameters for target-side positional encoding, "
"or `modality.timing` if not provided.")
]
def class_or_method_args():
return [
Flag("input_tarball",
dtype=Flag.TYPE.STRING,
default=None,
help="The original tarball."),
Flag("excluded_file",
dtype=Flag.TYPE.STRING,
default=None,
help="A file containing transcriptions or translations "
"that would be removed when reading the corpus "
"(for filtering out testsets)."),
ModuleFlag(FeatureExtractor.REGISTRY_NAME,
default=None,
help="The audio feature extractor.")
]
def class_or_method_args():
return [
Flag(
"data_files",
dtype=Flag.TYPE.STRING,
help="A dict of data files. The key is the dataset name while "
"the value is a dict containing arguments indicating data files."
),
Flag("data_class",
dtype=Flag.TYPE.STRING,
help="The dataset class for the data files."),
Flag("common_properties",
dtype=Flag.TYPE.STRING,
default=None,
help="Other common properties for building a dataset."),
ModuleFlag(DataSampler.REGISTRY_NAME,
default=None,
help="The sampler for unbalanced datasets.")
]
def class_or_method_args():
return [
Flag(
"tb_log_dir",
dtype=Flag.TYPE.STRING,
default=None,
help=
"The path to store tensorboard summary, or `model_dir`/validation by default."
),
Flag("waiting_interval",
dtype=Flag.TYPE.INTEGER,
default=120,
help="The waiting interval between two evaluation steps."),
Flag("maximum_waiting_time",
dtype=Flag.TYPE.INTEGER,
default=3600,
help="The maximum waiting time(in seconds)."),
ModuleFlag(Validator.REGISTRY_NAME,
help="The validation process during training."),
]
def class_or_method_args():
return [
Flag(
"data_files",
dtype=Flag.TYPE.STRING,
help=
"A dict of parallel data files. The key is the dataset name while "
"the value is a dict containing `src_file` and `trg_file`."),
Flag("data_is_processed",
dtype=Flag.TYPE.BOOLEAN,
help="Whether the text data is already processed."),
Flag("src_lang",
dtype=Flag.TYPE.STRING,
default=None,
help="The source language"),
Flag("trg_lang",
dtype=Flag.TYPE.STRING,
default=None,
help="The target language"),
ModuleFlag(DataSampler.REGISTRY_NAME,
default=None,
help="The sampler for unbalanced datasets.")
]
def class_or_method_args():
return [
ModuleFlag(Criterion.REGISTRY_NAME,
help="The criterion for evaluation."),
]
def class_or_method_args():
return [
ModuleFlag(Criterion.REGISTRY_NAME,
help="The criterion for training or evaluation."),
ModuleFlag(OPTIMIZER_REGISTRY_NAME,
help="The optimizer for training."),
ModuleFlag(LR_SCHEDULE_REGISTRY_NAME,
help="The learning schedule for training."),
ModuleFlag(Validator.REGISTRY_NAME,
help="The validation process while training."),
Flag(
"tb_log_dir",
dtype=Flag.TYPE.STRING,
default=None,
help=
"The path to store tensorboard summary, or `model_dir`/train by default."
),
Flag("train_steps",
dtype=Flag.TYPE.INTEGER,
default=10000000,
help="The maximum steps for training loop."),
Flag(
"summary_steps",
dtype=Flag.TYPE.INTEGER,
default=200,
help="Doing summary(logging & tensorboard) this every steps."),
Flag("save_checkpoint_steps",
dtype=Flag.TYPE.INTEGER,
default=1000,
help="Saving checkpoints this every steps."),
Flag("checkpoints_max_to_keep",
dtype=Flag.TYPE.INTEGER,
default=8,
help="The maximum checkpoints to be kept."),
Flag("initial_global_step",
dtype=Flag.TYPE.INTEGER,
default=None,
help="The manually specified initial global step."),
Flag(
"pretrain_model",
dtype=Flag.TYPE.STRING,
default=None,
multiple=True,
help=
"The path to a pretrained model directory(a seq2seq model, bert model, etc.). "
"Multiple pretrain models are also available."),
Flag(
"pretrain_variable_pattern",
dtype=Flag.TYPE.STRING,
default=None,
multiple=True,
help=
"One can restore specified variables in the `pretrain_model` by this regular expression."
"Multiple pattern are also available, but must match to `pretrain_model`."
),
Flag(
"update_cycle",
dtype=Flag.TYPE.INTEGER,
default=1,
help=
"Training step with this many batches (Gradient Accumulation)."
),
Flag("clip_value",
dtype=Flag.TYPE.FLOAT,
default=None,
help="Gradient clipping by value."),
Flag("clip_norm",
dtype=Flag.TYPE.FLOAT,
default=None,
help="Gradient clipping by norm."),
Flag(
"experimental_count_batch_num",
dtype=Flag.TYPE.BOOLEAN,
default=None,
help=
"Pre-scan the dataset for training and count the number of batches."
)
]