Openspeech’s configurations¶
This page describes all configurations in Openspeech
.
audio
¶
mfcc
¶
name
: Name of dataset.sample_rate
: Sampling rate of audioframe_length
: Frame length for spectrogramframe_shift
: Length of hop between STFTdel_silence
: Flag indication whether to apply delete silence or notnum_mels
: The number of mfc coefficients to retain.apply_spec_augment
: Flag indication whether to apply spec augment or notapply_noise_augment
: Flag indication whether to apply noise augment or notapply_time_stretch_augment
: Flag indication whether to apply time stretch augment or notapply_joining_augment
: Flag indication whether to apply audio joining augment or not
melspectrogram
¶
name
: Name of dataset.sample_rate
: Sampling rate of audioframe_length
: Frame length for spectrogramframe_shift
: Length of hop between STFTdel_silence
: Flag indication whether to apply delete silence or notnum_mels
: The number of mfc coefficients to retain.apply_spec_augment
: Flag indication whether to apply spec augment or notapply_noise_augment
: Flag indication whether to apply noise augment or notapply_time_stretch_augment
: Flag indication whether to apply time stretch augment or notapply_joining_augment
: Flag indication whether to apply audio joining augment or not
fbank
¶
name
: Name of dataset.sample_rate
: Sampling rate of audioframe_length
: Frame length for spectrogramframe_shift
: Length of hop between STFTdel_silence
: Flag indication whether to apply delete silence or notnum_mels
: The number of mfc coefficients to retain.apply_spec_augment
: Flag indication whether to apply spec augment or notapply_noise_augment
: Flag indication whether to apply noise augment or notapply_time_stretch_augment
: Flag indication whether to apply time stretch augment or notapply_joining_augment
: Flag indication whether to apply audio joining augment or not
spectrogram
¶
name
: Name of dataset.sample_rate
: Sampling rate of audioframe_length
: Frame length for spectrogramframe_shift
: Length of hop between STFTdel_silence
: Flag indication whether to apply delete silence or notnum_mels
: Spectrogram is independent of mel, but uses the ‘num_mels’ variable to unify feature size variablesapply_spec_augment
: Flag indication whether to apply spec augment or notapply_noise_augment
: Flag indication whether to apply noise augment or notapply_time_stretch_augment
: Flag indication whether to apply time stretch augment or notapply_joining_augment
: Flag indication whether to apply audio joining augment or not
augment
¶
default
¶
apply_spec_augment
: Flag indication whether to apply spec augment or notapply_noise_augment
: Flag indication whether to apply noise augment or not Noise augment requiresnoise_dataset_path
.noise_dataset_dir
should be contain audio files.apply_joining_augment
: Flag indication whether to apply joining augment or not If true, create a new audio file by connecting two audio randomlyapply_time_stretch_augment
: Flag indication whether to apply spec augment or notfreq_mask_para
: Hyper Parameter for freq masking to limit freq masking lengthfreq_mask_num
: How many freq-masked area to maketime_mask_num
: How many time-masked area to makenoise_dataset_dir
: How many time-masked area to makenoise_level
: Noise adjustment leveltime_stretch_min_rate
: Minimum rate of audio time stretchtime_stretch_max_rate
: Maximum rate of audio time stretch
dataset
¶
kspon
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasettest_dataset_path
: Path of evaluation datasetmanifest_file_path
: Path of manifest filetest_manifest_dir
: Path of directory contains test manifest filespreprocess_mode
: KsponSpeech preprocess mode {phonetic, spelling}
libri
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasetdataset_download
: Flag indication whether to download dataset or not.manifest_file_path
: Path of manifest file
aishell
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasetdataset_download
: Flag indication whether to download dataset or not.manifest_file_path
: Path of manifest file
ksponspeech
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasettest_dataset_path
: Path of evaluation datasetmanifest_file_path
: Path of manifest filetest_manifest_dir
: Path of directory contains test manifest filespreprocess_mode
: KsponSpeech preprocess mode {phonetic, spelling}
librispeech
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasetdataset_download
: Flag indication whether to download dataset or not.manifest_file_path
: Path of manifest file
lm
¶
dataset
: Select dataset for training (librispeech, ksponspeech, aishell, lm)dataset_path
: Path of datasetvalid_ratio
: Ratio of validation datatest_ratio
: Ratio of test data
model
¶
listen_attend_spell
¶
model_name
: Model namenum_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.hidden_state_dim
: The hidden state dimension of encoder.encoder_dropout_p
: The dropout probability of encoder.encoder_bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)joint_ctc_attention
: Flag indication joint ctc attention or notmax_length
: Max decoding length.num_attention_heads
: The number of attention heads.decoder_dropout_p
: The dropout probability of decoder.decoder_attn_mechanism
: The attention mechanism for decoder.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
listen_attend_spell_with_location_aware
¶
model_name
: Model namenum_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.hidden_state_dim
: The hidden state dimension of encoder.encoder_dropout_p
: The dropout probability of encoder.encoder_bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)joint_ctc_attention
: Flag indication joint ctc attention or notmax_length
: Max decoding length.num_attention_heads
: The number of attention heads.decoder_dropout_p
: The dropout probability of decoder.decoder_attn_mechanism
: The attention mechanism for decoder.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
listen_attend_spell_with_multi_head
¶
model_name
: Model namenum_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.hidden_state_dim
: The hidden state dimension of encoder.encoder_dropout_p
: The dropout probability of encoder.encoder_bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)joint_ctc_attention
: Flag indication joint ctc attention or notmax_length
: Max decoding length.num_attention_heads
: The number of attention heads.decoder_dropout_p
: The dropout probability of decoder.decoder_attn_mechanism
: The attention mechanism for decoder.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
joint_ctc_listen_attend_spell
¶
model_name
: Model namenum_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.hidden_state_dim
: The hidden state dimension of encoder.encoder_dropout_p
: The dropout probability of encoder.encoder_bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)joint_ctc_attention
: Flag indication joint ctc attention or notmax_length
: Max decoding length.num_attention_heads
: The number of attention heads.decoder_dropout_p
: The dropout probability of decoder.decoder_attn_mechanism
: The attention mechanism for decoder.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
deep_cnn_with_joint_ctc_listen_attend_spell
¶
model_name
: Model namenum_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.hidden_state_dim
: The hidden state dimension of encoder.encoder_dropout_p
: The dropout probability of encoder.encoder_bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)extractor
: The CNN feature extractor.activation
: Type of activation functionjoint_ctc_attention
: Flag indication joint ctc attention or notmax_length
: Max decoding length.num_attention_heads
: The number of attention heads.decoder_dropout_p
: The dropout probability of decoder.decoder_attn_mechanism
: The attention mechanism for decoder.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
deepspeech2
¶
model_name
: Model namernn_type
: Type of rnn cell (rnn, lstm, gru)num_rnn_layers
: The number of rnn layersrnn_hidden_dim
: Hidden state dimenstion of RNN.dropout_p
: The dropout probability of model.bidirectional
: If True, becomes a bidirectional encodersactivation
: Type of activation functionoptimizer
: Optimizer for training.
lstm_lm
¶
model_name
: Model namenum_layers
: The number of encoder layers.hidden_state_dim
: The hidden state dimension of encoder.dropout_p
: The dropout probability of encoder.rnn_type
: Type of rnn cell (rnn, lstm, gru)max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.optimizer
: Optimizer for training.
rnn_transducer
¶
model_name
: Model nameencoder_hidden_state_dim
: Dimension of encoder.decoder_hidden_state_dim
: Dimension of decoder.num_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.encoder_dropout_p
: The dropout probability of encoder.decoder_dropout_p
: The dropout probability of decoder.bidirectional
: If True, becomes a bidirectional encodersrnn_type
: Type of rnn cell (rnn, lstm, gru)output_dim
: Dimension of outputsoptimizer
: Optimizer for training.
transformer_lm
¶
model_name
: Model namenum_layers
: The number of encoder layers.d_model
: The dimension of model.d_ff
: The dimenstion of feed forward network.num_attention_heads
: The number of attention heads.dropout_p
: The dropout probability of encoder.max_length
: Max decoding length.optimizer
: Optimizer for training.
transformer
¶
model_name
: Model named_model
: Dimension of model.d_ff
: Dimenstion of feed forward network.num_attention_heads
: The number of attention heads.num_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.encoder_dropout_p
: The dropout probability of encoder.decoder_dropout_p
: The dropout probability of decoder.ffnet_style
: Style of feed forward network. (ff, conv)max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.joint_ctc_attention
: Flag indication joint ctc attention or notoptimizer
: Optimizer for training.
joint_ctc_transformer
¶
model_name
: Model nameextractor
: The CNN feature extractor.d_model
: Dimension of model.d_ff
: Dimenstion of feed forward network.num_attention_heads
: The number of attention heads.num_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.encoder_dropout_p
: The dropout probability of encoder.decoder_dropout_p
: The dropout probability of decoder.ffnet_style
: Style of feed forward network. (ff, conv)max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.joint_ctc_attention
: Flag indication joint ctc attention or notoptimizer
: Optimizer for training.
transformer_with_ctc
¶
model_name
: Model named_model
: Dimension of model.d_ff
: Dimenstion of feed forward network.num_attention_heads
: The number of attention heads.num_encoder_layers
: The number of encoder layers.encoder_dropout_p
: The dropout probability of encoder.ffnet_style
: Style of feed forward network. (ff, conv)optimizer
: Optimizer for training.
vgg_transformer
¶
model_name
: Model nameextractor
: The CNN feature extractor.d_model
: Dimension of model.d_ff
: Dimenstion of feed forward network.num_attention_heads
: The number of attention heads.num_encoder_layers
: The number of encoder layers.num_decoder_layers
: The number of decoder layers.encoder_dropout_p
: The dropout probability of encoder.decoder_dropout_p
: The dropout probability of decoder.ffnet_style
: Style of feed forward network. (ff, conv)max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.joint_ctc_attention
: Flag indication joint ctc attention or notoptimizer
: Optimizer for training.
conformer
¶
model_name
: Model nameencoder_dim
: Dimension of encoder.num_encoder_layers
: The number of encoder layers.num_attention_heads
: The number of attention heads.feed_forward_expansion_factor
: The expansion factor of feed forward module.conv_expansion_factor
: The expansion factor of convolution module.input_dropout_p
: The dropout probability of inputs.feed_forward_dropout_p
: The dropout probability of feed forward module.attention_dropout_p
: The dropout probability of attention module.conv_dropout_p
: The dropout probability of convolution module.conv_kernel_size
: The kernel size of convolution.half_step_residual
: Flag indication whether to use half step residual or notoptimizer
: Optimizer for training.
conformer_lstm
¶
model_name
: Model nameencoder_dim
: Dimension of encoder.num_encoder_layers
: The number of encoder layers.num_attention_heads
: The number of attention heads.feed_forward_expansion_factor
: The expansion factor of feed forward module.conv_expansion_factor
: The expansion factor of convolution module.input_dropout_p
: The dropout probability of inputs.feed_forward_dropout_p
: The dropout probability of feed forward module.attention_dropout_p
: The dropout probability of attention module.conv_dropout_p
: The dropout probability of convolution module.conv_kernel_size
: The kernel size of convolution.half_step_residual
: Flag indication whether to use half step residual or notnum_decoder_layers
: The number of decoder layers.decoder_dropout_p
: The dropout probability of decoder.max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.rnn_type
: Type of rnn cell (rnn, lstm, gru)decoder_attn_mechanism
: The attention mechanism for decoder.optimizer
: Optimizer for training.
conformer_transducer
¶
model_name
: Model nameencoder_dim
: Dimension of encoder.num_encoder_layers
: The number of encoder layers.num_attention_heads
: The number of attention heads.feed_forward_expansion_factor
: The expansion factor of feed forward module.conv_expansion_factor
: The expansion factor of convolution module.input_dropout_p
: The dropout probability of inputs.feed_forward_dropout_p
: The dropout probability of feed forward module.attention_dropout_p
: The dropout probability of attention module.conv_dropout_p
: The dropout probability of convolution module.conv_kernel_size
: The kernel size of convolution.half_step_residual
: Flag indication whether to use half step residual or notnum_decoder_layers
: The number of decoder layers.decoder_dropout_p
: The dropout probability of decoder.max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.rnn_type
: Type of rnn cell (rnn, lstm, gru)decoder_hidden_state_dim
: Hidden state dimension of decoder.decoder_output_dim
: Output dimension of decoder.optimizer
: Optimizer for training.
joint_ctc_conformer_lstm
¶
model_name
: Model nameencoder_dim
: Dimension of encoder.num_encoder_layers
: The number of encoder layers.num_attention_heads
: The number of attention heads.feed_forward_expansion_factor
: The expansion factor of feed forward module.conv_expansion_factor
: The expansion factor of convolution module.input_dropout_p
: The dropout probability of inputs.feed_forward_dropout_p
: The dropout probability of feed forward module.attention_dropout_p
: The dropout probability of attention module.conv_dropout_p
: The dropout probability of convolution module.conv_kernel_size
: The kernel size of convolution.half_step_residual
: Flag indication whether to use half step residual or notnum_decoder_layers
: The number of decoder layers.decoder_dropout_p
: The dropout probability of decoder.num_decoder_attention_heads
: The number of decoder attention heads.max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.rnn_type
: Type of rnn cell (rnn, lstm, gru)decoder_attn_mechanism
: The attention mechanism for decoder.optimizer
: Optimizer for training.
transformer_transducer
¶
model_name
: Model nameencoder_dim
: Dimension of encoder named_ff
: Dimension of feed forward networknum_audio_layers
: Number of audio layersnum_label_layers
: Number of label layersnum_attention_heads
: Number of attention headsaudio_dropout_p
: Dropout probability of audio layerlabel_dropout_p
: Dropout probability of label layerdecoder_hidden_state_dim
: Hidden state dimension of decoderdecoder_output_dim
: Dimension of model output.conv_kernel_size
: Kernel size of convolution layer.max_positional_length
: Max length of positional encoding.optimizer
: Optimizer for training.
quartznet5x5
¶
model_name
: Model namenum_blocks
: Number of quartznet blocksnum_sub_blocks
: Number of quartznet sub blocksin_channels
: Input channels of jasper blocksout_channels
: Output channels of jasper block’s convolutionkernel_size
: Kernel size of jasper block’s convolutiondilation
: Dilation of jasper block’s convolutiondropout_p
: Dropout probabilityoptimizer
: Optimizer for training.
quartznet10x5
¶
model_name
: Model namenum_blocks
: Number of quartznet blocksnum_sub_blocks
: Number of quartznet sub blocksin_channels
: Input channels of jasper blocksout_channels
: Output channels of jasper block’s convolutionkernel_size
: Kernel size of jasper block’s convolutiondilation
: Dilation of jasper block’s convolutiondropout_p
: Dropout probabilityoptimizer
: Optimizer for training.
quartznet15x5
¶
model_name
: Model namenum_blocks
: Number of quartznet5x5 blocksnum_sub_blocks
: Number of quartznet5x5 sub blocksin_channels
: Input channels of jasper blocksout_channels
: Output channels of jasper block’s convolutionkernel_size
: Kernel size of jasper block’s convolutiondilation
: Dilation of jasper block’s convolutiondropout_p
: Dropout probabilityoptimizer
: Optimizer for training.
contextnet
¶
model_name
: Model namemodel_size
: Model sizeinput_dim
: Dimension of input vectornum_encoder_layers
: The number of convolution layerskernel_size
: Value of convolution kernel sizenum_channels
: The number of channels in the convolution filterencoder_dim
: Dimension of encoder output vectoroptimizer
: Optimizer for training
contextnet_lstm
¶
model_name
: Model namemodel_size
: Model sizeinput_dim
: Dimension of input vectornum_encoder_layers
: The number of convolution layersnum_decoder_layers
: The number of decoder layers.kernel_size
: Value of convolution kernel sizenum_channels
: The number of channels in the convolution filterencoder_dim
: Dimension of encoder output vectornum_attention_heads
: The number of attention heads.attention_dropout_p
: The dropout probability of attention module.decoder_dropout_p
: The dropout probability of decoder.max_length
: Max decoding length.teacher_forcing_ratio
: The ratio of teacher forcing.rnn_type
: Type of rnn cell (rnn, lstm, gru)decoder_attn_mechanism
: The attention mechanism for decoder.optimizer
: Optimizer for training.
contextnet_transducer
¶
model_name
: Model namemodel_size
: Model sizeinput_dim
: Dimension of input vectornum_encoder_layers
: The number of convolution layersnum_decoder_layers
: The number of rnn layerskernel_size
: Value of convolution kernel sizenum_channels
: The number of channels in the convolution filterhidden_dim
: The number of features in the decoder hidden stateencoder_dim
: Dimension of encoder output vectordecoder_output_dim
: Dimension of decoder output vectordropout
: Dropout probability of decoderrnn_type
: Type of rnn celloptimizer
: Optimizer for training
jasper5x3
¶
model_name
: Model namenum_blocks
: Number of jasper blocksnum_sub_blocks
: Number of jasper sub blocksin_channels
: Input channels of jasper blocksout_channels
: Output channels of jasper block’s convolutionkernel_size
: Kernel size of jasper block’s convolutiondilation
: Dilation of jasper block’s convolutiondropout_p
: Dropout probabilityoptimizer
: Optimizer for training.
jasper10x5
¶
model_name
: Model namenum_blocks
: Number of jasper blocksnum_sub_blocks
: Number of jasper sub blocksin_channels
: Input channels of jasper blocksout_channels
: Output channels of jasper block’s convolutionkernel_size
: Kernel size of jasper block’s convolutiondilation
: Dilation of jasper block’s convolutiondropout_p
: Dropout probabilityoptimizer
: Optimizer for training.
criterion
¶
label_smoothed_cross_entropy
¶
criterion_name
: Criterion name for training.reduction
: Reduction method of criterionsmoothing
: Ratio of smoothing loss (confidence = 1.0 - smoothing)
joint_ctc_cross_entropy
¶
criterion_name
: Criterion name for training.reduction
: Reduction method of criterionctc_weight
: Weight of ctc loss for training.cross_entropy_weight
: Weight of cross entropy loss for training.smoothing
: Ratio of smoothing loss (confidence = 1.0 - smoothing)zero_infinity
: Whether to zero infinite losses and the associated gradients.
perplexity
¶
criterion_name
: Criterion name for trainingreduction
: Reduction method of criterion
transducer
¶
criterion_name
: Criterion name for training.reduction
: Reduction method of criteriongather
: Reduce memory consumption.
ctc
¶
criterion_name
: Criterion name for trainingreduction
: Reduction method of criterionzero_infinity
: Whether to zero infinite losses and the associated gradients.
cross_entropy
¶
criterion_name
: Criterion name for trainingreduction
: Reduction method of criterion
lr_scheduler
¶
reduce_lr_on_plateau
¶
lr
: Learning ratescheduler_name
: Name of learning rate scheduler.lr_patience
: Number of epochs with no improvement after which learning rate will be reduced.lr_factor
: Factor by which the learning rate will be reduced. new_lr = lr * factor.
warmup
¶
lr
: Learning ratescheduler_name
: Name of learning rate scheduler.peak_lr
: Maximum learning rate.init_lr
: Initial learning rate.warmup_steps
: Warmup the learning rate linearly for the first N updatestotal_steps
: Total training steps.
warmup_reduce_lr_on_plateau
¶
lr
: Learning ratescheduler_name
: Name of learning rate scheduler.lr_patience
: Number of epochs with no improvement after which learning rate will be reduced.lr_factor
: Factor by which the learning rate will be reduced. new_lr = lr * factor.peak_lr
: Maximum learning rate.init_lr
: Initial learning rate.warmup_steps
: Warmup the learning rate linearly for the first N updates
tri_stage
¶
lr
: Learning ratescheduler_name
: Name of learning rate scheduler.init_lr
: Initial learning rate.init_lr_scale
: Initial learning rate scale.final_lr_scale
: Final learning rate scalephase_ratio
: Automatically sets warmup/hold/decay steps to the ratio specified here from max_updates. the ratios must add up to 1.0total_steps
: Total training steps.
transformer
¶
lr
: Learning ratescheduler_name
: Name of learning rate scheduler.peak_lr
: Maximum learning rate.final_lr
: Final learning rate.final_lr_scale
: Final learning rate scalewarmup_steps
: Warmup the learning rate linearly for the first N updatesdecay_steps
: Steps in decay stages
trainer
¶
cpu
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPU
gpu
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPUauto_select_gpus
: If enabled and gpus is an integer, pick available gpus automatically.
tpu
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPUuse_tpu
: If set True, will train with GPUtpu_cores
: Number of TPU cores
gpu-fp16
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPUauto_select_gpus
: If enabled and gpus is an integer, pick available gpus automatically.precision
: Double precision (64), full precision (32) or half precision (16). Can be used on CPU, GPU or TPUs.amp_backend
: The mixed precision backend to use (“native” or “apex”)
tpu-fp16
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPUuse_tpu
: If set True, will train with GPUtpu_cores
: Number of TPU coresprecision
: Double precision (64), full precision (32) or half precision (16). Can be used on CPU, GPU or TPUs.amp_backend
: The mixed precision backend to use (“native” or “apex”)
cpu-fp64
¶
seed
: Seed for training.accelerator
: Previously known as distributed_backend (dp, ddp, ddp2, etc…).accumulate_grad_batches
: Accumulates grads every k batches or as set up in the dict.num_workers
: The number of cpu coresbatch_size
: Size of batchcheck_val_every_n_epoch
: Check val every n train epochs.gradient_clip_val
: 0 means don’t clip.logger
: Training logger. {wandb, tensorboard}max_epochs
: Stop training once this number of epochs is reached.auto_scale_batch_size
: If set to True, will initially run a batch size finder trying to find the largest batch size that fits into memory.name
: Trainer namedevice
: Training device.use_cuda
: If set True, will train with GPUprecision
: Double precision (64), full precision (32) or half precision (16). Can be used on CPU, GPU or TPUs.amp_backend
: The mixed precision backend to use (“native” or “apex”)
tokenizer
¶
libri_subword
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.vocab_size
: Size of vocabulary.vocab_path
: Path of vocabulary file.
libri_character
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.vocab_path
: Path of vocabulary file.
aishell_character
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.vocab_path
: Path of vocabulary file.
kspon_subword
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.sp_model_path
: Path of sentencepiece model.vocab_size
: Size of vocabulary.
kspon_grapheme
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.vocab_path
: Path of vocabulary file.
kspon_character
¶
sos_token
: Start of sentence tokeneos_token
: End of sentence tokenpad_token
: Pad tokenblank_token
: Blank token (for CTC training)encoding
: Encoding of vocabunit
: Unit of vocabulary.vocab_path
: Path of vocabulary file.