conf/conformer_transducer_bpe_large.yaml

# It contains the default values for training a Conformer-Transducer ASR model, large size (~120M) with Transducer loss and sub-word encoding.

# Architecture and training config:
# Default learning parameters in this config are set for effective batch size of 2K. To train it with smaller effective
# batch sizes, you may need to re-tune the learning parameters or use higher accumulate_grad_batches.
# Here are the recommended configs for different variants of Conformer-Transducer, other parameters are the same as in this config file.
#
#  +-------------+---------+---------+----------+--------------+--------------------------+
#  | Model       | d_model | n_heads | n_layers | weight_decay | pred_hidden/joint_hidden |
#  +=============+=========+========+===========+==============+==========================+
#  | Small  (14M)|   176   |    4   |    16     |     0.0      |           320            |
#  +-------------+---------+--------+-----------+--------------+--------------------------+
#  | Medium (32M)|   256   |    4   |    16     |     1e-3     |           640            |
#  +-------------+---------+--------+-----------+--------------+--------------------------+
#  | Large (120M)|   512   |    8   |    17     |     1e-3     |           640            |
#  +-----------------------------------------------------------+--------------------------+
#

# You may find more info about Conformer-Transducer here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/models.html#conformer-transducer
# Pre-trained models of Conformer-Transducer can be found here: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/results.html
# The checkpoint of the large model trained on NeMo ASRSET with this recipe can be found here: https://ngc.nvidia.com/catalog/models/nvidia:nemo:stt_en_conformer_transducer_large

name: "Conformer-Transducer-BPE"

model:
  sample_rate: 16000
  compute_eval_loss: false # eval samples can be very long and exhaust memory. Disable computation of transducer loss during validation/testing with this flag.
  log_prediction: true # enables logging sample predictions in the output during training
  skip_nan_grad: false

  model_defaults:
    enc_hidden: ${model.encoder.d_model}
    pred_hidden: 640
    joint_hidden: 640

  train_ds:
    manifest_filepath: ???
    sample_rate: ${model.sample_rate}
    batch_size: 16 # you may increase batch_size if your memory allows
    shuffle: true
    num_workers: 8
    pin_memory: true
    use_start_end_token: false
    trim_silence: false
    max_duration: 16.7 # it is set for LibriSpeech, you may need to update it for your dataset
    min_duration: 0.1
    # tarred datasets
    is_tarred: false
    tarred_audio_filepaths: null
    shuffle_n: 2048
    # bucketing params
    bucketing_strategy: "synced_randomized"
    bucketing_batch_size: null

  validation_ds:
    manifest_filepath: ???
    sample_rate: ${model.sample_rate}
    batch_size: 16
    shuffle: false
    num_workers: 8
    pin_memory: true
    use_start_end_token: false

  test_ds:
    manifest_filepath: null
    sample_rate: ${model.sample_rate}
    batch_size: 16
    shuffle: false
    num_workers: 8
    pin_memory: true
    use_start_end_token: false

  # You may find more detail on how to train a tokenizer at: /scripts/tokenizers/process_asr_text_tokenizer.py
  tokenizer:
    dir: ???  # path to directory which contains either tokenizer.model (bpe) or vocab.txt (for wpe)
    type: bpe  # Can be either bpe (SentencePiece tokenizer) or wpe (WordPiece tokenizer)

  preprocessor:
    _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
    sample_rate: ${model.sample_rate}
    normalize: "per_feature"
    window_size: 0.025
    window_stride: 0.01
    window: "hann"
    features: 80
    n_fft: 512
    frame_splicing: 1
    dither: 0.00001
    pad_to: 0

  spec_augment:
    _target_: nemo.collections.asr.modules.SpectrogramAugmentation
    freq_masks: 2 # set to zero to disable it
    time_masks: 10 # set to zero to disable it
    freq_width: 27
    time_width: 0.05

  encoder:
    _target_: nemo.collections.asr.modules.ConformerEncoder
    feat_in: ${model.preprocessor.features}
    feat_out: -1 # you may set it if you need different output size other than the default d_model
    n_layers: 17
    d_model: 512

    # Sub-sampling params
    subsampling: striding # vggnet, striding, stacking or stacking_norm, dw_striding
    subsampling_factor: 4 # must be power of 2 for striding and vggnet
    subsampling_conv_channels: -1 # set to -1 to make it equal to the d_model
    causal_downsampling: false

    # Feed forward module's params
    ff_expansion_factor: 4

    # Multi-headed Attention Module's params
    self_attention_model: rel_pos # rel_pos or abs_pos
    n_heads: 8 # may need to be lower for smaller d_models
    # [left, right] specifies the number of steps to be seen from left and right of each step in self-attention
    att_context_size: [-1, -1] # -1 means unlimited context
    att_context_style: regular # regular or chunked_limited
    xscaling: true # scales up the input embeddings by sqrt(d_model)
    untie_biases: true # unties the biases of the TransformerXL layers
    pos_emb_max_len: 5000

    # Convolution module's params
    conv_kernel_size: 31
    conv_norm_type: 'batch_norm' # batch_norm or layer_norm or groupnormN (N specifies the number of groups)
    # conv_context_size can be"causal" or a list of two integers while conv_context_size[0]+conv_context_size[1]+1==conv_kernel_size
    # null means [(kernel_size-1)//2, (kernel_size-1)//2], and 'causal' means [(kernel_size-1), 0]
    conv_context_size: null

    ### regularization
    dropout: 0.1 # The dropout used in most of the Conformer Modules
    dropout_emb: 0.0 # The dropout used for embeddings
    dropout_att: 0.1 # The dropout for multi-headed attention modules

  decoder:
    _target_: nemo.collections.asr.modules.RNNTDecoder
    normalization_mode: null # Currently only null is supported for export.
    random_state_sampling: false # Random state sampling: https://arxiv.org/pdf/1910.11455.pdf
    blank_as_pad: true # This flag must be set in order to support exporting of RNNT models + efficient inference.

    prednet:
      pred_hidden: ${model.model_defaults.pred_hidden}
      pred_rnn_layers: 1
      t_max: null
      dropout: 0.2

  joint:
    _target_: nemo.collections.asr.modules.RNNTJoint
    log_softmax: null  # 'null' would set it automatically according to CPU/GPU device
    preserve_memory: false  # dramatically slows down training, but might preserve some memory

    # Fuses the computation of prediction net + joint net + loss + WER calculation
    # to be run on sub-batches of size `fused_batch_size`.
    # When this flag is set to true, consider the `batch_size` of *_ds to be just `encoder` batch size.
    # `fused_batch_size` is the actual batch size of the prediction net, joint net and transducer loss.
    # Using small values here will preserve a lot of memory during training, but will make training slower as well.
    # An optimal ratio of fused_batch_size : *_ds.batch_size is 1:1.
    # However, to preserve memory, this ratio can be 1:8 or even 1:16.
    # Extreme case of 1:B (i.e. fused_batch_size=1) should be avoided as training speed would be very slow.
    fuse_loss_wer: true
    fused_batch_size: 16

    jointnet:
      joint_hidden: ${model.model_defaults.joint_hidden}
      activation: "relu"
      dropout: 0.2

  decoding:
    strategy: "greedy_batch" # can be greedy, greedy_batch, beam, tsd, alsd.

    # greedy strategy config
    greedy:
      max_symbols: 10

    # beam strategy config
    beam:
      beam_size: 2
      return_best_hypothesis: False
      score_norm: true
      tsd_max_sym_exp: 50  # for Time Synchronous Decoding
      alsd_max_target_len: 2.0  # for Alignment-Length Synchronous Decoding

  loss:
    loss_name: "default"

    warprnnt_numba_kwargs:
      # FastEmit regularization: https://arxiv.org/abs/2010.11148
      # You may enable FastEmit to reduce the latency of the model for streaming
      fastemit_lambda: 0.0  # Recommended values to be in range [1e-4, 1e-2], 0.001 is a good start.
      clamp: -1.0  # if > 0, applies gradient clamping in range [-clamp, clamp] for the joint tensor only.

  # Adds Gaussian noise to the gradients of the decoder to avoid overfitting
  variational_noise:
    start_step: 0
    std: 0.0

  optim:
    name: adamw
    lr: 5.0
    # optimizer arguments
    betas: [0.9, 0.98]
    weight_decay: 1e-3

    # scheduler setup
    sched:
      name: NoamAnnealing
      d_model: ${model.encoder.d_model}
      # scheduler config override
      warmup_steps: 10000
      warmup_ratio: null
      min_lr: 1e-6