examples/asr/conf/contextnet_rnnt/contextnet_rnnt_multilang.yaml

# This config contains the default values for training a modified ContextNet model with Transducer loss and BPE-based vocabulary.
# It also uses the AggregateTokenizer, so that the model is trained on more than one language, one language per tokenizer
# In contrast to original ContextNet, the same number of filters is used throughout the model.
# Default learning parameters in this config are set for effective batch size of 1k on 32 GPUs.
# To train it with smaller batch sizes, you may need to re-tune the learning parameters or use higher accumulate_grad_batches.

# It contains the default values for training a ContextNet ASR model, large size (~144M) with Transducer loss and sub-word encoding.

# Architecture and training config:
# Default learning parameters in this config are set for effective batch size of 1K. 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 ContextNet, other parameters are the same as in this config file.
#
#  +-------------+---------+------------+
#  | Model       | filters | time_masks |
#  +=============+=========+============+
#  | Small  (14M)|   256   |    2       |
#  +-------------+---------+------------+
#  | Medium (40M)|   512   |    5       |
#  +-------------+---------+------------+
#  | Large (145M)|   1024  |   10       |
#  +-------------------------------------

name: &name "ContextNet-8x-Stride-RNNT-mla"

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.

  train_ds:
    manifest_filepath: ???
    sample_rate: ${model.sample_rate}
    batch_size: 16  # Can be increased if memory allows or when using smaller model
    trim_silence: false
    max_duration: 16.7
    shuffle: true
    use_start_end_token: false
    num_workers: 16
    pin_memory: true
    # tarred datasets
    is_tarred: false
    tarred_audio_filepaths: null
    tarred_shard_strategy: "scatter"
    shuffle_n: 2048
    # bucketing params
    bucketing_strategy: "synced_randomized"
    bucketing_batch_size: null
  validation_ds:
    manifest_filepath: ???
    sample_rate: ${model.sample_rate}
    batch_size: 8
    shuffle: false
    use_start_end_token: false
    num_workers: 16
    pin_memory: true

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

  model_defaults:
    filters: 1024
    repeat: 5
    dropout: 0.1
    separable: true
    se: true
    se_context_size: -1
    kernel_size_factor: 1.0
    # encoder / decoder / joint values
    enc_hidden: 640
    pred_hidden: 640
    joint_hidden: 640

  tokenizer:
    type: agg  # The AggregateTokenizer is an ordered dict of N monolingual tokenizers, one per language id
    langs:
      en: # this language id must match the 'lang' field for english samples in the manifest
        dir: ??? # path to the en tokenizer which contains either tokenizer.model (bpe) or vocab.txt (for wpe)
        type: ??? # Can be either bpe or wpe
      es: # this language id must match the 'lang' field for spanish samples in the manifest
        dir: ??? # path to the es tokenizer which contains either tokenizer.model (bpe) or vocab.txt (for wpe)
        type: ??? # Can be either bpe or wpe

  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: &n_mels 80
    n_fft: 512
    frame_splicing: 1
    dither: 0.00001
    pad_to: 16
    stft_conv: false

  spec_augment:
    _target_: nemo.collections.asr.modules.SpectrogramAugmentation
    freq_masks: 2  # should be kept at 2
    time_masks: 10  # can be 5 for small-med models, 10 for larger models.
    freq_width: 27
    time_width: 0.05

  encoder:
    _target_: nemo.collections.asr.modules.ConvASREncoder
    feat_in: *n_mels
    activation: swish
    conv_mask: true
    init_mode: "tds_uniform"

    jasper:
      - filters: ${model.model_defaults.filters}
        repeat: 1
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: 0.0
        residual: false
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [2]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        stride_last: true
        residual_mode: "stride_add"
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [2]  # *stride
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        stride_last: true
        residual_mode: "stride_add"
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [2]  # stride
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        stride_last: true
        residual_mode: "stride_add"
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.filters}
        repeat: ${model.model_defaults.repeat}
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: ${model.model_defaults.dropout}
        residual: true
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}

      - filters: ${model.model_defaults.enc_hidden}
        repeat: 1
        kernel: [5]
        stride: [1]
        dilation: [1]
        dropout: 0.0
        residual: false
        separable: ${model.model_defaults.separable}
        se: ${model.model_defaults.se}
        se_context_size: ${model.model_defaults.se_context_size}
        kernel_size_factor: ${model.model_defaults.kernel_size_factor}


  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  # only 1 layer LSTM networks are exportable.
      t_max: null  # Maximum possible target seq length used for Chrono Initialization - https://arxiv.org/abs/1804.11188. Disabled by default.
      dropout: 0.1

  joint:
    _target_: nemo.collections.asr.modules.RNNTJoint
    log_softmax: null  # sets it 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: 0.1

  # RNNT decoding strategy
  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: 4
      score_norm: true
      return_best_hypothesis: False
      softmax_temperature: 1.0  # scale the logits by some temperature prior to softmax
      tsd_max_sym_exp: 10  # for Time Synchronous Decoding, int > 0
      alsd_max_target_len: 5.0  # for Alignment-Length Synchronous Decoding, float > 1.0
      maes_num_steps: 2  # for modified Adaptive Expansion Search, int > 0
      maes_prefix_alpha: 1  # for modified Adaptive Expansion Search, int > 0
      maes_expansion_beta: 2  # for modified Adaptive Expansion Search, int >= 0
      maes_expansion_gamma: 2.3  # for modified Adaptive Expansion Search, float >= 0

  # RNNT loss config
  loss:
    loss_name: "default"

    warprnnt_numba_kwargs:
      # FastEmit regularization: https://arxiv.org/abs/2010.11148
      fastemit_lambda: 0.001  # Values can be in range [1e-4, 1e-2]. Generally, 0.001 is good start.
      clamp: -1.0  # if > 0, applies gradient clamping in range [-clamp, clamp] for the joint tensor only.

  optim:
    name: novograd
    lr: 0.05

    # optimizer arguments
    betas: [0.9, 0.0]
    weight_decay: 0.001

    # scheduler setup
    sched:
      name: CosineAnnealing

      # scheduler config override
      warmup_steps: 5000
      warmup_ratio: null
      min_lr: 1e-6
      last_epoch: -1

trainer:
  devices: -1 # number of GPUs, -1 would use all available GPUs
  max_epochs: 100
  max_steps: -1 # computed at runtime if not set
  num_nodes: 1  # Should be set via SLURM variable `SLURM_JOB_NUM_NODES`
  accelerator: auto
  strategy: ddp
  accumulate_grad_batches: 1
  enable_checkpointing: False  # Provided by exp_manager
  logger: false  # Provided by exp_manager
  log_every_n_steps: 100  # Interval of logging.
  val_check_interval: 1.0 # Set to 0.25 to check 4 times per epoch, or an int for number of iterations
  check_val_every_n_epoch: 1  # RNNT decoding is slower than CTC, so eval takes longer. Increase value to speed up training slightly.
  precision: 32  # RNNT requires a lot of memory, so precision 16 is very important. Use very small batch size for precision 32.
  gradient_clip_val: 1.0  # Gradient norm clip value
  sync_batchnorm: true
  benchmark: false # needs to be false for models with variable-length speech input as it slows down training


exp_manager:
  exp_dir: null
  name: *name
  create_tensorboard_logger: true
  create_checkpoint_callback: true
  checkpoint_callback_params:
    monitor: "val_wer"
    mode: "min"
    save_top_k: 3
    always_save_nemo: true
  create_wandb_logger: false
  wandb_logger_kwargs:
    name: null
    project: null
    entity: null
  resume_if_exists: false
  resume_ignore_no_checkpoint: false