transformer.yaml

# ############################################################################
# Model: E2E ASR with Transformer
# Encoder: Transformer Encoder
# Decoder: Transformer Decoder + (CTC/ATT joint) beamsearch + TransformerLM
# Tokens: unigram
# losses: CTC + KLdiv (Label Smoothing loss)
# Training: Switchboard
# Authors:  Jianyuan Zhong, Titouan Parcollet, Samuele Cornell, Dominik Wagner
# ############################################################################
# Seed needs to be set at top of yaml, before objects with parameters are made

seed: 1312
__set_seed: !apply:torch.manual_seed [!ref <seed>]
output_folder: !ref results/transformer/<seed>
output_wer_folder: !ref <output_folder>/
save_folder: !ref <output_folder>/save
train_log: !ref <output_folder>/train_log.txt

# Language model (LM) pretraining
# NB: To avoid mismatch, the speech recognizer must be trained with the same
# tokenizer used for LM training. You can download everything from the
# Speechbrain HuggingFace repository, or you can provide a local
# path pointing to a directory containing the lm.ckpt and tokenizer.ckpt.
pretrained_lm_tokenizer_path: !PLACEHOLDER
tokenizer_file: !ref <pretrained_lm_tokenizer_path>/tokenizer.ckpt
lm_file: !ref <pretrained_lm_tokenizer_path>/lm.ckpt

# Data files
# Set the local path to the Switchboard dataset (e.g. /nfs/data/swbd) here
data_folder: !PLACEHOLDER
splits: [train, dev]
split_ratio: [99, 1]
skip_prep: False
# We don't need the Fisher corpus, here since
# it is only used for Tokenizer and LM training
add_fisher_corpus: False
# Remove optional/deletable parts of the transcript
normalize_words: True
# Maximum number of times the same utterance is allowed to appear
# in the training data.
max_utt: 300
train_csv: !ref <output_folder>/train.csv
valid_csv: !ref <output_folder>/dev.csv
# The test data is split into the full test set (test.csv),
# the Switchboard portion of the data (test_swbd.csv),
# and the Callhome portion of the data (test_callhome.csv).
test_csv:
  - !ref <output_folder>/test_swbd.csv
  - !ref <output_folder>/test_callhome.csv
  - !ref <output_folder>/test.csv

ckpt_interval_minutes: 30  # save checkpoint every N min

####################### Training Parameters ####################################
# To make Transformers converge, the global batch size should be large enough.
# The global batch size is computed as:
# batch_size * n_gpus * grad_accumulation_factor.
# Empirically, we found that this value should be >= 128.
# Please, set your parameters accordingly.
number_of_epochs: 100
batch_size: 48  # This works for 1x GPU with 40GB
ctc_weight: 0.3
grad_accumulation_factor: 2
max_grad_norm: 5.0
loss_reduction: batchmean
sorting: random
avg_checkpoints: 5

#dynamic_batching: False
#
#dynamic_batch_sampler:
#  feats_hop_size: 0.01
#  max_batch_len: 100000    # in terms of frames
#  num_buckets: 200
#  shuffle_ex: False    # re-creates batches at each epoch shuffling examples.
#  batch_ordering: descending
#  max_batch_ex: -1

# stages related parameters
lr_adam: 0.006

# Feature parameters
sample_rate: 16000
n_fft: 400
n_mels: 80

# Dataloader options
train_dataloader_opts:
  batch_size: !ref <batch_size>
  shuffle: True
  num_workers: 12

valid_dataloader_opts:
  batch_size: 1

test_dataloader_opts:
  batch_size: 1

####################### Model Parameters  ###########################
# Transformer
transformer_input_size: 1280
d_model: 256
nhead: 4
num_encoder_layers: 12
num_decoder_layers: 6
d_ffn: 2048
transformer_dropout: 0.1
activation: !name:torch.nn.GELU
output_neurons: 2000

# Outputs
blank_index: 0
label_smoothing: 0.1
pad_index: 0
bos_index: 1
eos_index: 2
# unk_index: 0

# Decoding parameters
min_decode_ratio: 0.0
max_decode_ratio: 1.0
valid_search_interval: 10
valid_beam_size: 10
lm_weight: 0.30
test_beam_size: 60
ctc_weight_decode: 0.30
temperature: 1.0
temperature_lm: 1.0
using_eos_threshold: False
eos_threshold: 1.5
length_normalization: True
using_max_attn_shift: False
max_attn_shift: 30
scorer_beam_scale: 0.3

CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
  input_shape: (8, 10, 80)
  num_blocks: 3
  num_layers_per_block: 1
  out_channels: (64, 64, 64)
  kernel_sizes: (5, 5, 1)
  strides: (2, 2, 1)
  residuals: (False, False, True)

Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
  input_size: !ref <transformer_input_size>
  tgt_vocab: !ref <output_neurons>
  d_model: !ref <d_model>
  nhead: !ref <nhead>
  num_encoder_layers: !ref <num_encoder_layers>
  num_decoder_layers: !ref <num_decoder_layers>
  d_ffn: !ref <d_ffn>
  dropout: !ref <transformer_dropout>
  activation: !ref <activation>
  encoder_module: transformer
  attention_type: regularMHA
  normalize_before: True
  causal: False

lm_model: !new:speechbrain.lobes.models.transformer.TransformerLM.TransformerLM # yamllint disable-line rule:line-length
  vocab: !ref <output_neurons>
  d_model: 264
  d_embedding: 128
  nhead: 12
  num_encoder_layers: 12
  num_decoder_layers: 0
  d_ffn: 1024
  dropout: 0.1
  activation: !name:torch.nn.ReLU
  normalize_before: False

tokenizer: !new:sentencepiece.SentencePieceProcessor

ctc_lin: !new:speechbrain.nnet.linear.Linear
  input_size: !ref <d_model>
  n_neurons: !ref <output_neurons>

seq_lin: !new:speechbrain.nnet.linear.Linear
  input_size: !ref <d_model>
  n_neurons: !ref <output_neurons>

modules:
  CNN: !ref <CNN>
  Transformer: !ref <Transformer>
  seq_lin: !ref <seq_lin>
  ctc_lin: !ref <ctc_lin>
  normalize: !ref <normalize>

model: !new:torch.nn.ModuleList
  - [!ref <CNN>, !ref <Transformer>, !ref <seq_lin>, !ref <ctc_lin>]
Adam: !name:torch.optim.Adam
  lr: !ref <lr_adam>
  betas: (0.9, 0.98)
  eps: 0.000000001

transformerlm_scorer: !new:speechbrain.decoders.scorer.TransformerLMScorer
  language_model: !ref <lm_model>
  temperature: !ref <temperature_lm>

# Scorer
ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
  eos_index: !ref <eos_index>
  blank_index: !ref <blank_index>
  ctc_fc: !ref <ctc_lin>

scorer: !new:speechbrain.decoders.scorer.ScorerBuilder
  full_scorers: [!ref <transformerlm_scorer>, !ref <ctc_scorer>]
  weights:
    transformerlm: !ref <lm_weight>
    ctc: !ref <ctc_weight_decode>
  scorer_beam_scale: !ref <scorer_beam_scale>

valid_search: !new:speechbrain.decoders.S2STransformerBeamSearcher
  modules: [!ref <Transformer>, !ref <seq_lin>]
  bos_index: !ref <bos_index>
  eos_index: !ref <eos_index>
  min_decode_ratio: !ref <min_decode_ratio>
  max_decode_ratio: !ref <max_decode_ratio>
  beam_size: !ref <valid_beam_size>
  using_eos_threshold: !ref <using_eos_threshold>
  length_normalization: !ref <length_normalization>
  using_max_attn_shift: !ref <using_max_attn_shift>
  max_attn_shift: !ref <max_attn_shift>
  scorer: !ref <scorer>
  temperature: !ref <temperature>

test_search: !new:speechbrain.decoders.S2STransformerBeamSearcher
  modules: [!ref <Transformer>, !ref <seq_lin>]
  bos_index: !ref <bos_index>
  eos_index: !ref <eos_index>
  min_decode_ratio: !ref <min_decode_ratio>
  max_decode_ratio: !ref <max_decode_ratio>
  beam_size: !ref <test_beam_size>
  using_eos_threshold: !ref <using_eos_threshold>
  using_max_attn_shift: !ref <using_max_attn_shift>
  max_attn_shift: !ref <max_attn_shift>
  eos_threshold: !ref <eos_threshold>
  temperature: !ref <temperature>
  length_normalization: !ref <length_normalization>
  scorer: !ref <scorer>


log_softmax: !new:torch.nn.LogSoftmax
  dim: -1

ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
  blank_index: !ref <blank_index>
  reduction: !ref <loss_reduction>

seq_cost: !name:speechbrain.nnet.losses.kldiv_loss
  label_smoothing: !ref <label_smoothing>
  reduction: !ref <loss_reduction>

noam_annealing: !new:speechbrain.nnet.schedulers.NoamScheduler
  lr_initial: !ref <lr_adam>
  n_warmup_steps: 25000
    # model_size: !ref <d_model>

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: !ref <save_folder>
  recoverables:
    model: !ref <model>
    noam_scheduler: !ref <noam_annealing>
    normalizer: !ref <normalize>
    counter: !ref <epoch_counter>

epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
  limit: !ref <number_of_epochs>

normalize: !new:speechbrain.processing.features.InputNormalization
  norm_type: global
  update_until_epoch: 4

############################## Augmentations ###################################

# Speed perturbation
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
  orig_freq: !ref <sample_rate>
  speeds: [95, 100, 105]

# Time Drop
time_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
  drop_length_low: 15
  drop_length_high: 25
  drop_count_low: 5
  drop_count_high: 5

# Frequency Drop
freq_drop: !new:speechbrain.augment.freq_domain.SpectrogramDrop
  drop_length_low: 25
  drop_length_high: 35
  drop_count_low: 2
  drop_count_high: 2
  dim: 2

# Time warp
time_warp: !new:speechbrain.augment.freq_domain.Warping

fea_augment: !new:speechbrain.augment.augmenter.Augmenter
  min_augmentations: 3
  max_augmentations: 3
  augment_prob: 1.0
  augmentations: [
    !ref <time_drop>,
    !ref <freq_drop>,
    !ref <time_warp>]

do_speed_perturb: True

compute_features: !new:speechbrain.lobes.features.Fbank
  sample_rate: !ref <sample_rate>
  n_fft: !ref <n_fft>
  n_mels: !ref <n_mels>

train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: !ref <train_log>

error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
acc_computer: !name:speechbrain.utils.Accuracy.AccuracyStats

# The pretrainer allows a mapping between pretrained files and instances that
# are declared in the yaml. E.g here, we will download the file lm.ckpt
# and it will be loaded into "lm" which is pointing to the <lm_model> defined
# before.
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  collect_in: !ref <save_folder>
  loadables:
    lm: !ref <lm_model>
    tokenizer: !ref <tokenizer>
  paths:
    lm: !ref <lm_file>
    tokenizer: !ref <tokenizer_file>