semi_wavlm_large_tunisian_ctc/1234/hyperparams.yaml

# Generated 2023-09-25 from:
# /home/salah/Tunisian-Speech-Recognition/train_semi.yaml
# yamllint disable
# ################################
# Model: wav2vec2 + DNN + CTC
# Augmentation: SpecAugment
# Authors: Titouan Parcollet 2021
# ################################

# Seed needs to be set at top of yaml, before objects with parameters are made
seed: 1234
__set_seed: !!python/object/apply:torch.manual_seed [1234]
output_folder: semi_wavlm_large_tunisian_ctc/1234
wer_file: semi_wavlm_large_tunisian_ctc/1234/wer.txt
save_folder: semi_wavlm_large_tunisian_ctc/1234/save
train_log: semi_wavlm_large_tunisian_ctc/1234/train_log.txt

# URL for the biggest LeBenchmark wav2vec french.
wav2vec2_folder: semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint

# Data files
data_folder: /path/to/data  # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
train_tsv_file: /path/to/data/train.tsv       # Standard CommonVoice .tsv files
dev_tsv_file: /path/to/data/dev.tsv       # Standard CommonVoice .tsv files
test_tsv_file: /path/to/data/test.tsv       # Standard CommonVoice .tsv files
accented_letters: true
language: fr # use 'it' for Italian, 'rw' for Kinyarwanda, 'en' for english
test_csv:
- /path/to/test_data

skip_prep: true # Skip data preparation

use_language_modelling: true
ngram_lm_path: outdomain.arpa

# We remove utterance slonger than 10s in the train/dev/test sets as
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 10.0
avoid_if_shorter_than: 1.2


# Training parameters
number_of_epochs: 12
lr: 1.0
lr_wav2vec: 0.0001
sorting: ascending
auto_mix_prec: false
sample_rate: 16000
ckpt_interval_minutes: 30 # save checkpoint every N min

# With data_parallel batch_size is split into N jobs
# With DDP batch_size is multiplied by N jobs
# Must be 6 per GPU to fit 16GB of VRAM
batch_size: 10
test_batch_size: 4

dataloader_options:
  batch_size: 10
  num_workers: 6
test_dataloader_options:
  batch_size: 4
  num_workers: 6

# BPE parameters
token_type: char  # ["unigram", "bpe", "char"]
character_coverage: 1.0

# Model parameters
# activation: !name:torch.nn.LeakyReLU
wav2vec_output_dim: 1024
dnn_neurons: 1024
freeze_wav2vec: false
freeze_feature_extractor: true
dropout: 0.15
warmup_steps: 500 # The wav2vec 2 model isn't updated for this amount of steps

# Outputs
output_neurons: 40  # BPE size, index(blank/eos/bos) = 0

# Decoding parameters
# Be sure that the bos and eos index match with the BPEs ones
blank_index: 0
unk_index: 1

#
# Functions and classes
#
epoch_counter: &id007 !new:speechbrain.utils.epoch_loop.EpochCounter

  limit: 12

augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
  sample_rate: 16000
  speeds: [95, 100, 105]

enc: &id002 !new:speechbrain.nnet.containers.Sequential
  input_shape: [null, null, 1024]
  linear1: !name:speechbrain.nnet.linear.Linear
    n_neurons: 1024
    bias: true
  bn1: !name:speechbrain.nnet.normalization.BatchNorm1d
  activation: !new:torch.nn.LeakyReLU
  drop: !new:torch.nn.Dropout
    p: 0.15
  linear2: !name:speechbrain.nnet.linear.Linear
    n_neurons: 1024
    bias: true
  bn2: !name:speechbrain.nnet.normalization.BatchNorm1d
  activation2: !new:torch.nn.LeakyReLU
  drop2: !new:torch.nn.Dropout
    p: 0.15
  linear3: !name:speechbrain.nnet.linear.Linear
    n_neurons: 1024
    bias: true
  bn3: !name:speechbrain.nnet.normalization.BatchNorm1d
  activation3: !new:torch.nn.LeakyReLU

wav2vec2: &id001 !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
  source: wavlm-large/
  output_norm: false
  freeze: false
  freeze_feature_extractor: true
  save_path: semi_wavlm_large_tunisian_ctc/1234/save/wav2vec2_checkpoint


ctc_lin: &id003 !new:speechbrain.nnet.linear.Linear

  input_size: 1024
  n_neurons: 40

log_softmax: !new:speechbrain.nnet.activations.Softmax
  apply_log: true

ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
  blank_index: 0

modules:
  wav2vec2: *id001
  enc: *id002
  ctc_lin: *id003
model: &id004 !new:torch.nn.ModuleList
- [*id002, *id003]
model_opt_class: !name:torch.optim.Adadelta
  lr: 1.0
  rho: 0.95
  eps: 1.e-8

wav2vec_opt_class: !name:torch.optim.Adam
  lr: 0.0001

lr_annealing_model: &id005 !new:speechbrain.nnet.schedulers.NewBobScheduler
  initial_value: 1.0
  improvement_threshold: 0.0025
  annealing_factor: 0.8
  patient: 0

lr_annealing_wav2vec: &id006 !new:speechbrain.nnet.schedulers.NewBobScheduler
  initial_value: 0.0001
  improvement_threshold: 0.0025
  annealing_factor: 0.9
  patient: 0

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: semi_wavlm_large_tunisian_ctc/1234/save
  recoverables:
    wav2vec2: *id001
    model: *id004
    scheduler_model: *id005
    scheduler_wav2vec: *id006
    counter: *id007
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: semi_wavlm_large_tunisian_ctc/1234/train_log.txt

error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats

cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
  split_tokens: true