Training/hyperparams.yaml

# Generated 2022-05-27 from:
# /data/n.abdoumohamed/dvoice-africa/speechbrain/recipes/DVoice/ASR/CTC/hparams/train_amharic.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: 1249
__set_seed: !!python/object/apply:torch.manual_see
output_folder: results/wav2vec2_ctc_AMHARIC/1249
wer_file: results/wav2vec2_ctc_AMHARIC/1249/wer.txt
save_folder: results/wav2vec2_ctc_AMHARIC/1249/save
train_log: results/wav2vec2_ctc_AMHARIC/1249/train_log.txt

# URL for the biggest LeBenchmark wav2vec french.
wav2vec2_hub: facebook/wav2vec2-large-xlsr-53

# Data files
data_folder: ASR/AMHARIC/data # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
train_csv_file: ASR/AMHARIC/data/train.csv # Standard CommonVoice .tsv files
dev_csv_file: ASR/AMHARIC/data/dev.csv # Standard CommonVoice .tsv files
test_csv_file: ASR/AMHARIC/data/test.csv # Standard CommonVoice .tsv files
accented_letters: true
language: amharic
train_csv: results/wav2vec2_ctc_AMHARIC/save/train.csv
valid_csv: results/wav2vec2_ctc_AMHARIC/save/dev.csv
test_csv: results/wav2vec2_ctc_AMHARIC/save/test.csv
skip_prep: false # Skip data preparation
data_augmentation: false # Skip data augmentation

# We remove utterance slonger than 10s in the train/dev/test sets as
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 15.0

# Training parameters
number_of_epochs: 30
number_of_ctc_epochs: 15
lr: 1.0
lr_wav2vec: 0.0001
ctc_weight: 0.3
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: 4
test_batch_size: 4

dataloader_options:
  batch_size: 4
  num_workers: 2
test_dataloader_options:
  batch_size: 4
  num_workers: 2

# 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

# Outputs
output_neurons: 224  # 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
bos_index: 1
eos_index: 2
min_decode_ratio: 0.0
max_decode_ratio: 1.0
beam_size: 80
eos_threshold: 1.5
using_max_attn_shift: true
max_attn_shift: 140
ctc_weight_decode: 0.0
temperature: 1.50

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

  limit: 30

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: facebook/wav2vec2-large-xlsr-53
  output_norm: true
  freeze: false
  save_path: results/wav2vec2_ctc_AMHARIC/1249/save/wav2vec2_checkpoint

#####
# Uncomment this block if you prefer to use a Fairseq pretrained model instead
# of a HuggingFace one. Here, we provide an URL that is obtained from the
# Fairseq github for the multilingual XLSR.
#
#wav2vec2_url: https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr_53_56k.pt
#wav2vec2: !new:speechbrain.lobes.models.fairseq_wav2vec.FairseqWav2Vec2
#    pretrained_path: !ref <wav2vec2_url>
#    output_norm: True
#    freeze: False
#    save_path: !ref <save_folder>/wav2vec2_checkpoint/model.pt
#####


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

  input_size: 1024
  n_neurons: 224

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: results/wav2vec2_ctc_AMHARIC/1249/save
  recoverables:
    wav2vec2: *id001
    model: *id004
    scheduler_model: *id005
    scheduler_wav2vec: *id006
    counter: *id007
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: results/wav2vec2_ctc_AMHARIC/1249/train_log.txt

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

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