results/non_semi_final_stac/hyperparams.yaml

# Generated 2023-09-20 from:
# /home/salah/Code_Switched_Tunisian_Speech_Recognition/cs.yaml
# yamllint disable
# Generated 2023-08-03 from:
# /home/salah/new_tunisian_model/hparams/train_tunisian_withwavlm.yaml
# yamllint disable
# ################################
# Model: wav2vec2 + DNN + CTC
# Augmentation: SpecAugment
# Authors: Titouan Parcollet 2021
# ################################

seed: 1994
__set_seed: !!python/object/apply:torch.manual_seed [1234]
output_folder: results/non_semi_final_stac
wer_file: results/non_semi_final_stac/wer.txt
save_folder: results/non_semi_final_stac/save
train_log: results/non_semi_final_stac/train_log.txt



# Data files
data_folder: junk  # e.g, /localscratch/cv-corpus-5.1-2020-06-22/fr
train_tsv_file: junk/train.tsv                # Standard CommonVoice .tsv files
dev_tsv_file: junk/dev.tsv                # Standard CommonVoice .tsv files
test_tsv_file: junk/test.tsv                # Standard CommonVoice .tsv files
accented_letters: true

csv_folder: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean/
train_csv: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean//train.csv
valid_csv: /gpfsscratch/rech/nou/uzn19yk/switched_data/extended_clean//dev.csv
test_csv:
- all_tests/cs_test.csv
- all_tests/stac_test.csv

# We remove utterance slonger than 10s in the train/dev/test sets as
# longer sentences certainly correspond to "open microphones".
avoid_if_longer_than: 13.0
avoid_if_shorter_than: 0.5

# Training parameters
number_of_epochs: 20
lr: 0.0002
lr_weights: 0.01
sorting: ascending
auto_mix_prec: false
sample_rate: 16000
language_modelling: true
ngram_lm_path: arpas/pluslanguages_everything.arpa

# With data_parallel batch_size is split into N jobs
# With DDP batch_size is multiplied by N jobs
# Must be 3 per GPU to fit 32GB of VRAM
batch_size: 3
test_batch_size: 4

# Dataloader options
dataloader_options:
  batch_size: 3
  num_workers: 6

test_dataloader_options:
  batch_size: 4
  num_workers: 6

# Model parameters
activation: !name:torch.nn.Sigmoid
dnn_layers: 1
dnn_neurons: 768
freeze_encoder: true

# Outputs
output_neurons: 76  # BPE size, index(blank/eos/bos) = 0

# Functions and classes
#
epoch_counter: &id006 !new:speechbrain.utils.epoch_loop.EpochCounter
  limit: 20

encoder_dim: 3217
enc: &id001 !new:speechbrain.nnet.RNN.LSTM
  input_shape: [null, null, 3217]
  num_layers: 2
  bidirectional: true
  dropout: 0.2
  hidden_size: 1024

ctc_lin: &id002 !new:speechbrain.nnet.linear.Linear

  input_size: 2048
  n_neurons: 76

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

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

modules:
  enc: *id001
  ctc_lin: *id002
model: &id003 !new:torch.nn.ModuleList
- [*id001, *id002]
model_opt_class: !name:torch.optim.Adam
  lr: 0.0002

weights_opt_class: !name:torch.optim.Adam
  lr: 0.01

lr_annealing_model: &id004 !new:speechbrain.nnet.schedulers.NewBobScheduler
  initial_value: 0.0002
  improvement_threshold: 0.0025
  annealing_factor: 0.8
  patient: 0

lr_annealing_weights: &id005 !new:speechbrain.nnet.schedulers.NewBobScheduler
  initial_value: 0.01
  improvement_threshold: 0.0025
  annealing_factor: 0.9
  patient: 0

label_encoder: &id007 !new:speechbrain.dataio.encoder.CTCTextEncoder


checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: results/non_semi_final_stac/save
  recoverables:
    model: *id003
    scheduler_model: *id004
    scheduler_encoder: *id005
    counter: *id006
    tokenizer: *id007
blank_index: 0
unk_index: 1


train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: results/non_semi_final_stac/train_log.txt

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

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