# Generated 2022-08-17 from:
# /netscratch/sagar/thesis/speechbrain/recipes/CommonVoice_de/LM/hparams/RNNLM.yaml
# yamllint disable
# ############################################################################
# Model: RNNLM of E2E ASR
# Tokens: unigram/char
# losses: NLL
# Training: Librispeech 960h transcript + LM corpus
# Authors:  Ju-Chieh Chou 2020, Jianyuan Zhong 2021
# ############################################################################

# Seed needs to be set at top of yaml, before objects with parameters are made
seed: 2995
__set_seed: !apply:torch.manual_seed [2995]
output_folder: results/RNN/2995
save_folder: results/RNN/2995/save
train_log: results/RNN/2995/train_log.txt

# Data files
# The data_folder is needed because we train the LM on the training
# transcriptions of LibriSpeech as well.
data_folder: ../cv_de_acc/

# path to the lm_corpus
# if set to null, it will automatically download from the internet
# in the case when there is no internet access, set this to your local file
lm_corpus_path: lm_corpus/orignal_data/de_lm_normalized.txt         # 17M sentences


# Tokenizer model
tokenizer_file: ../Tokenizer/results/unigram/1000_unigram.model

# Training parameters
number_of_epochs: 20
batch_size: 128
lr: 0.001
accu_steps: 1 # Gradient accumulation to simulate large batch training
ckpt_interval_minutes: 15 # save checkpoint every N min

# Dataloader options
train_dataloader_opts:
  batch_size: 128
  shuffle: true

valid_dataloader_opts:
  batch_size: 1

test_dataloader_opts:
  batch_size: 1

# Model parameters
emb_size: 128
activation: &id001 !name:torch.nn.LeakyReLU
dropout: 0.0
rnn_layers: 2
rnn_neurons: 2048
dnn_blocks: 1
dnn_neurons: 512

# Outputs
output_neurons: 1000 # index(blank/eos/bos) = 0 | char: 32 | unigram: 1000
# blank_index: 0
bos_index: 0
eos_index: 0


# Functions
model: &id002 !new:speechbrain.lobes.models.RNNLM.RNNLM

  output_neurons: 1000
  embedding_dim: 128
  activation: *id001
  dropout: 0.0
  rnn_layers: 2
  rnn_neurons: 2048
  dnn_blocks: 1
  dnn_neurons: 512

modules:
  model: *id002
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: results/RNN/2995/save
  recoverables:
    model: *id002
    scheduler: &id003 !new:speechbrain.nnet.schedulers.NewBobScheduler
      initial_value: 0.001
      improvement_threshold: 0.0025
      annealing_factor: 0.8
      patient: 0

    counter: &id004 !new:speechbrain.utils.epoch_loop.EpochCounter

      limit: 20

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

optimizer: !name:torch.optim.Adam
  lr: 0.001
  betas: (0.9, 0.98)
  eps: 0.000000001

lr_annealing: *id003
epoch_counter: *id004
compute_cost: !name:speechbrain.nnet.losses.nll_loss

train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: results/RNN/2995/train_log.txt

tokenizer: &id005 !new:sentencepiece.SentencePieceProcessor

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  collect_in: results/RNN/2995/save
  loadables:
    tokenizer: *id005
  paths:
    tokenizer: ../Tokenizer/results/unigram/1000_unigram.model