hyperparams.yaml

# Pretrain folder (HuggingFace)
pretrained_path: chaanks/transformer-s2ut-fr-en-hubert-k100

sample_rate: 16000

# URL for the HuggingFace model we want to load as encoder
wav2vec2_hub: LeBenchmark/wav2vec2-FR-7K-large

# Outputs
vocab_size: 103
blank_index: 102
bos_index: 100
eos_index: 101
pad_index: 102
label_smoothing: 0.0

# Encoder
features_dim: 1024

# Length Regulator
enc_kernel_size: 3
enc_stride: 2

# Transformer decoder
embedding_size: 512
d_model: 512
nhead: 8
num_encoder_layers: 0
num_decoder_layers: 6
d_ffn: 2048
transformer_dropout: 0.1
activation: !name:torch.nn.GELU
output_neurons: !ref <vocab_size>
attention_type: "regularMHA"

# Decoding parameters
min_decode_ratio: 0.0
max_decode_ratio: 1.0

wav2vec2: !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
    source: !ref <wav2vec2_hub>
    output_norm: True
    freeze: True
    freeze_feature_extractor: True
    apply_spec_augment : False
    save_path: wav2vec2_checkpoints

length_regulator: !new:speechbrain.nnet.CNN.Conv1d
  input_shape: [null, null, !ref <features_dim>]
  out_channels: !ref <embedding_size>
  kernel_size: !ref <enc_kernel_size>
  stride: !ref <enc_stride>

transformer_decoder: !new:speechbrain.lobes.models.transformer.TransformerST.TransformerST # yamllint disable-line rule:line-length
    input_size: !ref <embedding_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>
    attention_type: !ref <attention_type>
    normalize_before: True
    causal: False

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

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

model: !new:torch.nn.ModuleList
    - [!ref <length_regulator>, !ref <transformer_decoder>, !ref <seq_lin>]

encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    wav2vec2: !ref <wav2vec2>
    length_regulator: !ref <length_regulator>

decoder_beamsearch: !new:speechbrain.decoders.seq2seq.S2STransformerBeamSearch
    modules: [!ref <transformer_decoder>, !ref <seq_lin>, null]
    bos_index: !ref <bos_index>
    eos_index: !ref <eos_index>
    blank_index: !ref <blank_index>
    min_decode_ratio: !ref <min_decode_ratio>
    max_decode_ratio: !ref <max_decode_ratio>
    beam_size: 10
    temperature: 1.0

modules:
    encoder: !ref <encoder>
    decoder: !ref <decoder_beamsearch>

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
      model: !ref <model>
      wav2vec2: !ref <wav2vec2>
    paths:
        wav2vec2: !ref <pretrained_path>/wav2vec2.ckpt
        model: !ref <pretrained_path>/model.ckpt