detect.yaml

##########################################
# Hyperparameters for detecting words
#
# Authors
#  * Peter Plantinga 2021
##########################################

seed: 1237
__set_seed: !apply:torch.manual_seed [!ref <seed>]

data_folder: !PLACEHOLDER
output_folder: !ref results/detect/<seed>
save_folder: !ref <output_folder>/save
train_log: !ref <output_folder>/train_log.txt
stats_file: !ref <output_folder>/stats.txt
encoder_save_file: ./encoder.txt

train_manifest: ./train_aligned.json
valid_manifest: ./valid_aligned.json
train_alignments: ./alignment.json
valid_alignments: ./alignment_valid.json

number_of_epochs: 30
n_mels: 40
learning_rate: 0.0001

frozen_module_keys: [] #[encoder]

dataloader_opts:
    batch_size: 4

encoder: !new:time_freq_crdnn.TimeFreqCRDNN
    input_size: !ref <n_mels>
    cnn_channels: [64, 128]
    rnn_neurons: 256
    rnn_bidirectional: True
    dnn_neurons: 256
    dropout: 0.2

phoneme_embedding: !new:speechbrain.nnet.embedding.Embedding
    num_embeddings: 40
    embedding_dim: 100

phn2word_embedding: !new:speechbrain.nnet.RNN.LSTM
    input_size: !ref <phoneme_embedding[embedding_dim]>
    hidden_size: 100
    bidirectional: True

prediction_model: !new:speechbrain.nnet.RNN.AttentionalRNNDecoder
    rnn_type: gru
    attn_type: location
    hidden_size: 100
    attn_dim: 100
    num_layers: 1
    enc_dim: !ref <encoder[dnn_neurons]>
    input_size: !ref <phn2word_embedding[hidden_size]> * 2
    kernel_size: 20
    channels: 20

# Output for existence
existence_output: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <prediction_model[hidden_size]>
    n_neurons: 1

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

counter: !new:speechbrain.utils.epoch_loop.EpochCounter
    limit: !ref <number_of_epochs>

compute_feats: !new:speechbrain.lobes.features.Fbank
    n_mels: !ref <n_mels>
normalize: !new:speechbrain.processing.features.InputNormalization
spec_augment: !new:speechbrain.lobes.augment.SpecAugment
    n_freq_mask: 4
    freq_mask_width: 5
    time_mask: False
    replace_with_zero: False

modules:
    encoder: !ref <encoder>
    phoneme_embedding: !ref <phoneme_embedding>
    phn2word_embedding: !ref <phn2word_embedding>
    prediction_model: !ref <prediction_model>
    existence_output: !ref <existence_output>

opt_class: !name:torch.optim.Adam
    lr: !ref <learning_rate>

lr_annealing: !new:speechbrain.nnet.schedulers.NewBobScheduler
    initial_value: !ref <learning_rate>
    improvement_threshold: 0.0025
    annealing_factor: 0.8
    patient: 0

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
    checkpoints_dir: !ref <save_folder>
    recoverables:
        encoder: !ref <encoder>
        phoneme_embedding: !ref <phoneme_embedding>
        phn2word_embedding: !ref <phn2word_embedding>
        prediction_model: !ref <prediction_model>
        existence_output: !ref <existence_output>
        counter: !ref <counter>

ctc_stats: !name:speechbrain.utils.metric_stats.MetricStats
    metric: !name:speechbrain.nnet.losses.ctc_loss
        reduction: batch

train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
    save_file: !ref <train_log>

pretrained: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        recognizer: !ref <encoder>