conf.yaml

# Copyright (C) 2023 Mitsubishi Electric Research Laboratories (MERL)
#
# SPDX-License-Identifier: AGPL-3.0-or-later

data:
 train_batch_size: 10
 train_chunk_sec: 3.2
 eval_batch_size: 10
 num_workers: 4

model:
 # Mask-inference model
 n_sources: 7
 n_layers: 4
 hidden_dim: 600
 n_freq_bins: 257
 dropout: 0.3
 bidirectional: true

 # Hyperbolic-related
 hyperbolic_dim: 2
 hyperbolic_train: false
 hyperbolic_k: 1.0

training:
 epochs: 300
 num_gpu: 1
 learning_rate: 0.001
 seed: 2

features:
 window_length: 512  # STFT frame-size in samples
 fft_size: 512  # FFT size in samples
 hop_length: 256  # STFT hop-size in samples
 normalize: true  # normalize stft, 'window' divides window by its sum, 'default' divides by sqrt of window size, None uses no normalization

loss:
 loss_type: mask # type of loss, choices = ['mask', 'waveform', 'spectrogram']
 mag_weights: true # whether or not weighting the bins based off their spectrogram magnitude (only for loss_type==mask)