config.yaml

# Dataset options: true for BBC, false for DAPS
urban_noise: true

# DAPS
version_A: clean
version_B: ipad_livingroom1
dataset_size: .inf 

# BBC.16K
noise_dataset: BBC.16K
snr: 5
noise_class_ids:
- 0
num_noise_files: 10
city: London

# Data preprocessing options 
clip_size: 32000
sr: 16000
b_size: 20
num_workers: 10
pure_noise_a: 0
pure_noise_b: 0.3

# Optimization options
epochs_a: 700
epochs_b: 1500

lr: 0.0001
beta1: 0.5
beta2: 0.999
weight_decay: 0.0001
step_size: 1000
gamma: 0.5

# Reconstruction loss
recon_loss_type: L2
# EQ
EQ: true
all_EQ: true

# Loss weights
# Straight paths
recon_a_w: 1
recon_b_w: 1
kl_a_w: 0.001
kl_b_w: 0.001

# Cross paths
cross_w: 0.001
cyc_recon_b_w: 10
cyc_kl_b_w: 0.001
recon_latent_b_w: 1

# STFT options
filter_length: 1024
hop_length: 256
win_length: 1024
window: hann

# Convolutional layers
conv_dim: 1
num_layers: 4
dim: 512
latent_dim: 64

# Custom intermediate dimensions for convolutional layers
# first_dim = dim, last_dim = latent_dim
customize_dim_a: False
inter_dims_a: 
- 512
- 256
- 128
- 64

customize_dim_b: True
inter_dims_b: 
- 512
- 400
- 300
- 200
- 100
- 64

# Conv1d (time dimension)
time:
    kernel_size: 7
    stride: 1
    padding: 3

# Magnitude spectrogram power (to reduce dynamic range)
mag_power: 0.5