torchscript/model_config.yaml

# Configuration for MQGAN training

project_name: "MQGAN"

data:
  data_dir: '../hifispeech4_CORS'
  output_dir: 'logs/mqgan_speech4_varcrop_newd'
  validation_split: 0.02
  crop_len: [256, 192, 128]
  batch_size: 16
  num_workers: 0

model:
  mel_channels: 128 # Number of mel frequency channels
  generator:
    channels: [512, 512, 512, 768]
    kernel_sizes: [3, 3, 5, 7]
    dropout: 0.1
    fsq_levels: [8, 5, 5, 5]
    refiner_base_channels: 64
    refiner_depth: 3
  discriminator_patch:
    hidden_channels: [256, 256, 384, 512, 512]
    kernel_sizes: [5,   5,   5,   3,   3,   3]
    strides: [[1,2], [2,2], [2,2], [2,1], [2,1], [2,1]]
  discriminator_multibin:
    hidden_channels: [128, 128, 256, 256, 384]
    kernel_sizes: [7, 5, 3, 3, 3, 3]
    n_bins: 8
    n_no_strides: 2

training:
  num_epochs: 1000
  lr: 0.0001
  beta1: 0.9
  beta2: 0.999
  lr_d_factor: 1.15
  d_beta1: 0.5
  d_beta2: 0.999
  warmup_steps: 1000
  discriminator_train_start_epoch: 10
  loss_weights:
    fm_lambda: 0.25
    Gloss_lambda: 15.0
    recon_lambda: 15.0
  use_fm_loss: False
  seed: 42
  no_cuda: False
  pretrained: null # path to pretrained model, or null

logging:
  eval_interval: 2
  save_interval: 2
  num_plot_examples: 10
  wandb:
    entity: null # Your wandb entity
    project: "MQGAN"