model.yaml

version: 1.0.0

model:
  base_learning_rate: 0.00001
  target: mug.diffusion.diffusion.DDPM
  params:
    linear_start: 0.0001
    linear_end: 0.02
    log_every_t: 100
    timesteps: 1000
    z_channels: 16
    z_length: 512
    parameterization: eps
    loss_type: smooth_l1
    monitor: val/loss_simple

    unet_config:
      target: mug.diffusion.unet.UNetModel
      params:
        in_channels: 16
        model_channels: 128
        out_channels: 16
        attention_resolutions: [ 8,4,2 ]
        num_res_blocks: 2
        channel_mult: [ 1,2,3,4 ]
        num_heads: 8
        context_dim: 128
        dropout: 0.0
        lstm_last: false
        lstm_layer: false
        s4_layer: true
        audio_channels: [ 256,512,512,512 ]
        use_checkpoint: false

    first_stage_config:
      target: mug.firststage.autoencoder.AutoencoderKL
      params:
        monitor: "val/loss"
        kl_weight: 0.000001
        ddconfig:
          x_channels: 16 # key_count * 4
          middle_channels: 64
          z_channels: 16
          num_groups: 8
          channel_mult: [ 1,2,4,4 ]  # num_down = len(ch_mult)-1
          num_res_blocks: 1
        lossconfig:
          target: torch.nn.Identity
          # target: mug.firststage.losses.ManiaReconstructLoss
          # params:
          #   weight_start_offset: 0.5
          #   weight_holding: 0.5
          #   weight_end_offset: 0.2
          #   label_smoothing: 0.001


    cond_stage_config:
      target: mug.cond.feature.BeatmapFeatureEmbedder
      params:
        path_to_yaml: "configs/mug/mania_beatmap_features.yaml"
        embed_dim: 128

    wave_stage_config:
      target: mug.cond.wave.MelspectrogramScaleEncoder1D
      params:
        n_freq: 128
        middle_channels: 128
        attention_resolutions: [ 128,256,512 ]
        num_res_blocks: 2
        num_heads: 8
        num_groups: 32
        dropout: 0.0
        use_checkpoint: true
        channel_mult: [ 1,1,1,1,2,2,2,4,4,4 ]


data:
  target: main.DataModuleFromConfig
  params:
    batch_size: 48
    wrap: False
    # num_workers: 0
    num_workers: 7
    common_params:
      txt_file: [  ]
      sr: 22050
      n_fft: 512
      max_audio_frame: 32768
      audio_note_window_ratio: 8
      n_mels: 128
      cache_dir: "data/audio_cache/"
      with_audio: true
      with_feature: true
      feature_yaml: "configs/mug/mania_beatmap_features.yaml"
      # audio_window_frame = n_fft / sr / 4 = 0.00580499 s
      # note_window_frame = audio_note_window_ratio * audio_window_frame = 0.04643990 s
      # max_duration = audio_window_frame * max_audio_frame = 190.2179 s = 3 min 10 s
      # max_note_frame = max_audio_frame / audio_note_window_ratio = 4096

      # old ===========
      # audio_window_frame = n_fft / sr / 4 = 0.02321995 s
      # note_window_frame = audio_note_window_ratio * audio_window_frame = 0.04643990 s
      # max_duration = audio_window_frame * max_audio_frame = 380.4357 s = 6 min 20 s
      # max_note_frame = max_audio_frame / audio_note_window_ratio = 8192
    train:
      target: mug.data.dataset.OsuTrainDataset
      params:
        mirror_p: 0.5
        feature_dropout_p: 0.5
        mirror_at_interval_p: 0
        rate_p: 0.2
        rate: [ 0.75,1.3 ]
        freq_mask_p: 0.0
        freq_mask_num: 15

    validation:
      target: mug.data.dataset.OsuValidDataset
      params: {}
#        test_txt_file: "data\\mug\\local_mania_4k_test.txt"


lightning:
  callbacks:
    beatmap_logger:
      target: mug.data.dataset.BeatmapLogger
      params:
        log_batch_idx: [ 0 ]
        splits: [ 'val' ]
        count: 16

  trainer:
    benchmark: True
    accelerator: dp
    accumulate_grad_batches: 1
    # precision: 16