configs/video2audio-cfm1-cfg-LargeDiT1-moe.yaml

model:
  base_learning_rate: 3.0e-06
  target: ldm.models.diffusion.cfm1_audio.CFM
  params:
    linear_start: 0.00085
    linear_end: 0.012
    num_timesteps_cond: 1
    log_every_t: 200
    timesteps: 1000
    first_stage_key: "mix_spec"
    cond_stage_key: "mix_video_feat"
    mel_dim: 20
    mel_length: 256
    channels: 0
    cond_stage_trainable: True
    conditioning_key: crossattn
    monitor: val/loss_simple_ema
    scale_by_std: true
    use_ema: false
    scheduler_config:
      target: ldm.lr_scheduler.LambdaLinearScheduler
      params:
        warm_up_steps:
        - 10000
        cycle_lengths:
        - 10000000000000
        f_start:
        - 1.0e-06
        f_max:
        - 1.0
        f_min:
        - 1.0
    unet_config:
      target: ldm.modules.diffusionmodules.flag_large_dit_moe.VideoFlagLargeDiT
      params:
        in_channels: 20
        context_dim: 768
        hidden_size: 768
        num_heads: 32
        depth: 16
        max_len: 1000
        num_experts: 4


    first_stage_config:
      target: ldm.models.autoencoder1d.AutoencoderKL
      params:
        embed_dim: 20
        monitor: val/rec_loss
        ckpt_path: /apdcephfs_intern/share_1316500/nlphuang/results/Text_to_audio/ldm_src/ckpt/epoch=000032.ckpt
        ddconfig:
          double_z: true
          in_channels: 80
          out_ch: 80
          z_channels: 20
          kernel_size: 5
          ch: 384
          ch_mult:
          - 1
          - 2
          - 4
          num_res_blocks: 2
          attn_layers:
          - 3
          down_layers:
          - 0
          dropout: 0.0
        lossconfig:
          target: torch.nn.Identity
    cond_stage_config:
      target: ldm.modules.encoders.modules.Video_Feat_Encoder_NoPosembed
      params:
        origin_dim: 512
        embed_dim: 768
        seq_len: 40

lightning:
  callbacks:
    image_logger:
      target: main.AudioLogger
      params:
        sample_rate: 16000
        for_specs: true
        increase_log_steps: false
        batch_frequency: 5000
        max_images: 8
        melvmin: -5
        melvmax: 1.5
        vocoder_cfg:
          target: vocoder.bigvgan.models.VocoderBigVGAN
          params:
            ckpt_vocoder: /apdcephfs_intern/share_1316500/nlphuang/results/Text_to_audio/ldm_src/ckpt/bigvnat
  trainer:
    benchmark: True
    gradient_clip_val: 1.0
  modelcheckpoint:
    params:
      monitor: epoch
      mode: max
      save_top_k: 10
      every_n_epochs: 5
      
data:
  target: main.DataModuleFromConfig
  params:
    batch_size: 16 # originally 220
    num_workers: 10
    wrap: True
    train:
      target: ldm.data.video_spec_maa2_dataset.audio_video_spec_fullset_Dataset_Train
      params:
        dataset_cfg:
          dataset1:
            dataset_name: VGGSound
            data_dir: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/
            video_dir: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/           # Not Necessary Except for Inference
            split_txt_path: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/split_txt

          feat_type: CAVP_feat
          sr: 16000
          duration: 10
          truncate: 131072
          fps: 4
          hop_len: 256
          drop: 0.2

    validation:
      target: ldm.data.video_spec_maa2_dataset.audio_video_spec_fullset_Dataset_Valid
      params:
        dataset_cfg:
          dataset1:
            dataset_name: VGGSound
            data_dir: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/
            video_dir: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/           # Not Necessary Except for Inference
            split_txt_path: /apdcephfs_intern/share_1316500/nlphuang/data/video_to_audio/vggsound/split_txt

          feat_type: CAVP_feat
          sr: 16000
          duration: 10
          truncate: 131072
          fps: 4
          hop_len: 256