configs/text_cond_diffuser_asl/text-ASLDM-256.yaml

model:
  target: michelangelo.models.asl_diffusion.clip_asl_diffuser_pl_module.ClipASLDiffuser
  params:
    first_stage_config:
      target: michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule
      params:
        shape_module_cfg:
          target: michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver
          params:
            num_latents: &num_latents 256
            embed_dim: &embed_dim 64
            point_feats: 3   # normal
            num_freqs: 8
            include_pi: false
            heads: 12
            width: 768
            num_encoder_layers: 8
            num_decoder_layers: 16
            use_ln_post: true
            init_scale: 0.25
            qkv_bias: false
            use_checkpoint: true
        aligned_module_cfg:
          target: michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule
          params:
            clip_model_version: "/home/user/app/checkpoints/clip/clip-vit-large-patch14"

        loss_cfg:
          target: torch.nn.Identity

    cond_stage_config:
      target: michelangelo.models.conditional_encoders.encoder_factory.FrozenAlignedCLIPTextEmbedder
      params:
        version: "/home/user/app/checkpoints/clip/clip-vit-large-patch14"
        zero_embedding_radio: 0.1
        max_length: 77

    first_stage_key: "surface"
    cond_stage_key: "text"
    scale_by_std: false

    denoiser_cfg:
      target: michelangelo.models.asl_diffusion.asl_udt.ConditionalASLUDTDenoiser
      params:
        input_channels: *embed_dim
        output_channels: *embed_dim
        n_ctx: *num_latents
        width: 768
        layers: 8   # 2 * 6 + 1 = 13
        heads: 12
        context_dim: 768
        init_scale: 1.0
        skip_ln: true
        use_checkpoint: true

    scheduler_cfg:
      guidance_scale: 7.5
      num_inference_steps: 50
      eta: 0.0

      noise:
        target: diffusers.schedulers.DDPMScheduler
        params:
          num_train_timesteps: 1000
          beta_start: 0.00085
          beta_end: 0.012
          beta_schedule: "scaled_linear"
          variance_type: "fixed_small"
          clip_sample: false
      denoise:
        target: diffusers.schedulers.DDIMScheduler
        params:
          num_train_timesteps: 1000
          beta_start: 0.00085
          beta_end: 0.012
          beta_schedule: "scaled_linear"
          clip_sample: false   # clip sample to -1~1
          set_alpha_to_one: false
          steps_offset: 1

    optimizer_cfg:
      optimizer:
        target: torch.optim.AdamW
        params:
          betas: [0.9, 0.99]
          eps: 1.e-6
          weight_decay: 1.e-2

      scheduler:
        target: michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
        params:
          warm_up_steps: 5000
          f_start: 1.e-6
          f_min: 1.e-3
          f_max: 1.0

    loss_cfg:
      loss_type: "mse"