configs/deploy/clip_sp+pk_aslperceiver=256_01_4096_8_udt=03.yaml

name: "0428_clip_subsp+pk_sal_perceiver=256_01_4096_8_udt=03"
#wandb:
#  project: "image_diffuser"
#  offline: false

training:
  steps: 500000
  use_amp: true
  ckpt_path: ""
  base_lr: 1.e-4
  gradient_clip_val: 5.0
  gradient_clip_algorithm: "norm"
  every_n_train_steps: 5000
  val_check_interval: 1024
  limit_val_batches: 16

# dataset
dataset:
  target: michelangelo.data.asl_torch_dataset.MultiAlignedShapeImageTextModule
  params:
    batch_size: 38
    num_workers: 4
    val_num_workers: 4
    buffer_size: 256
    return_normal: true
    random_crop: false
    surface_sampling: true
    pc_size: &pc_size 4096
    image_size: 384
    mean: &mean [0.5, 0.5, 0.5]
    std: &std [0.5, 0.5, 0.5]

    cond_stage_key: "text"

    meta_info:
      3D-FUTURE:
        render_folder: "/root/workspace/cq_workspace/datasets/3D-FUTURE/renders"
        tar_folder: "/root/workspace/datasets/make_tars/3D-FUTURE"

      ABO:
        render_folder: "/root/workspace/cq_workspace/datasets/ABO/renders"
        tar_folder: "/root/workspace/datasets/make_tars/ABO"

      GSO:
        render_folder: "/root/workspace/cq_workspace/datasets/GSO/renders"
        tar_folder: "/root/workspace/datasets/make_tars/GSO"

      TOYS4K:
        render_folder: "/root/workspace/cq_workspace/datasets/TOYS4K/TOYS4K/renders"
        tar_folder: "/root/workspace/datasets/make_tars/TOYS4K"

      3DCaricShop:
        render_folder: "/root/workspace/cq_workspace/datasets/3DCaricShop/renders"
        tar_folder: "/root/workspace/datasets/make_tars/3DCaricShop"

      Thingi10K:
        render_folder: "/root/workspace/cq_workspace/datasets/Thingi10K/renders"
        tar_folder: "/root/workspace/datasets/make_tars/Thingi10K"

      shapenet:
        render_folder: "/root/workspace/cq_workspace/datasets/shapenet/renders"
        tar_folder: "/root/workspace/datasets/make_tars/shapenet"

      pokemon:
        render_folder: "/root/workspace/cq_workspace/datasets/pokemon/renders"
        tar_folder: "/root/workspace/datasets/make_tars/pokemon"

      objaverse:
        render_folder: "/root/workspace/cq_workspace/datasets/objaverse/renders"
        tar_folder: "/root/workspace/datasets/make_tars/objaverse"

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:
        # ckpt_path: "/root/workspace/cq_workspace/michelangelo/experiments/aligned_shape_latents/clip_aslperceiver_sp+pk_01_01/ckpt/ckpt-step=00230000.ckpt"
        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: "/mnt/shadow_cv_training/stevenxxliu/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: "/mnt/shadow_cv_training/stevenxxliu/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"

logger:
  target: michelangelo.utils.trainings.mesh_log_callback.TextConditionalASLDiffuserLogger
  params:
    step_frequency: 1000
    num_samples: 4
    sample_times: 4
    bounds: [-1.1, -1.1, -1.1, 1.1, 1.1, 1.1]
    octree_depth: 7
    num_chunks: 10000