XPart/partgen/config/infer.yaml

name: "Xpart Pipeline release"

ckpt_path: checkpoints/xpart.pt

shapevae:
  target: partgen.models.autoencoders.VolumeDecoderShapeVAE
  params:
    num_latents: &num_latents 1024
    embed_dim: 64
    num_freqs: 8
    include_pi: false
    heads: 16
    width: 1024
    num_encoder_layers: 8
    num_decoder_layers: 16
    qkv_bias: false
    qk_norm: true
    scale_factor: &z_scale_factor 1.0039506158752403
    geo_decoder_mlp_expand_ratio: 4
    geo_decoder_downsample_ratio: 1
    geo_decoder_ln_post: true
    point_feats: 4
    pc_size: &pc_size 81920 
    pc_sharpedge_size: &pc_sharpedge_size 0

bbox_predictor:
  target: partgen.bbox_estimator.auto_mask_api.AutoMask
  params:
    ckpt_path: checkpoints/p3sam.ckpt
conditioner:
  target: partgen.models.conditioner.condioner_release.Conditioner
  params:
    use_geo: true
    use_obj: true
    use_seg_feat: true
    geo_cfg:
      target: partgen.models.conditioner.part_encoders.PartEncoder
      output_dim: &cross2_output_dim 1024
      params:
        use_local: true
        local_feat_type: latents_shape # [latents,miche-point-query-structural-vae]
        num_tokens_cond: &num_tokens_cond 4096 # num_tokens :2048 for holopart conditioner
        local_geo_cfg:
          target: partgen.models.autoencoders.VolumeDecoderShapeVAE
          params:
            num_latents: *num_tokens_cond
            embed_dim: 64
            num_freqs: 8
            include_pi: false
            heads: 16
            width: 1024
            num_encoder_layers: 8
            num_decoder_layers: 16
            qkv_bias: false
            qk_norm: true
            scale_factor: *z_scale_factor
            geo_decoder_mlp_expand_ratio: 4
            geo_decoder_downsample_ratio: 1
            geo_decoder_ln_post: true
            point_feats: 4
            pc_size:  &pc_size_bbox 81920 
            pc_sharpedge_size: &pc_sharpedge_size_bbox 0

    obj_encoder_cfg:
      target: partgen.models.autoencoders.VolumeDecoderShapeVAE
      output_dim: &cross1_output_dim 1024
      params:
        num_latents: 4096
        embed_dim: 64
        num_freqs: 8
        include_pi: false
        heads: 16
        width: 1024
        num_encoder_layers: 8
        num_decoder_layers: 16
        qkv_bias: false
        qk_norm: true
        scale_factor: 1.0039506158752403
        geo_decoder_mlp_expand_ratio: 4
        geo_decoder_downsample_ratio: 1
        geo_decoder_ln_post: true
        point_feats: 4
        pc_size: *pc_size
        pc_sharpedge_size: *pc_sharpedge_size
    seg_feat_cfg:
      target: partgen.models.conditioner.sonata_extractor.SonataFeatureExtractor

model:
  target: partgen.models.partformer_dit.PartFormerDITPlain
  params:
    use_self_attention: true
    use_cross_attention: true
    use_cross_attention_2: true
    # cond 
    use_bbox_cond: false
    num_freqs: 8
    use_part_embed: true
    valid_num: 50 #*valid_num
    # para
    input_size: *num_latents
    in_channels: 64
    hidden_size: 2048
    encoder_hidden_dim: *cross1_output_dim # for object mesh
    encoder_hidden2_dim: *cross2_output_dim # for part in bbox
    depth: 21
    num_heads: 16
    qk_norm: true
    qkv_bias: false
    qk_norm_type: 'rms'
    with_decoupled_ca: false
    decoupled_ca_dim: *num_tokens_cond
    decoupled_ca_weight: 1.0
    use_attention_pooling: false
    use_pos_emb: false
    num_moe_layers: 6
    num_experts: 8
    moe_top_k: 2

scheduler:
  target: partgen.models.diffusion.schedulers.FlowMatchEulerDiscreteScheduler
  params:
    num_train_timesteps: 1000