group: shape
name: shape_recon
load: 

batch_size: 28
debug: false
profile: false
image_size: [224,224]
gpu: 0
max_epoch: 15
output_root: output
resume: false
seed: 0
yaml:

pretrain:
    depth: weights/depth.ckpt

arch:
    # general
    num_heads: 8
    latent_dim: 256
    win_size: 16
    # depth
    depth:
        encoder: resnet
        n_blocks: 12
        dsp: 2
        pretrained: model/depth/pretrained_weights/omnidata_dpt_depth_v2.ckpt
    # rgb
    rgb:
        encoder: 
        n_blocks: 12 
    # implicit
    impl:
        n_channels: 256
        # attention-related
        att_blocks: 2
        mlp_ratio: 4.
        posenc_perlayer: false
        # mlp-related
        mlp_layers: 8
        posenc_3D: 0
        skip_in: [2,4,6]

eval:
    batch_size: 2
    brute_force: false
    n_vis: 50
    vox_res: 64
    num_points: 10000
    range: [-1.5,1.5]
    icp: false
    f_thresholds: [0.005, 0.01, 0.02, 0.05, 0.1, 0.2]

data:
    num_classes_test: 15
    max_img_cat: 
    dataset_train: synthetic
    dataset_test: synthetic
    num_workers: 6
    bgcolor: 1
    pix3d:                       
        cat: 
    ocrtoc:                       
        cat: 
        erode_mask: 
    synthetic:        
        subset: objaverse_LVIS,ShapeNet55
        percentage: 1
    train_sub:
    val_sub:

training:
    n_sdf_points: 4096
    shape_loss:
        impt_weight: 1
        impt_thres: 0.01
    depth_loss:
        grad_reg: 0.1
        depth_inv: true
        mask_shrink: false

loss_weight:
    shape: 1
    depth: 
    intr: 
    
optim:
    lr: 3.e-5
    lr_ft: 1.e-5
    weight_decay: 0.05
    fix_dpt: false
    fix_clip: true
    clip_norm: 
    amp: false
    accum: 1
    sched: false

tb:
    num_images: [4,8]

freq:
    print: 200
    print_eval: 100
    scalar: 1000 # iterations
    vis: 1000 # iterations
    save_vis: 1000
    ckpt_latest: 1000 # iterations
    eval: 1