livingrooms_mixed.yaml

data:
    dataset_type: "cached_threedfront"
    encoding_type: "cached_diffusion_cosin_angle_wocm"
    dataset_directory: "livingroom"
    annotation_file: "livingroom_threed_front_splits.csv"
    augmentations: ["fixed_rotations"]
    train_stats: "dataset_stats.txt"
    room_layout_size: "64,64"

network:
    type: "diffusion_scene_layout_mixed"

    # encoding dim
    sample_num_points: 21 # max_length 
    angle_dim: 2 

    # room mask condition
    room_mask_condition: true
    room_latent_dim: 64
    
    # position condition
    learnable_embedding: true
    position_condition: true
    position_emb_dim: 64

    # diffusion config
    time_num: 1000
    diffusion_semantic_kwargs:
        att_1: 0.99999 
        att_T: 0.000009
        ctt_1: 0.000009
        ctt_T: 0.99999
        model_output_type: 'x0'
        mask_weight: 1
        auxiliary_loss_weight: 0.0005
        adaptive_auxiliary_loss: True
    diffusion_geometric_kwargs:
        schedule_type: 'linear'
        beta_start: 0.0001
        beta_end: 0.02
        loss_type: 'mse'
        model_mean_type: 'eps'
        model_var_type: 'fixedsmall'

    # denoising net
    net_type: "transformer"
    net_kwargs:
        seperate_all: True
        n_layer: 8
        n_embd: 512
        n_head: 8
        dim_feedforward: 2048
        dropout: 0.1
        activate: 'GELU'
        timestep_type: 'adalayernorm_abs'
        mlp_type: 'fc'

feature_extractor:
    name: "pointnet_simple"
    feat_units: [4, 64, 64, 512, 64]

training:
    splits: ["train", "val"]
    epochs: 100000
    batch_size: 128
    save_frequency: 2000
    max_grad_norm: 10
    # optimizer
    optimizer: Adam
    weight_decay: 0.0
    # schedule
    schedule: 'step'
    lr: 0.0002
    lr_step: 15000
    lr_decay: 0.5

validation:
    splits: ["test"]
    frequency: 100
    save_frequency: 1000
    batch_size: 128

logger:
    type: "wandb"
    project: "MiDiffusion"