base.yaml

# Default config for the whole project

# Outermost level configuration switches
exp_name: base
mocking: False
detect_anomaly: False
fix_random: False
allow_tf32: True
deterministic: False # deterministic training (debug only)
benchmark: True # when benchmarking, slow, after, fast
profiler_cfg:
    enabled: False # no profiling by default
print_model: False # defaults to a compact interface
preparing_parser: False
print_parameters: True

# Top level model building
model_cfg:
    type: VolumetricVideoModel
    camera_cfg:
        type: NoopCamera
    sampler_cfg:
        type: ImportanceSampler
    network_cfg:
        type: MultilevelNetwork
        # <<: *network_cfg # is this ok?
        parameterizer_cfg:
            type: ContractRegressor
            in_dim: 3
        xyzt_embedder_cfg:
            type: ComposedXyztEmbedder
            xyz_embedder_cfg:
                type: PositionalEncodingEmbedder
                multires: 10
            t_embedder_cfg:
                type: LatentCodeEmbedder
        xyz_embedder_cfg:
            type: EmptyEmbedder
        dir_embedder_cfg:
            type: PositionalEncodingEmbedder
            multires: 4
        rgb_embedder_cfg:
            type: EmptyEmbedder
        deformer_cfg:
            type: EmptyRegressor
        geometry_cfg:
            type: SplitRegressor
            width: 512
            depth: 8
        appearance_cfg:
            type: MlpRegressor
            width: 256
            depth: 2
            out_dim: 3
            out_actvn: sigmoid
        network_cfgs:
            # - &network_cfg # coarse network configuration
            - type: VolumetricVideoNetwork
              geometry_cfg:
                  type: SplitRegressor
                  width: 128
                  depth: 4
              appearance_cfg:
                  type: EmptyRegressor
            # - <<: *network_cfg # fine network configuration
            - type: VolumetricVideoNetwork
              # seems to be hierarchically overwritting, good
    renderer_cfg:
        type: VolumeRenderer
    supervisor_cfg:
        type: SequentialSupervisor

dataloader_cfg: &dataloader_cfg # we see the term "dataloader" a one word?
    type: VolumetricVideoDataloader
    dataset_cfg: &dataset_cfg
        type: VolumetricVideoDataset
        split: TRAIN
        data_root: data/dataset/sequence
        n_rays: 512
        supply_decoded: True # pass the image to the network directly
        encode_ext: .png # save memory during training
        frame_sample: [0, null, 1]
        view_sample: [0, null, 1]
        intri_file: intri.yml
        extri_file: extri.yml
        bodymodel_file: output/cfg_exp.yml
        motion_file: 'motion.npz'
        append_gt_prob: 0.1
        extra_src_pool: 1
        bounds: [[-10.0, -10.0, -10.0], [10.0, 10.0, 10.0]]
    sampler_cfg:
        type: RandomSampler
        frame_sample: [0, null, 1]
        view_sample: [0, null, 1]
    batch_sampler_cfg:
        type: BatchSampler
        batch_size: 8

val_dataloader_cfg: # we see the term "dataloader" a one word?
    <<: *dataloader_cfg
    max_iter: -1
    dataset_cfg:
        <<: *dataset_cfg
        type: WillChangeToNoopIfGUIDataset
        split: VAL
        supply_decoded: True # pass the image to the network directly
        encode_ext: .png # save bandwidth for rendering
        frame_sample: [0, null, 50]
        view_sample: [0, null, 5]
        append_gt_prob: 1.0
        extra_src_pool: 0
    sampler_cfg:
        type: SequentialSampler
        frame_sample: [0, null, 1]
        view_sample: [0, null, 1]
        # Please modify dataset_cfg instead of this
    batch_sampler_cfg:
        type: BatchSampler
        batch_size: 1

runner_cfg: &runner_cfg
    type: VolumetricVideoRunner
    epochs: 400
    # decay_epochs: 400
    ep_iter: 500
    optimizer_cfg:
        type: ConfigurableOptimizer
    scheduler_cfg:
        type: ExponentialLR
    moderator_cfg:
        type: NoopModerator
    visualizer_cfg:
        type: VolumetricVideoVisualizer
        types: [RENDER, DEPTH, ALPHA]
        result_dir: data/result
        save_tag: ''
    evaluator_cfg:
        type: VolumetricVideoEvaluator
    recorder_cfg:
        type: TensorboardRecorder

viewer_cfg:
    type: VolumetricVideoViewer