configs/papers/define/scannet_temporal_test_context_1.yaml

wrapper:
    seed: 42
    min_epochs: 0
    max_epochs: 0
    find_unused_parameters: True
    flip_lr_prob: 0.5
    validate_first: True
    validate_flipped: False
    sync_batch_norm: True
evaluation:
    rgb:
        only_first: False
    depth:
        crop: ''
        min_depth: 0.2
        max_depth: 10.0
        scale_output: resize
        median_scaling: True
        post_process: False
arch:
    model:
        checkpoint: /data/vidar/models/scannet_full.ckpt
        file: perceiver/DefineGenericModel
        use_pose_noise: []
        use_virtual_cameras: []
        virtual_cameras_eval: False
        use_virtual_rgb: False
        augment_canonical: False
        encode_train: all
        encode_eval: all
        decode_train: all
        decode_eval: all
        decode_encodes: False
        task_weights: [1.0,1.0]
        scale_loss: True
        sample_decoded_queries: 0.0
        network:
            tasks: [depth]
            depth_range: [0.1,10.]
            image_shape: [128,192]
            num_bands_orig: 20
            num_bands_dirs: 10
            max_resolution_orig: 60
            max_resolution_dirs: 60
            to_world: True
            d_latents: 512
            num_latents: 1024
            hidden_dropout_prob: 0.25
            num_cross_attention_heads: 1
            num_self_attends_per_block: 8
            num_self_attention_heads: 8
            decoder_num_heads: 1
            rgb_feat_dim: 960
            rgb_feat_type: resnet_all
            downsample_encoder: 4
            downsample_decoder: 4
            upsample_convex: 'convex'
            encoder_with_rgb: True
            decoder_with_rgb: False
            output_mode: inv_depth
            sample_encoding_rays: 0
            with_monodepth: False
datasets:
    validation:
        name: [ScanNetTemporal]
        path: [/data/vidar/scannet_processed]
        split: [test]
        context: [1]
        stride: [10]
        labels: [depth,pose]
        cameras: [[0]]
        dataloader:
            batch_size: 1
            pin_memory: True
            num_workers: 4
        augmentation:
            resize: [128,192]
            resize_supervision: True
            preserve_depth: True