| name: u_vit3d_pose | |
| channels: [128, 256, 512, 1024] | |
| emb_channels: 1024 | |
| patch_size: 2 | |
| block_types: ["ResBlock", "ResBlock", "TransformerBlock", "TransformerBlock"] | |
| block_dropouts: [0.0, 0.0, 0.1, 0.1] | |
| num_updown_blocks: [3, 3, 3] | |
| num_mid_blocks: 16 | |
| num_heads: 4 | |
| pos_emb_type: rope | |
| use_checkpointing: [false, false, false, false] | |
| conditioning: | |
| dim: null # this will be overwritten at dfot_video_3d.py | |
| external_cond_dropout: 0.1 # probability of dropping a camera pose of each video during training |