Could you tell me about auxiliary inputs in hash encodings

[HWP97]

Hi, I'm a student who just started to learn computer vision. I'm very impressed for your research and felt overwhelmingly grateful for pytorch bindings for your work. I got curious while using your hash encodings for NeRF. And I found there is no auxiliary inputs(in paper) in your hash encodings. I know that the auxiliary inputs varies from tasks to tasks. But due to my poor understandings for encodings and python codings, I can't figure out to implement codes as you explained in your paper(T.Muller. et al. 'Instant Neural Graphics Primitives with a Multiresolution Hash Encoding'). I wonder if you can tell me some details about implementation of auxiliary inputs. thank you so much!

[Tom94]

Hi, you can compose multiple encodings (and thereby achieve auxiliary inputs) using the Composite encoding in the json config passed to tiny-cuda-nn.

Here's the encoding config we used for neural radiance caching:

    "encoding": {
        "otype": "Composite",
        "nested": [
            {
                "n_dims_to_encode": 3, # Position
                "otype": "HashGrid",
                "per_level_scale": 2.0,
                "log2_hashmap_size": 15,
                "base_resolution": 16,
                "n_levels": 16
            },
            {
                "n_dims_to_encode": 5, # Interesting conditionals
                "otype": "OneBlob",
                "n_bins": 4
            },
            {
                "n_dims_to_encode": 6, # Linear conditionals that should be identity encoded
                "otype": "Identity"
            }
        ]
    },

[HWP97]

Your reply helped me a lot. Thank you!