`constant_tensor` in `BevDecoder`

[LIIHWF]

Hi,

I noticed that the attribute constant_tensor is randomly assigned in the class BevDecoder, and BevDecoder is enabled in Mile by default.

mile/mile/models/common.py

Line 197 in c8b245c

self.constant_tensor = nn.Parameter(torch.randn((n_channels, *constant_size), dtype=torch.float32))

I wonder what the role of this attribute is, and is it necessary to make its value random?

Introducing randomness into the system may also introduce unknown risks.

[anthonyhu]

This parameter corresponds to the learnable latent grid in bird's-eye view, which is upsampled by the BeV decoder and injects the signal from the input with instance normalization. The parameter is randomly initialised, similarly to a weight in the neural net, and is optimised over the course of training