depth_align_embs is not saved or optimized

[LIJUNCHENG001]

According to the paper: we apply the learnable queries [Qt1, Qt2, Qt3] corresponding to three-view operational images.
However, in the code, depth_align_embs is defined as:

self.depth_align_embs = nn.Parameter(torch.randn(...)).to(torch.bfloat16)

Parameter.to() returns a regular Tensor, so it is absent from named_parameters(), optimizer groups, and checkpoints. It is randomly
recreated during loading, making depth output seed-dependent.

I have checked the pretrained model weights, depth_align_embs is indeed not saved.

Is this a bug or an intended design?

[PushpakAg]

It seems like depth_align_embs isn't being saved due to how it's defined. I'd check if its absence from named_parameters() is causing any unexpected behavior during training or inference.