Confusion about the query_embedding layer in DETR

[baibaidj]

When I read the code of DETR, I found that the self.query_embedding was created as a nn.Embedding layer but only its weight was used in the model. I was a bit confused. See here:

mmdetection/mmdet/models/dense_heads/detr_head.py

Line 259 in 1a90fa8

outs_dec, _ = self.transformer(x, masks, self.query_embedding.weight,

If only the weight is needed, why not initialize the self.query_embedding by nn.Parameter(num_queries, embed_dims)?

[Li-Qingyun]

if it is initialized by nn.Parameter(num_queries, embed_dims)
it may need registry