models.transformer

DETR Transformer class.

Copy-paste from torch.nn.Transformer with modifications:

• positional encodings are passed in MHattention

• extra LN at the end of encoder is removed

• decoder returns a stack of activations from all decoding layers