请详细解释DETR的decoder过程，必要可以给出代码或者图片来解释

DETR（对抗性变换检测）的解码器过程是一种在计算机视觉任务中用于检测物体的技术。它使用一个可以并行计算的非常深的神经网络，能够更快地检测出图像中的多个物体。解码器是一个编码器-解码器网络，它接收输入图像，然后编码它，以提取该图像的特征，然后给出检测结果。解码器的代码如下：class DETRDecoder(nn.Module):
def __init__(self, num_classes, num_queries):
super().__init__()
self.num_queries = num_queries
self.cls_embed = nn.Linear(num_classes, num_queries)
self.bbox_embed = nn.Linear(4, num_queries)
self.query_embed = nn.Linear(256, num_queries)
self.cls_logits = nn.Linear(num_queries, num_classes)
self.bbox_delta = nn.Linear(num_queries, 4)

def forward(self, features):
# query embedding
query = self.query_embed(features) # (B, num_queries, 256)
query = query.mean(dim=1) # (B, num_queries)
# class embedding
cls_embed = self.cls_embed.weight # (num_classes, num_queries)
cls_logits = torch.einsum('bq,cq->bc', [query, cls_embed]) # (B, num_classes)
# bbox embedding
bbox_embed = self.bbox_embed.weight # (4, num_queries)
bbox_delta = torch.einsum('bq,ab->ba', [query, bbox_embed]) # (B, 4)
return {'cls_logits': cls_logits, 'bbox_delta': bbox_delta}