[pytorch] detr源码浅析

为之后SAM的代码分析做铺垫

1. backbone部分

• detr.py中的DETR class

class DETR(nn.Module):
	def __init__(self, backbone, transformer, num_classes, num_queries, aux_loss=False):
		...
	def forward(self, samples: NestedTensor):
		features, pos = self.backbone(samples)
        # 第一步，从图像提取特征
        # 返回值：特征图，pos位置编码（当前得到的特征图编码.不是对原始图像）
        # 跳到backbone - backbone.py里的Joiner函数

• backbone.py 中的 Joiner class

class Joiner(nn.Sequential):
    def __init__(self, backbone, position_embedding):
        super().__init__(backbone, position_embedding)

    def forward(self, tensor_list: NestedTensor):
        # print(tensor_list.tensor.shape)
        xs = self[0](tensor_list)
        # 输入图像经过resnet
        # 得到特征图
        out: List[NestedTensor] = []
        pos = []
        for name, x in xs.items():
            out.append(x)
            pos.append(self[1](x).to(x.tensors.dtype))
            # 跳到position encoding.py

        return out, pos

• position encoding.py
  第一种：Attention Is All You Need中的正余弦编码方式，不用学习，默认方法
  

class PositionEmbeddingSine(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """
    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
        super().__init__()
        self.num_pos_feats = num_pos_feats
        self.temperature = temperature # 经验值
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

    def forward(self, tensor_list: NestedTensor):
        x = tensor_list.tensors
        # 输入特征图大小 batch, c, h, w    resnet50 c = 2048
        mask = tensor_list.mask
        # mask 表示实际的特征true  还是padding出来的false
        # 大小 batch，h，w
        assert mask is not None
        not_mask = ~mask
        y_embed = not_mask.cumsum(1, dtype=torch.float32) # 行方向累加 最后一位为累加的得到的最大值
        x_embed = not_mask.cumsum(2, dtype=torch.float32) # 列方向累加 最后一位为累加的得到的最大值
        if self.normalize:
            eps = 1e-6
            # 归一化
            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        # 算奇数维度或者偶数维度 公式不一样
        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4)