两个不同尺度的特征进行融合时x1(2,128,640,640)x2(2,512,320,320),他们不应该是一视同仁的,一定有重要的不重要的,那就要找出权重来进行重新分配,融合成新的特征 x3=ax1’+bx2’
#--------------------------------------------#
# 上采样+通道变化(上采样倍率默认为2)
#--------------------------------------------#
class Upsample(nn.Module):
"""Applies convolution followed by upsampling."""
def __init__(self, c1, c2, scale_factor=2):
super().__init__()
# self.cv1 = Conv(c1, c2, 1)
# self.upsample = nn.Upsample(scale_factor=scale_factor, mode='nearest') # or model='bilinear' non-deterministic
if scale_factor == 2:
self.cv1 = nn.ConvTranspose2d(c1, c2, 2, 2, 0, bias=True) # 如果下采样率为2,就用Stride为2的2×2卷积来实现2次下采样
elif scale_factor == 4:
self.cv1 = nn.ConvTranspose2d(c1, c2, 4, 4, 0, bias=True) # 如果下采样率为4,就用Stride为4的4×4卷积来实现4次下采样
def forward(self, x):
# return self.upsample(self.cv1(x))
return self.cv1(x)
#--------------------------------------------#
# 自适应空间融合(ASFF)
# eg:以 level1 为例子
# 如果特征x1,x2要进行ASFF 融合后到level1层
# 1.需要x2变成与x1通道,空间形状完全一样的x2'
# 2.之后将x1,x2'通道数分别进行降维到8,
# 3.将压缩后的两个特征进行拼接后进行通道降维到2 ([b,2,h,w])
# 4.此时就已经取得了x1,x2'两个特征的不同权重
# 5.将权重拆分为两部分 分别乘x1,x2',得到的特征再求和
# 6.最后经过一层卷积将融合的新特征的形状变得和x1一样
# (同理 level2 需要变得和x2形状一样)
#--------------------------------------------#
class ASFF(nn.Module):
"""ASFF2 module for YOLO AFPN head https://arxiv.org/abs/2306.15988"""
def __init__(self, c1, level=0):
super().__init__()
#-----------------------------------#
# c1 :[c1,c2] 输入特征x1,x2 的通道
# level :融合到哪一层中去
# inter_dim :最终输出的通道数
# compress_c :中间被压缩的通道数
#-----------------------------------#
c1, c2 = c1[0], c1[1]
self.level = level
self.dim = c1, c2
self.inter_dim = self.dim[self.level]
compress_c = 8
#------------------------------------#
# level1 x2变为和x1形状相同
# level2 x1变为和x2形状相同
#------------------------------------#
if level == 0:
self.stride_level_1 = Upsample(c2, self.inter_dim)
if level == 1:
self.stride_level_0 = Conv(c1, self.inter_dim, 2, 2, 0) # stride=2 下采样为2倍
#------------------------------------#
# 分别对两特征通道降维
#------------------------------------#
self.weight_level_0 = Conv(self.inter_dim, compress_c, 1, 1)
self.weight_level_1 = Conv(self.inter_dim, compress_c, 1, 1)
#------------------------------------#
# 将降维后的特征合并并降维到2,得到每个特征的权重
#------------------------------------#
self.weights_levels = nn.Conv2d(compress_c * 2, 2, kernel_size=1, stride=1, padding=0)
self.conv = Conv(self.inter_dim, self.inter_dim, 3, 1)
def forward(self, x):
#--------------------------------#
# 获得两个特征
#--------------------------------#
x0, x1 = x[0], x[1]
#--------------------------------#
# 根据要融合到不同的level x1/x2进行形状变化
#--------------------------------#
if self.level == 0:
level_0_resized = x0
level_1_resized = self.stride_level_1(x1)
elif self.level == 1:
level_0_resized = self.stride_level_0(x0)
level_1_resized = x1
#----------------------------------#
# 通道降维
#----------------------------------#
level_0_weight_v = self.weight_level_0(level_0_resized)
level_1_weight_v = self.weight_level_1(level_1_resized)
#----------------------------------#
# 特征合并并降维到2,得到不同特征的权重
#----------------------------------#
levels_weight_v = torch.cat((level_0_weight_v, level_1_weight_v), 1)
levels_weight = self.weights_levels(levels_weight_v)
#----------------------------------#
# 权重归一化
#----------------------------------#
levels_weight = F.softmax(levels_weight, dim=1)
#----------------------------------#
# 将权重分配到不同的特征层
#----------------------------------#
fused_out_reduced = level_0_resized * levels_weight[:, 0:1] + level_1_resized * levels_weight[:, 1:2]
return self.conv(fused_out_reduced)
model=ASFF2([128,512],0)
a=torch.Tensor(2,128,640,640)
b=torch.Tensor(2,512,320,320)
c=model([a,b])
print(c.shape)