yolov5中head修改為decouple head詳解
yolov5的head修改為decouple head
yolox的decoupled head結構
本來想將yolov5的head修改為decoupled head，與yolox的decouple head對齊，但是沒註意，該成瞭如下結構：
感謝少年肩上楊柳依依的指出，如還有問題歡迎指出
1.修改models下的yolo.py文件中的Detect
class Detect(nn.Module):
stride = None # strides computed during build
onnx_dynamic = False # ONNX export parameter
def __init__(self, nc=80, anchors=(), ch=(), inplace=True): # detection layer
super().__init__()
self.nc = nc # number of classes
self.no = nc + 5 # number of outputs per anchor
self.nl = len(anchors) # number of detection layers
self.na = len(anchors[0]) // 2 # number of anchors
self.grid = [torch.zeros(1)] * self.nl # init grid
self.anchor_grid = [torch.zeros(1)] * self.nl # init anchor grid
self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2)) # shape(nl,na,2)
# self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
self.m_box = nn.ModuleList(nn.Conv2d(256, 4 * self.na, 1) for x in ch) # output conv
self.m_conf = nn.ModuleList(nn.Conv2d(256, 1 * self.na, 1) for x in ch) # output conv
self.m_labels = nn.ModuleList(nn.Conv2d(256, self.nc * self.na, 1) for x in ch) # output conv
self.base_conv = nn.ModuleList(BaseConv(in_channels = x, out_channels = 256, ksize = 1, stride = 1) for x in ch)
self.cls_convs = nn.ModuleList(BaseConv(in_channels = 256, out_channels = 256, ksize = 3, stride = 1) for x in ch)
self.reg_convs = nn.ModuleList(BaseConv(in_channels = 256, out_channels = 256, ksize = 3, stride = 1) for x in ch)
# self.m = nn.ModuleList(nn.Conv2d(x, 4 * self.na, 1) for x in ch, nn.Conv2d(x, 1 * self.na, 1) for x in ch,nn.Conv2d(x, self.nc * self.na, 1) for x in ch)
self.inplace = inplace # use in-place ops (e.g. slice assignment)self.ch = ch
def forward(self, x):
z = [] # inference output
for i in range(self.nl):
# # x[i] = self.m[i](x[i]) # convs
# print("&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&", i)
# print(x[i].shape)
# print(self.base_conv[i])
# print("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
x_feature = self.base_conv[i](x[i])
# x_feature = x[i]
cls_feature = self.cls_convs[i](x_feature)
reg_feature = self.reg_convs[i](x_feature)
# reg_feature = x_feature
m_box = self.m_box[i](reg_feature)
m_conf = self.m_conf[i](reg_feature)
m_labels = self.m_labels[i](cls_feature)
x[i] = torch.cat((m_box,m_conf, m_labels),1)
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
if not self.training: # inference
if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
y = x[i].sigmoid()
if self.inplace:
y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i] # xy
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
y = torch.cat((xy, wh, y[..., 4:]), -1)
z.append(y.view(bs, -1, self.no))
return x if self.training else (torch.cat(z, 1), x)
2.在yolo.py中添加
def get_activation(name="silu", inplace=True):
if name == "silu":
module = nn.SiLU(inplace=inplace)
elif name == "relu":
module = nn.ReLU(inplace=inplace)
elif name == "lrelu":
module = nn.LeakyReLU(0.1, inplace=inplace)
else:
raise AttributeError("Unsupported act type: {}".format(name))
return module
class BaseConv(nn.Module):
"""A Conv2d -> Batchnorm -> silu/leaky relu block"""
def __init__(
self, in_channels, out_channels, ksize, stride, groups=1, bias=False, act="silu"
):
super().__init__()
# same padding
pad = (ksize - 1) // 2
self.conv = nn.Conv2d(
in_channels,
out_channels,
kernel_size=ksize,
stride=stride,
padding=pad,
groups=groups,
bias=bias,
)
self.bn = nn.BatchNorm2d(out_channels)
self.act = get_activation(act, inplace=True)
def forward(self, x):
# print(self.bn(self.conv(x)).shape)
return self.act(self.bn(self.conv(x)))
# return self.bn(self.conv(x))
def fuseforward(self, x):
return self.act(self.conv(x))
decouple head的特點：
由於訓練模型時，應該是channels = 256的地方改成瞭channels = x(失誤)，所以在decoupled head的部分參數量比yolox要大一些，以下的結果是在channels= x的情況下得出
比yolov5s參數多，計算量大，在我自己的2.5萬的數據量下map提升瞭3%多
1.模型給出的目標cls較高，需要將conf的閾值設置較大（0.5），不然準確率較低
parser.add_argument('--conf-thres', type=float, default=0.5, help='confidence threshold')
2.對於少樣本的檢測效果較好，召回率的提升比準確率多
3.在conf設置為0.25時，召回率比yolov5s高，但是準確率低；在conf設置為0.5時，召回率與準確率比yolov5s高
4.比yolov5s參數多，計算量大，在2.5萬的數據量下map提升瞭3%多
對於decouple head的改進
改進：
1.將紅色框中的conv去掉，縮小參數量和計算量；
2.channels =256 ，512 ，1024是考慮不增加參數，不進行featuremap的信息壓縮
class Detect(nn.Module):
stride = None # strides computed during build
onnx_dynamic = False # ONNX export parameter
def __init__(self, nc=80, anchors=(), ch=(), inplace=True): # detection layer
super().__init__()
self.nc = nc # number of classes
self.no = nc + 5 # number of outputs per anchor
self.nl = len(anchors) # number of detection layers
self.na = len(anchors[0]) // 2 # number of anchors
self.grid = [torch.zeros(1)] * self.nl # init grid
self.anchor_grid = [torch.zeros(1)] * self.nl # init anchor grid
self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2)) # shape(nl,na,2)
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
self.inplace = inplace # use in-place ops (e.g. slice assignment)
def forward(self, x):
z = [] # inference output
for i in range(self.nl):
x[i] = self.m[i](x[i]) # conv
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
if not self.training: # inference
if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
y = x[i].sigmoid()
if self.inplace:
y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i] # xy
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i] # xy
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
y = torch.cat((xy, wh, y[..., 4:]), -1)
z.append(y.view(bs, -1, self.no))
return x if self.training else (torch.cat(z, 1), x)
特點
1.模型給出的目標cls較高，需要將conf的閾值設置較大（0.4），不然準確率較低
2.對於少樣本的檢測效果較好，準確率的提升比召回率多
3. 準確率的提升比召回率多，
該改進不如上面的模型提升多，但是參數量小，計算量小少9Gflop，占用顯存少
decoupled head指標提升的原因：由於yolov5s原本的head不能完全的提取featuremap中的信息，decoupled head能夠較為充分的提取featuremap的信息；
疑問
為什麼decoupled head目標的cls會比較高，沒想明白
為什麼去掉base_conv，召回率要比準確率提升少
總結
到此這篇關於yolov5中head修改為decouple head的文章就介紹到這瞭,更多相關yolov5 head修改為decouple head內容請搜索WalkonNet以前的文章或繼續瀏覽下面的相關文章希望大傢以後多多支持WalkonNet！
推薦閱讀：
- 通過底層源碼理解YOLOv5的Backbone
- YOLOv5改進之添加SE註意力機制的詳細過程
- pytorch查看網絡參數顯存占用量等操作
- PyTorch零基礎入門之構建模型基礎
- Pytorch實現ResNet網絡之Residual Block殘差塊