out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False,
),
)
result.add_module("bn", nn.BatchNorm2d(num_features=out_channels))
return result
class DepthWiseConv(nn.Module): def __init__(self, inc, kernel_size, stride=1): super().init() padding = 1 if kernel_size == 1: padding = 0 # self.conv = nn.Sequential( # nn.Conv2d(inc, inc, kernel_size, stride, padding, groups=inc, bias=False,), # nn.BatchNorm2d(inc), # ) self.conv = conv_bn(inc, inc, kernel_size, stride, padding, inc)
def forward(self, x):
return self.conv(x)
class PointWiseConv(nn.Module): def __init__(self, inc, outc): super().init() # self.conv = nn.Sequential( # nn.Conv2d(inc, outc, 1, 1, 0, bias=False), # nn.BatchNorm2d(outc), # ) self.conv = conv_bn(inc, outc, 1, 1, 0)
def forward(self, x):
return self.conv(x)
class MobileOneBlock(nn.Module): def __init__( self, in_channels, out_channels, k, stride=1, dilation=1, padding_mode="zeros", deploy=False, use_se=False, ): super(MobileOneBlock, self).init() self.deploy = deploy self.in_channels = in_channels self.out_channels = out_channels self.deploy = deploy kernel_size = 3 padding = 1 assert kernel_size == 3 assert padding == 1 self.k = k padding_11 = padding - kernel_size // 2
self.nonlinearity = nn.ReLU()
if use_se:
# self.se = SEBlock(out\_channels, internal\_neurons=out\_channels // 16)
...
else:
self.se = nn.Identity()
if deploy:
self.dw_reparam = nn.Conv2d(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=in_channels,
bias=True,
padding_mode=padding_mode,
)
self.pw_reparam = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
bias=True,
)
else:
self.dw_bn_layer = (
nn.BatchNorm2d(in_channels)
if out_channels == in_channels and stride == 1
else None
)
for k_idx in range(k):
setattr(
self,
f"dw\_3x3\_{k\_idx}",
DepthWiseConv(in_channels, 3, stride=stride),
)
self.dw_1x1 = DepthWiseConv(in_channels, 1, stride=stride)
self.pw_bn_layer = (
nn.BatchNorm2d(in_channels)
if out_channels == in_channels and stride == 1
else None
)
for k_idx in range(k):
setattr(
self, f"pw\_1x1\_{k\_idx}", PointWiseConv(in_channels, out_channels)
)
def forward(self, inputs):
if self.deploy:
x = self.dw_reparam(inputs)
x = self.nonlinearity(x)
x = self.pw_reparam(x)
x = self.nonlinearity(x)
return x
if self.dw_bn_layer is None:
id_out = 0
else:
id_out = self.dw_bn_layer(inputs)
x_conv_3x3 = []
for k_idx in range(self.k):
x = getattr(self, f"dw\_3x3\_{k\_idx}")(inputs)
# print(x.shape)
x_conv_3x3.append(x)
x_conv_1x1 = self.dw_1x1(inputs)
# print(x\_conv\_1x1.shape, x\_conv\_3x3[0].shape)
# print(x\_conv\_1x1.shape)
# print(id\_out)
x = id_out + x_conv_1x1 + sum(x_conv_3x3)
x = self.nonlinearity(self.se(x))
# 1x1 conv
if self.pw_bn_layer is None:
id_out = 0
else:
id_out = self.pw_bn_layer(x)
x_conv_1x1 = []
for k_idx in range(self.k):
x_conv_1x1.append(getattr(self, f"pw\_1x1\_{k\_idx}")(x))
x = id_out + sum(x_conv_1x1)
x = self.nonlinearity(x)
return x
# Optional. This improves the accuracy and facilitates quantization.
# 1. Cancel the original weight decay on rbr\_dense.conv.weight and rbr\_1x1.conv.weight.
# 2. Use like this.
# loss = criterion(....)
# for every RepVGGBlock blk:
# loss += weight\_decay\_coefficient \* 0.5 \* blk.get\_cust\_L2()
# optimizer.zero\_grad()
# loss.backward()
def get\_custom\_L2(self):
# K3 = self.rbr\_dense.conv.weight
# K1 = self.rbr\_1x1.conv.weight
# t3 = (self.rbr\_dense.bn.weight / ((self.rbr\_dense.bn.running\_var + self.rbr\_dense.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
# t1 = (self.rbr\_1x1.bn.weight / ((self.rbr\_1x1.bn.running\_var + self.rbr\_1x1.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
# l2\_loss\_circle = (K3 \*\* 2).sum() - (K3[:, :, 1:2, 1:2] \*\* 2).sum() # The L2 loss of the "circle" of weights in 3x3 kernel. Use regular L2 on them.
# eq\_kernel = K3[:, :, 1:2, 1:2] \* t3 + K1 \* t1 # The equivalent resultant central point of 3x3 kernel.
# l2\_loss\_eq\_kernel = (eq\_kernel \*\* 2 / (t3 \*\* 2 + t1 \*\* 2)).sum() # Normalize for an L2 coefficient comparable to regular L2.
# return l2\_loss\_eq\_kernel + l2\_loss\_circle
...
# This func derives the equivalent kernel and bias in a DIFFERENTIABLE way.
# You can get the equivalent kernel and bias at any time and do whatever you want,
# for example, apply some penalties or constraints during training, just like you do to the other models.
# May be useful for quantization or pruning.
def get\_equivalent\_kernel\_bias(self):
# kernel3x3, bias3x3 = self.\_fuse\_bn\_tensor(self.rbr\_dense)
# kernel1x1, bias1x1 = self.\_fuse\_bn\_tensor(self.rbr\_1x1)
# kernelid, biasid = self.\_fuse\_bn\_tensor(self.rbr\_identity)
# return kernel3x3 + self.\_pad\_1x1\_to\_3x3\_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
dw_kernel_3x3 = []
dw_bias_3x3 = []
for k_idx in range(self.k):
k3, b3 = self._fuse_bn_tensor(getattr(self, f"dw\_3x3\_{k\_idx}").conv)
# print(k3.shape, b3.shape)
dw_kernel_3x3.append(k3)
dw_bias_3x3.append(b3)
dw_kernel_1x1, dw_bias_1x1 = self._fuse_bn_tensor(self.dw_1x1.conv)
dw_kernel_id, dw_bias_id = self._fuse_bn_tensor(
self.dw_bn_layer, self.in_channels
)
dw_kernel = (
sum(dw_kernel_3x3)
+ self._pad_1x1_to_3x3_tensor(dw_kernel_1x1)
+ dw_kernel_id
)
dw_bias = sum(dw_bias_3x3) + dw_bias_1x1 + dw_bias_id
# pw
pw_kernel = []
pw_bias = []
for k_idx in range(self.k):
k1, b1 = self._fuse_bn_tensor(getattr(self, f"pw\_1x1\_{k\_idx}").conv)
# print(k1.shape)
pw_kernel.append(k1)
pw_bias.append(b1)
pw_kernel_id, pw_bias_id = self._fuse_bn_tensor(self.pw_bn_layer, 1)
pw_kernel_1x1 = sum(pw_kernel) + pw_kernel_id
pw_bias_1x1 = sum(pw_bias) + pw_bias_id
return dw_kernel, dw_bias, pw_kernel_1x1, pw_bias_1x1
def \_pad\_1x1\_to\_3x3\_tensor(self, kernel1x1):
if kernel1x1 is None:
return 0
else:
return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1])
def \_fuse\_bn\_tensor(self, branch, groups=None):
if branch is None:
return 0, 0
if isinstance(branch, nn.Sequential):
kernel = branch.conv.weight
bias = branch.conv.bias
running_mean = branch.bn.running_mean
running_var = branch.bn.running_var
gamma = branch.bn.weight
beta = branch.bn.bias
eps = branch.bn.eps
else:
assert isinstance(branch, nn.BatchNorm2d)
# if not hasattr(self, 'id\_tensor'):
input_dim = self.in_channels // groups # self.groups
if groups == 1:
ks = 1
else:
ks = 3
kernel_value = np.zeros(
(self.in_channels, input_dim, ks, ks), dtype=np.float32
)
for i in range(self.in_channels):
if ks == 1:
kernel_value[i, i % input_dim, 0, 0] = 1
else:
kernel_value[i, i % input_dim, 1, 1] = 1
self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
kernel = self.id_tensor
running_mean = branch.running_mean
running_var = branch.running_var
gamma = branch.weight
beta = branch.bias
eps = branch.eps
std = (running_var + eps).sqrt()
t = (gamma / std).reshape(-1, 1, 1, 1)
return kernel \* t, beta - running_mean \* gamma / std
def switch\_to\_deploy(self):
dw_kernel, dw_bias, pw_kernel, pw_bias = self.get_equivalent_kernel_bias()
self.dw_reparam = nn.Conv2d(
in_channels=self.pw_1x1_0.conv.conv.in_channels,
out_channels=self.pw_1x1_0.conv.conv.in_channels,
kernel_size=self.dw_3x3_0.conv.conv.kernel_size,
stride=self.dw_3x3_0.conv.conv.stride,
padding=self.dw_3x3_0.conv.conv.padding,
groups=self.dw_3x3_0.conv.conv.in_channels,
bias=True,
)
self.pw_reparam = nn.Conv2d(
in_channels=self.pw_1x1_0.conv.conv.in_channels,
out_channels=self.pw_1x1_0.conv.conv.out_channels,
kernel_size=1,
stride=1,
bias=True,
)
self.dw_reparam.weight.data = dw_kernel
self.dw_reparam.bias.data = dw_bias
self.pw_reparam.weight.data = pw_kernel
self.pw_reparam.bias.data = pw_bias
for para in self.parameters():
para.detach_()
self.__delattr__("dw\_1x1")
for k_idx in range(self.k):
self.__delattr__(f"dw\_3x3\_{k\_idx}")
self.__delattr__(f"pw\_1x1\_{k\_idx}")
if hasattr(self, "dw\_bn\_layer"):
self.__delattr__("dw\_bn\_layer")
if hasattr(self, "pw\_bn\_layer"):
self.__delattr__("pw\_bn\_layer")
if hasattr(self, "id\_tensor"):
self.__delattr__("id\_tensor")
self.deploy = True
class MobileOneNet(nn.Module): def __init__( self, blocks, ks, channels, strides, width_muls, num_classes=None, deploy=False ): super().init()
self.stage_num = len(blocks)
# self.stage0 = MobileOneBlock(3, int(channels[0] \* width\_muls[0]), ks[0], stride=strides[0], deploy=deploy)
self.stage0 = nn.Sequential(
nn.Conv2d(3, int(channels[0] \* width_muls[0]), 3, 2, 1, bias=False),
nn.BatchNorm2d(int(channels[0] \* width_muls[0])),
nn.ReLU(),
)
in_channels = int(channels[0] \* width_muls[0])
for idx, block_num in enumerate(blocks[1:]):
idx += 1
module = []
out_channels = int(channels[idx] \* width_muls[idx])
for b_idx in range(block_num):
stride = strides[idx] if b_idx == 0 else 1
block = MobileOneBlock(
in_channels, out_channels, ks[idx], stride, deploy=deploy
)
in_channels = out_channels
module.append(block)
setattr(self, f"stage{idx}", nn.Sequential(\*module))
if num_classes is not None:
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc1 = nn.Sequential(
nn.Linear(
out_channels,
num_classes,
),
)
def forward(self, x):
# for s\_idx in range(self.stage\_num):
# x = getattr(self, f'stage{s\_idx}')(x)
x0 = self.stage0(x)
# print(x0[0,:,0,0])
# return x0
x1 = self.stage1(x0)
x2 = self.stage2(x1)
x3 = self.stage3(x2)
x4 = self.stage4(x3)
x5 = self.stage5(x4)
assert x5.shape[-1] == 7
x = self.avg_pool(x5)
x = torch.flatten(x, start_dim=1) # b, c
x = self.fc1(x)
return x
### 2.3 修改tasks.py
[详细的改进教程以及源码，戳这！戳这！！戳这！！！B站：AI学术叫叫兽 源码在相簿的链接中，动态中也有链接，感谢支持！祝科研遥遥领先！]( )
## 三、验证是否成功即可
执行命令
