一、导言
提出了一种为移动设备设计的高效神经网络主干结构,称为MobileOne。传统的优化方法通常侧重于减少计算量(FLOPs)和参数数量来提升模型效率,但这些指标并不总能准确反映实际部署到移动设备上的延迟性能。因此,研究团队通过在移动设备上部署多个对移动友好的网络并进行深入分析,识别并解决了现代高效神经网络架构中的架构与优化瓶颈。
MobileOne的设计旨在提供超低延迟的同时保持高精度,其变体能够在iPhone 12上实现低于1毫秒的推理时间,同时在ImageNet数据集上达到75.9%的顶尖-1准确率。相比同类高效架构,MobileOne不仅在精度上达到了最先进的水平,而且运行速度更快,比如比MobileFormer快38倍,比EfficientNet在相近延迟下高出2.3%的顶尖-1准确率。
此外,MobileOne模型还展现了良好的泛化能力,不仅在图像分类任务上表现优异,还在目标检测和语义分割等任务上展现出显著的延迟降低和准确率提升。研究中还探讨了训练策略对模型性能的影响,发现MobileOne模型对随机种子的选择不敏感,且较少依赖正则化技术即可达到竞争力的精度。作者还研究了过参数化对于不同规模模型的效果,并发现对于小型变体而言,额外的过参数化分支能够带来更显著的性能提升。
-
低延迟优化:MobileOne架构专门针对移动设备进行了设计和优化,其变体能够在iPhone 12上实现低于1毫秒的推理时间,同时在ImageNet数据集上达到75.9%的Top-1精度,表明其在实际部署时具有极低的延迟。
-
性能超越:与当前高效的架构相比,MobileOne不仅在准确性方面达到了最先进的水平,而且在移动设备上的运行速度快得多。例如,它比MobileFormer快38倍,同时在ImageNet上的表现相当;比EfficientNet在相似延迟下高出2.3%的Top-1精度。
-
多任务适应性:MobileOne模型不仅在图像分类任务上表现出色,还泛化到了对象检测和语义分割等其他任务,在移动设备上实现了显著的延迟和精度改进。
-
训练策略和正则化:研究通过详尽实验分析了不同训练策略对模型性能的影响,发现MobileOne模型即使使用较少的正则化技术也能达到与竞争模型如EfficientNet、MobileNetV3-L和MixNet-S相匹敌的精度。此外,采用逐步学习和权重衰减策略能进一步提升某些模型的表现。
-
稳定性与随机种子敏感性:模型训练稳健,不同随机种子下的表现一致,证明了方法的可靠性和重现性。
-
微观架构设计:MobileOne的微观变体设计不单纯追求FLOPs的优化,而是通过选择参数量更小且使用适度过参数化的架构,以达到竞争力的精度,这为资源有限的环境提供了有效解决方案。
-
过参数化效果:研究表明,对于较小的模型变体,额外的过参数化分支带来了更多益处,有助于提高模型性能。
-
扩展到物体检测:在物体检测任务中,通过使用SSDLite模型进行训练,并采用余弦学习率调度结合预热策略,展示了MobileOne在实际应用中的多功能性和潜力。
综上所述,MobileOne是一个针对移动平台优化的深度学习模型,它在确保高性能的同时极大提高了推理速度,为移动设备上的实时AI应用提供了新的可能性。相关代码和模型已开源供研究者使用。
二、准备工作
首先在YOLOv5/v7项目文件下的models文件夹下创建新的文件mobileone.py
导入如下代码
from models.common import *
import copy as copy2
from typing import Optional, List, Tuple
class SEBlock(nn.Module):
""" Squeeze and Excite module.
https://arxiv.org/pdf/1709.01507.pdf
"""
def __init__(self, in_channels: int, rd_ratio: float = 0.0625) -> None:
""" Construct a Squeeze and Excite Module.
:param in_channels: Number of input channels.
:param rd_ratio: Input channel reduction ratio.
"""
super(SEBlock, self).__init__()
self.reduce = nn.Conv2d(in_channels=in_channels,out_channels=int(in_channels * rd_ratio), kernel_size=1, stride=1, bias=True)
self.expand = nn.Conv2d(in_channels=int(in_channels * rd_ratio),out_channels=in_channels, kernel_size=1, stride=1, bias=True)
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
""" Apply forward pass. """
b, c, h, w = inputs.size()
x = F.avg_pool2d(inputs, kernel_size=[h, w])
x = self.reduce(x)
x = F.relu(x)
x = self.expand(x)
x = torch.sigmoid(x)
x = x.view(-1, c, 1, 1)
return inputs * x
class MobileOneBlock(nn.Module):
""" MobileOne building block. https://arxiv.org/pdf/2206.04040.pdf
"""
def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1,
padding: int = 0, dilation: int = 1, groups: int = 1, use_se: bool = False, num_conv_branches: int = 1, inference_mode: bool = False) -> None:
""" Construct a MobileOneBlock module.
:param in_channels: Number of channels in the input.
:param out_channels: Number of channels produced by the block.
:param kernel_size: Size of the convolution kernel.
:param stride: Stride size.
:param padding: Zero-padding size.
:param dilation: Kernel dilation factor.
:param groups: Group number.
:param inference_mode: If True, instantiates model in inference mode.
:param use_se: Whether to use SE-ReLU activations.
:param num_conv_branches: Number of linear conv branches.
"""
super(MobileOneBlock, self).__init__()
self.inference_mode = inference_mode
self.groups = groups
self.stride = stride
self.kernel_size = kernel_size
self.in_channels = in_channels
self.out_channels = out_channels
self.num_conv_branches = num_conv_branches # 4
# Check if SE-ReLU is requested
if use_se:
self.se = SEBlock(out_channels)
else:
self.se = nn.Identity()
self.activation = nn.ReLU()
if inference_mode:
self.reparam_conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,stride=stride, padding=padding, dilation=dilation, groups=groups, bias=True)
else:
# Re-parameterizable skip connection
self.rbr_skip = nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None # BN skip
# Re-parameterizable conv branches
rbr_conv = list()
for _ in range(self.num_conv_branches):
rbr_conv.append(self._conv_bn(kernel_size=kernel_size, padding=padding))
self.rbr_conv = nn.ModuleList(rbr_conv)
# Re-parameterizable scale branch
self.rbr_scale = None
if kernel_size > 1:
self.rbr_scale = self._conv_bn(kernel_size=1, padding=0)
def forward(self, x: torch.Tensor) -> torch.Tensor:
""" Apply forward pass. """
# Inference mode forward pass.
if self.inference_mode:
return self.activation(self.se(self.reparam_conv(x)))
# Multi-branched train-time forward pass.
# Skip branch output
identity_out = 0
if self.rbr_skip is not None:
identity_out = self.rbr_skip(x)
# Scale branch output
scale_out = 0
if self.rbr_scale is not None:
scale_out = self.rbr_scale(x)
# Other branches
out = scale_out + identity_out
for ix in range(self.num_conv_branches):
out += self.rbr_conv[ix](x)
return self.activation(self.se(out))
def reparameterize(self):
""" Following works like `RepVGG: Making VGG-style ConvNets Great Again` -
https://arxiv.org/pdf/2101.03697.pdf. We re-parameterize multi-branched
architecture used at training time to obtain a plain CNN-like structure
for inference.
"""
if self.inference_mode:
return
kernel, bias = self._get_kernel_bias()
self.reparam_conv = nn.Conv2d(in_channels=self.rbr_conv[0].conv.in_channels,
out_channels=self.rbr_conv[0].conv.out_channels,
kernel_size=self.rbr_conv[0].conv.kernel_size,
stride=self.rbr_conv[0].conv.stride,
padding=self.rbr_conv[0].conv.padding,
dilation=self.rbr_conv[0].conv.dilation,
groups=self.rbr_conv[0].conv.groups,
bias=True)
self.reparam_conv.weight.data = kernel
self.reparam_conv.bias.data = bias
# Delete un-used branches
for para in self.parameters():
para.detach_()
self.__delattr__('rbr_conv')
self.__delattr__('rbr_scale')
if hasattr(self, 'rbr_skip'):
self.__delattr__('rbr_skip')
self.inference_mode = True
def _get_kernel_bias(self) -> Tuple[torch.Tensor, torch.Tensor]:
""" Method to obtain re-parameterized kernel and bias.
Reference: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L83
:return: Tuple of (kernel, bias) after fusing branches.
"""
# get weights and bias of scale branch
kernel_scale = 0
bias_scale = 0
if self.rbr_scale is not None:
kernel_scale, bias_scale = self._fuse_bn_tensor(self.rbr_scale)
# Pad scale branch kernel to match conv branch kernel size.
pad = self.kernel_size // 2
kernel_scale = torch.nn.functional.pad(kernel_scale, [pad, pad, pad, pad])
# get weights and bias of skip branch
kernel_identity = 0
bias_identity = 0
if self.rbr_skip is not None:
kernel_identity, bias_identity = self._fuse_bn_tensor(self.rbr_skip)
# get weights and bias of conv branches
kernel_conv = 0
bias_conv = 0
for ix in range(self.num_conv_branches):
_kernel, _bias = self._fuse_bn_tensor(self.rbr_conv[ix])
kernel_conv += _kernel
bias_conv += _bias
kernel_final = kernel_conv + kernel_scale + kernel_identity
bias_final = bias_conv + bias_scale + bias_identity
return kernel_final, bias_final
def _fuse_bn_tensor(self, branch) -> Tuple[torch.Tensor, torch.Tensor]:
""" Method to fuse batchnorm layer with preceeding conv layer.
Reference: https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L95
:param branch:
:return: Tuple of (kernel, bias) after fusing batchnorm.
"""
if isinstance(branch, nn.Sequential):
kernel = branch.conv.weight
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 // self.groups
kernel_value = torch.zeros((self.in_channels, input_dim, self.kernel_size, self.kernel_size),
dtype=branch.weight.dtype, device=branch.weight.device)
for i in range(self.in_channels):
kernel_value[i, i % input_dim,self.kernel_size // 2, self.kernel_size // 2] = 1
self.id_tensor = kernel_value
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 _conv_bn(self, kernel_size: int, padding: int) -> nn.Sequential:
""" Helper method to construct conv-batchnorm layers.
:param kernel_size: Size of the convolution kernel.
:param padding: Zero-padding size.
:return: Conv-BN module.
"""
mod_list = nn.Sequential()
mod_list.add_module('conv', nn.Conv2d(in_channels=self.in_channels,out_channels=self.out_channels,
kernel_size=kernel_size, stride=self.stride, padding=padding, groups=self.groups, bias=False))
mod_list.add_module('bn', nn.BatchNorm2d(num_features=self.out_channels))
return mod_list
class MobileOne(nn.Module):
""" MobileOne Model https://arxiv.org/pdf/2206.04040.pdf """
def __init__(self,
in_channels, out_channels,
num_blocks_per_stage = 2, num_conv_branches: int = 1,
use_se: bool = False, num_se: int = 0,
inference_mode: bool = False, ) -> None:
""" Construct MobileOne model.
:param num_blocks_per_stage: List of number of blocks per stage.
:param num_classes: Number of classes in the dataset.
:param width_multipliers: List of width multiplier for blocks in a stage.
:param inference_mode: If True, instantiates model in inference mode.
:param use_se: Whether to use SE-ReLU activations.
:param num_conv_branches: Number of linear conv branches.
"""
super().__init__()
self.inference_mode = inference_mode
self.use_se = use_se
self.num_conv_branches = num_conv_branches
self.stage = self._make_stage(in_channels, out_channels, num_blocks_per_stage, num_se_blocks= num_se if use_se else 0)
def _make_stage(self, in_channels, out_channels, num_blocks: int, num_se_blocks: int) -> nn.Sequential:
""" Build a stage of MobileOne model.
:param planes: Number of output channels.
:param num_blocks: Number of blocks in this stage.
:param num_se_blocks: Number of SE blocks in this stage.
:return: A stage of MobileOne model.
"""
# Get strides for all layers
strides = [2] + [1]*(num_blocks-1)
blocks = []
for ix, stride in enumerate(strides): # 用于训练几个blocks
use_se = False
if num_se_blocks > num_blocks:
raise ValueError("Number of SE blocks cannot " "exceed number of layers.")
if ix >= (num_blocks - num_se_blocks):
use_se = True
# Depthwise conv
blocks.append(MobileOneBlock(in_channels=in_channels, out_channels=in_channels,
kernel_size=3, stride=stride, padding=1, groups=in_channels,
inference_mode=self.inference_mode, use_se=use_se, num_conv_branches=self.num_conv_branches))
# Pointwise conv
blocks.append(MobileOneBlock(in_channels=in_channels, out_channels=out_channels,
kernel_size=1, stride=1, padding=0, groups=1,
inference_mode=self.inference_mode, use_se=use_se, num_conv_branches=self.num_conv_branches))
in_channels = out_channels
return nn.Sequential(*blocks)
def forward(self, x: torch.Tensor) -> torch.Tensor:
""" Apply forward pass. """
x = self.stage(x)
return x
def reparameterize_model(model: torch.nn.Module) -> nn.Module:
""" Method returns a model where a multi-branched structure
used in training is re-parameterized into a single branch
for inference.
:param model: MobileOne model in train mode.
:return: MobileOne model in inference mode.
"""
# Avoid editing original graph
model = copy2.deepcopy(model)
for module in model.modules():
if hasattr(module, 'reparameterize'):
module.reparameterize()
return model
其次在在YOLOv5/v7项目文件下的models/yolo.py中在文件首部添加代码
from models.mobileone import *
并搜索def parse_model(d, ch)
定位到如下行添加以下代码
elif m in [MobileOneBlock, MobileOne]:
c1, c2 = ch[f], args[0]
args = [c1, c2, *args[1:]]
三、YOLOv7-tiny改进工作
完成二后,在YOLOv5项目文件下的models文件夹下创建新的文件yolov7-tiny-mobileone.yaml,导入如下代码。
# parameters
nc: 80 # number of classes
depth_multiple: 1.0 # model depth multiple
width_multiple: 1.0 # layer channel multiple
# anchors
anchors:
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32
# yolov7-tiny backbone
backbone:
# [from, number, module, args] c2, k=1, s=1, p=None, g=1, act=True
[[-1, 1, MobileOneBlock, [48, 3, 2, 1]], # 0
[-1, 1, MobileOne, [48, 2, 4, False, 0]], # 1
[-1, 1, MobileOne, [128, 8, 4, False, 0]], # 2
[-1, 1, MobileOne, [256, 10, 4, False, 0]], # 3
[-1, 1, MobileOne, [512, 1, 4, False, 0]], # 4
]
# yolov7-tiny head
head:
[[-1, 1, Conv, [256, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-2, 1, Conv, [256, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, SP, [5]],
[-2, 1, SP, [9]],
[-3, 1, SP, [13]],
[[-1, -2, -3, -4], 1, Concat, [1]],
[-1, 1, Conv, [256, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -7], 1, Concat, [1]],
[-1, 1, Conv, [256, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # 13
[-1, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[3, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # route backbone P4
[[-1, -2], 1, Concat, [1]],
[-1, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-2, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [64, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [64, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -2, -3, -4], 1, Concat, [1]],
[-1, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # 23
[-1, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[2, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -2], 1, Concat, [1]], # 27
[-1, 1, Conv, [32, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-2, 1, Conv, [32, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [32, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [32, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -2, -3, -4], 1, Concat, [1]],
[-1, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # 33
[-1, 1, Conv, [128, 3, 2, None, 1, nn.LeakyReLU(0.1)]],
[[-1, 23], 1, Concat, [1]],
[-1, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-2, 1, Conv, [64, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [64, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [64, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -2, -3, -4], 1, Concat, [1]],
[-1, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # 41
[-1, 1, Conv, [256, 3, 2, None, 1, nn.LeakyReLU(0.1)]],
[[-1, 13], 1, Concat, [1]],
[-1, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-2, 1, Conv, [128, 1, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [128, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[-1, 1, Conv, [128, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[[-1, -2, -3, -4], 1, Concat, [1]],
[-1, 1, Conv, [256, 1, 1, None, 1, nn.LeakyReLU(0.1)]], # 49
[33, 1, Conv, [128, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[41, 1, Conv, [256, 3, 1, None, 1, nn.LeakyReLU(0.1)]],
[49, 1, Conv, [512, 3, 1, None, 1, nn.LeakyReLU(0.1)]], # 52
[[50,51,52], 1, IDetect, [nc, anchors]], # Detect(P3, P4, P5)
]
from n params module arguments
0 -1 1 1632 models.mobileone.MobileOneBlock [3, 48, 3, 2, 1]
1 -1 1 24000 models.mobileone.MobileOne [48, 48, 2, 4, False, 0]
2 -1 1 539472 models.mobileone.MobileOne [48, 128, 8, 4, False, 0]
3 -1 1 2634368 models.mobileone.MobileOne [128, 256, 10, 4, False, 0]
4 -1 1 540416 models.mobileone.MobileOne [256, 512, 1, 4, False, 0]
5 -1 1 131584 models.common.Conv [512, 256, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
6 -2 1 131584 models.common.Conv [512, 256, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
7 -1 1 0 models.common.SP [5]
8 -2 1 0 models.common.SP [9]
9 -3 1 0 models.common.SP [13]
10 [-1, -2, -3, -4] 1 0 models.common.Concat [1]
11 -1 1 262656 models.common.Conv [1024, 256, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
12 [-1, -7] 1 0 models.common.Concat [1]
13 -1 1 131584 models.common.Conv [512, 256, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
14 -1 1 33024 models.common.Conv [256, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
15 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
16 3 1 33024 models.common.Conv [256, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
17 [-1, -2] 1 0 models.common.Concat [1]
18 -1 1 16512 models.common.Conv [256, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
19 -2 1 16512 models.common.Conv [256, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
20 -1 1 36992 models.common.Conv [64, 64, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
21 -1 1 36992 models.common.Conv [64, 64, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
22 [-1, -2, -3, -4] 1 0 models.common.Concat [1]
23 -1 1 33024 models.common.Conv [256, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
24 -1 1 8320 models.common.Conv [128, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
25 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
26 2 1 8320 models.common.Conv [128, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
27 [-1, -2] 1 0 models.common.Concat [1]
28 -1 1 4160 models.common.Conv [128, 32, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
29 -2 1 4160 models.common.Conv [128, 32, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
30 -1 1 9280 models.common.Conv [32, 32, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
31 -1 1 9280 models.common.Conv [32, 32, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
32 [-1, -2, -3, -4] 1 0 models.common.Concat [1]
33 -1 1 8320 models.common.Conv [128, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
34 -1 1 73984 models.common.Conv [64, 128, 3, 2, None, 1, LeakyReLU(negative_slope=0.1)]
35 [-1, 23] 1 0 models.common.Concat [1]
36 -1 1 16512 models.common.Conv [256, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
37 -2 1 16512 models.common.Conv [256, 64, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
38 -1 1 36992 models.common.Conv [64, 64, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
39 -1 1 36992 models.common.Conv [64, 64, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
40 [-1, -2, -3, -4] 1 0 models.common.Concat [1]
41 -1 1 33024 models.common.Conv [256, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
42 -1 1 295424 models.common.Conv [128, 256, 3, 2, None, 1, LeakyReLU(negative_slope=0.1)]
43 [-1, 13] 1 0 models.common.Concat [1]
44 -1 1 65792 models.common.Conv [512, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
45 -2 1 65792 models.common.Conv [512, 128, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
46 -1 1 147712 models.common.Conv [128, 128, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
47 -1 1 147712 models.common.Conv [128, 128, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
48 [-1, -2, -3, -4] 1 0 models.common.Concat [1]
49 -1 1 131584 models.common.Conv [512, 256, 1, 1, None, 1, LeakyReLU(negative_slope=0.1)]
50 33 1 73984 models.common.Conv [64, 128, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
51 41 1 295424 models.common.Conv [128, 256, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
52 49 1 1180672 models.common.Conv [256, 512, 3, 1, None, 1, LeakyReLU(negative_slope=0.1)]
53 [50, 51, 52] 1 17132 models.yolo.IDetect [1, [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]], [128, 256, 512]]
Model Summary: 953 layers, 7290460 parameters, 7290460 gradients, 23.9 GFLOPS
运行后若打印出如上文本代表改进成功。
四、YOLOv5n改进工作
完成二后,在YOLOv5项目文件下的models文件夹下创建新的文件yolov5n-mobileone.yaml,导入如下代码。
# Parameters
nc: 80 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.25 # layer channel multiple
anchors:
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32
# YOLOv5 v6.0 backbone
backbone:
# [from, number, module, args]
[[-1, 1, MobileOneBlock, [48, 3, 2, 1]], # 0
[-1, 1, MobileOne, [48, 2, 4, False, 0]], # 1
[-1, 1, MobileOne, [128, 8, 4, False, 0]], # 2
[-1, 1, MobileOne, [256, 10, 4, False, 0]], # 3
[-1, 1, MobileOne, [512, 1, 4, False, 0]], # 4
]
# YOLOv5 v6.0 head
head:
[[-1, 1, Conv, [512, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 3], 1, Concat, [1]], # cat backbone P4
[-1, 3, C3, [512, False]], # 8
[-1, 1, Conv, [256, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 2], 1, Concat, [1]], # cat backbone P3
[-1, 3, C3, [256, False]], # 12 (P3/8-small)
[-1, 1, Conv, [256, 3, 2]],
[[-1, 9], 1, Concat, [1]], # cat head P4
[-1, 3, C3, [512, False]], # 15 (P4/16-medium)
[-1, 1, Conv, [512, 3, 2]],
[[-1, 5], 1, Concat, [1]], # cat head P5
[-1, 3, C3, [1024, False]], # 18 (P5/32-large)
[[12, 15, 18], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
]
from n params module arguments
0 -1 1 1632 models.mobileone.MobileOneBlock [3, 48, 3, 2, 1]
1 -1 1 24000 models.mobileone.MobileOne [48, 48, 2, 4, False, 0]
2 -1 1 539472 models.mobileone.MobileOne [48, 128, 8, 4, False, 0]
3 -1 1 2634368 models.mobileone.MobileOne [128, 256, 10, 4, False, 0]
4 -1 1 540416 models.mobileone.MobileOne [256, 512, 1, 4, False, 0]
5 -1 1 65792 models.common.Conv [512, 128, 1, 1]
6 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
7 [-1, 3] 1 0 models.common.Concat [1]
8 -1 1 107264 models.common.C3 [384, 128, 1, False]
9 -1 1 8320 models.common.Conv [128, 64, 1, 1]
10 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
11 [-1, 2] 1 0 models.common.Concat [1]
12 -1 1 27008 models.common.C3 [192, 64, 1, False]
13 -1 1 36992 models.common.Conv [64, 64, 3, 2]
14 [-1, 9] 1 0 models.common.Concat [1]
15 -1 1 74496 models.common.C3 [128, 128, 1, False]
16 -1 1 147712 models.common.Conv [128, 128, 3, 2]
17 [-1, 5] 1 0 models.common.Concat [1]
18 -1 1 296448 models.common.C3 [256, 256, 1, False]
19 [12, 15, 18] 1 8118 models.yolo.Detect [1, [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]], [64, 128, 256]]
Model Summary: 909 layers, 4512038 parameters, 4512038 gradients, 19.4 GFLOPs
运行后若打印出如上文本代表改进成功。
五、YOLOv5s改进工作
完成二后,在YOLOv5项目文件下的models文件夹下创建新的文件yolov5s-vanillanet.yaml,导入如下代码。
# Parameters
nc: 80 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors:
- [10,13, 16,30, 33,23] # P3/8
- [30,61, 62,45, 59,119] # P4/16
- [116,90, 156,198, 373,326] # P5/32
# YOLOv5 v6.0 backbone
backbone:
# [from, number, module, args]
[[-1, 1, MobileOneBlock, [48, 3, 2, 1]], # 0
[-1, 1, MobileOne, [48, 2, 4, False, 0]], # 1
[-1, 1, MobileOne, [128, 8, 4, False, 0]], # 2
[-1, 1, MobileOne, [256, 10, 4, False, 0]], # 3
[-1, 1, MobileOne, [512, 1, 4, False, 0]], # 4
]
# YOLOv5 v6.0 head
head:
[[-1, 1, Conv, [512, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 3], 1, Concat, [1]], # cat backbone P4
[-1, 3, C3, [512, False]], # 8
[-1, 1, Conv, [256, 1, 1]],
[-1, 1, nn.Upsample, [None, 2, 'nearest']],
[[-1, 2], 1, Concat, [1]], # cat backbone P3
[-1, 3, C3, [256, False]], # 12 (P3/8-small)
[-1, 1, Conv, [256, 3, 2]],
[[-1, 9], 1, Concat, [1]], # cat head P4
[-1, 3, C3, [512, False]], # 15 (P4/16-medium)
[-1, 1, Conv, [512, 3, 2]],
[[-1, 5], 1, Concat, [1]], # cat head P5
[-1, 3, C3, [1024, False]], # 18 (P5/32-large)
[[12, 15, 18], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
]
from n params module arguments
0 -1 1 1632 models.mobileone.MobileOneBlock [3, 48, 3, 2, 1]
1 -1 1 24000 models.mobileone.MobileOne [48, 48, 2, 4, False, 0]
2 -1 1 539472 models.mobileone.MobileOne [48, 128, 8, 4, False, 0]
3 -1 1 2634368 models.mobileone.MobileOne [128, 256, 10, 4, False, 0]
4 -1 1 540416 models.mobileone.MobileOne [256, 512, 1, 4, False, 0]
5 -1 1 131584 models.common.Conv [512, 256, 1, 1]
6 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
7 [-1, 3] 1 0 models.common.Concat [1]
8 -1 1 361984 models.common.C3 [512, 256, 1, False]
9 -1 1 33024 models.common.Conv [256, 128, 1, 1]
10 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
11 [-1, 2] 1 0 models.common.Concat [1]
12 -1 1 90880 models.common.C3 [256, 128, 1, False]
13 -1 1 147712 models.common.Conv [128, 128, 3, 2]
14 [-1, 9] 1 0 models.common.Concat [1]
15 -1 1 296448 models.common.C3 [256, 256, 1, False]
16 -1 1 590336 models.common.Conv [256, 256, 3, 2]
17 [-1, 5] 1 0 models.common.Concat [1]
18 -1 1 1182720 models.common.C3 [512, 512, 1, False]
19 [12, 15, 18] 1 16182 models.yolo.Detect [1, [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]], [128, 256, 512]]
Model Summary: 909 layers, 6590758 parameters, 6590758 gradients, 23.4 GFLOPs
运行后打印如上代码说明改进成功。
下一篇文章:【YOLOv5/v7改进系列】替换骨干网络为ShuffleNetv2
将会进行手把手的改进教学。
更多文章产出中,主打简洁和准确,欢迎关注我,共同探讨!