一、本文介绍
本文给大家带来的改进机制是利用TPAMI最新机制FreqFusion二次创新BiFPN，《Frequency-aware Feature Fusion for Dense Image Prediction》这篇文章的主要贡献是提出了一种新的特征融合方法（FreqFusion），旨在解决密集图像预测任务中的类别内不一致性和边界位移问题。本文将其和BiFPN进行结合实现二次创新BiFPN机制，相比于原始的YOLOv26本文的内容可以达到一定的轻量化，本文的内容在作者的多类别数据集上实现了涨点。
专栏链接：YOLOv26有效涨点专栏包含：Conv、注意力机制、主干/Backbone、损失函数、优化器、后处理等改进机制
目录
一、本文介绍
二、原理介绍
三、核心代码
四、添加方法
4.1 修改一
4.2 修改二
4.3 修改三
4.4 修改四
4.5 修改五
五、正式训练
5.1 yaml文件
5.2 训练代码
5.3 训练过程截图
五、本文总结
二、原理介绍
官方论文地址：官方论文地址点击此处即可跳转
官方代码地址：官方代码地址点击此处即可跳转
《Frequency-aware Feature Fusion for Dense Image Prediction》这篇文章的主要贡献是提出了一种新的特征融合方法，旨在解决密集图像预测任务中的类别内不一致性和边界位移问题。文章中的核心概念较多，以下是简要的总结和理解：
问题定义：
密集图像预测任务（例如语义分割、目标检测和实例分割）依赖于高精度的类别信息和空间边界。但传统的特征融合方法在类别内特征一致性和边界保留上表现不佳，容易导致类别内不一致（类别内部不同部分特征差异大）和边界模糊。解决方案——FreqFusion：
文章提出了一种**频率感知特征融合（FreqFusion），它通过三个主要组件来提升融合效果：
1. 自适应低通滤波器（ALPF）生成器：该模块通过生成空间可变的低通滤波器，平滑高层特征，减少类别内不一致。
2. 偏移生成器：通过重新采样，将类别一致性较高的特征替换掉不一致的特征，进一步增强边界的清晰度。
3. 自适应高通滤波器（AHPF）生成器：用于增强在下采样过程中丢失的高频信息，提升边界细节。方法优势：
提升类别内一致性：通过ALPF组件减少了对象内部特征的波动，提升了类别内的相似度。
边界优化：通过偏移生成器和AHPF组件修正了对象边界，使得边界更加清晰。
广泛的适用性：该方法在多个任务上验证了其有效性，如语义分割、目标检测和实例分割。实验结果：
在语义分割任务中，FreqFusion相比现有方法在多个数据集（如Cityscapes和ADE20K）上有显著的提升，例如在ADE20K上比现有最优方法提升了2.8 mIoU。
在目标检测任务中，使用Faster R-CNN的FreqFusion版本在MS COCO数据集上提升了1.8 AP。
实例分割和全景分割任务中，也实现了显著的性能提升。总结：
FreqFusion通过结合自适应低通和高通滤波器，解决了标准特征融合中的类别内不一致性和边界模糊问题，在多个计算机视觉任务上提升了预测性能。
三、核心代码
核心代码使用方式看章节四！
# TPAMI 2024：Frequency-aware Feature Fusion for Dense Image Prediction import torch import torch.nn as nn import torch.nn.functional as F from mmcv.ops.carafe import normal_init, xavier_init, carafe import warnings import numpy as np __all__ = ['FreqFusion'] def normal_init(module, mean=0, std=1, bias=0): if hasattr(module, 'weight') and module.weight is not None: nn.init.normal_(module.weight, mean, std) if hasattr(module, 'bias') and module.bias is not None: nn.init.constant_(module.bias, bias) def constant_init(module, val, bias=0): if hasattr(module, 'weight') and module.weight is not None: nn.init.constant_(module.weight, val) if hasattr(module, 'bias') and module.bias is not None: nn.init.constant_(module.bias, bias) def resize(input, size=None, scale_factor=None, mode='nearest', align_corners=None, warning=True): if warning: if size is not None and align_corners: input_h, input_w = tuple(int(x) for x in input.shape[2:]) output_h, output_w = tuple(int(x) for x in size) if output_h > input_h or output_w > input_w: if ((output_h > 1 and output_w > 1 and input_h > 1 and input_w > 1) and (output_h - 1) % (input_h - 1) and (output_w - 1) % (input_w - 1)): warnings.warn( f'When align_corners={align_corners}, ' 'the output would more aligned if ' f'input size {(input_h, input_w)} is `x+1` and ' f'out size {(output_h, output_w)} is `nx+1`') return F.interpolate(input, size, scale_factor, mode, align_corners) def hamming2D(M, N): """ 生成二维Hamming窗 参数： - M：窗口的行数 - N：窗口的列数 返回： - 二维Hamming窗 """ # 生成水平和垂直方向上的Hamming窗 # hamming_x = np.blackman(M) # hamming_x = np.kaiser(M) hamming_x = np.hamming(M) hamming_y = np.hamming(N) # 通过外积生成二维Hamming窗 hamming_2d = np.outer(hamming_x, hamming_y) return hamming_2d class FreqFusion(nn.Module): def __init__(self, channels, scale_factor=1, lowpass_kernel=5, highpass_kernel=3, up_group=1, encoder_kernel=3, encoder_dilation=1, compressed_channels=64, align_corners=False, upsample_mode='nearest', feature_resample=False, # use offset generator or not feature_resample_group=4, comp_feat_upsample=True, # use ALPF & AHPF for init upsampling use_high_pass=True, use_low_pass=True, hr_residual=True, semi_conv=True, hamming_window=True, # for regularization, do not matter really feature_resample_norm=True, **kwargs): super().__init__() hr_channels, lr_channels = channels self.scale_factor = scale_factor self.lowpass_kernel = lowpass_kernel self.highpass_kernel = highpass_kernel self.up_group = up_group self.encoder_kernel = encoder_kernel self.encoder_dilation = encoder_dilation self.compressed_channels = compressed_channels self.hr_channel_compressor = nn.Conv2d(hr_channels, self.compressed_channels,1) self.lr_channel_compressor = nn.Conv2d(lr_channels, self.compressed_channels,1) self.content_encoder = nn.Conv2d( # ALPF generator self.compressed_channels, lowpass_kernel ** 2 * self.up_group * self.scale_factor * self.scale_factor, self.encoder_kernel, padding=int((self.encoder_kernel - 1) * self.encoder_dilation / 2), dilation=self.encoder_dilation, groups=1) self.align_corners = align_corners self.upsample_mode = upsample_mode self.hr_residual = hr_residual self.use_high_pass = use_high_pass self.use_low_pass = use_low_pass self.semi_conv = semi_conv self.feature_resample = feature_resample self.comp_feat_upsample = comp_feat_upsample if self.feature_resample: self.dysampler = LocalSimGuidedSampler(in_channels=compressed_channels, scale=2, style='lp', groups=feature_resample_group, use_direct_scale=True, kernel_size=encoder_kernel, norm=feature_resample_norm) if self.use_high_pass: self.content_encoder2 = nn.Conv2d( # AHPF generator self.compressed_channels, highpass_kernel ** 2 * self.up_group * self.scale_factor * self.scale_factor, self.encoder_kernel, padding=int((self.encoder_kernel - 1) * self.encoder_dilation / 2), dilation=self.encoder_dilation, groups=1) self.hamming_window = hamming_window lowpass_pad=0 highpass_pad=0 if self.hamming_window: self.register_buffer('hamming_lowpass', torch.FloatTensor(hamming2D(lowpass_kernel + 2 * lowpass_pad, lowpass_kernel + 2 * lowpass_pad))[None, None,]) self.register_buffer('hamming_highpass', torch.FloatTensor(hamming2D(highpass_kernel + 2 * highpass_pad, highpass_kernel + 2 * highpass_pad))[None, None,]) else: self.register_buffer('hamming_lowpass', torch.FloatTensor([1.0])) self.register_buffer('hamming_highpass', torch.FloatTensor([1.0])) self.init_weights() def init_weights(self): for m in self.modules(): # print(m) if isinstance(m, nn.Conv2d): xavier_init(m, distribution='uniform') normal_init(self.content_encoder, std=0.001) if self.use_high_pass: normal_init(self.content_encoder2, std=0.001) def kernel_normalizer(self, mask, kernel, scale_factor=None, hamming=1): if scale_factor is not None: mask = F.pixel_shuffle(mask, self.scale_factor) n, mask_c, h, w = mask.size() mask_channel = int(mask_c / float(kernel**2)) # mask = mask.view(n, mask_channel, -1, h, w) # mask = F.softmax(mask, dim=2, dtype=mask.dtype) # mask = mask.view(n, mask_c, h, w).contiguous() mask = mask.view(n, mask_channel, -1, h, w) mask = F.softmax(mask, dim=2, dtype=mask.dtype) mask = mask.view(n, mask_channel, kernel, kernel, h, w) mask = mask.permute(0, 1, 4, 5, 2, 3).view(n, -1, kernel, kernel) # mask = F.pad(mask, pad=[padding] * 4, mode=self.padding_mode) # kernel + 2 * padding mask = mask * hamming mask /= mask.sum(dim=(-1, -2), keepdims=True) # print(hamming) # print(mask.shape) mask = mask.view(n, mask_channel, h, w, -1) mask = mask.permute(0, 1, 4, 2, 3).view(n, -1, h, w).contiguous() return mask def forward(self, x): hr_feat, lr_feat = x compressed_hr_feat = self.hr_channel_compressor(hr_feat) compressed_lr_feat = self.lr_channel_compressor(lr_feat) if self.semi_conv: if self.comp_feat_upsample: if self.use_high_pass: mask_hr_hr_feat = self.content_encoder2(compressed_hr_feat) mask_hr_init = self.kernel_normalizer(mask_hr_hr_feat, self.highpass_kernel, hamming=self.hamming_highpass) compressed_hr_feat = compressed_hr_feat + compressed_hr_feat - carafe(compressed_hr_feat, mask_hr_init, self.highpass_kernel, self.up_group, 1) mask_lr_hr_feat = self.content_encoder(compressed_hr_feat) mask_lr_init = self.kernel_normalizer(mask_lr_hr_feat, self.lowpass_kernel, hamming=self.hamming_lowpass) mask_lr_lr_feat_lr = self.content_encoder(compressed_lr_feat) mask_lr_lr_feat = F.interpolate( carafe(mask_lr_lr_feat_lr, mask_lr_init, self.lowpass_kernel, self.up_group, 2), size=compressed_hr_feat.shape[-2:], mode='nearest') mask_lr = mask_lr_hr_feat + mask_lr_lr_feat mask_lr_init = self.kernel_normalizer(mask_lr, self.lowpass_kernel, hamming=self.hamming_lowpass) mask_hr_lr_feat = F.interpolate( carafe(self.content_encoder2(compressed_lr_feat), mask_lr_init, self.lowpass_kernel, self.up_group, 2), size=compressed_hr_feat.shape[-2:], mode='nearest') mask_hr = mask_hr_hr_feat + mask_hr_lr_feat else: raise NotImplementedError else: mask_lr = self.content_encoder(compressed_hr_feat) + F.interpolate(self.content_encoder(compressed_lr_feat), size=compressed_hr_feat.shape[-2:], mode='nearest') if self.use_high_pass: mask_hr = self.content_encoder2(compressed_hr_feat) + F.interpolate(self.content_encoder2(compressed_lr_feat), size=compressed_hr_feat.shape[-2:], mode='nearest') else: compressed_x = F.interpolate(compressed_lr_feat, size=compressed_hr_feat.shape[-2:], mode='nearest') + compressed_hr_feat mask_lr = self.content_encoder(compressed_x) if self.use_high_pass: mask_hr = self.content_encoder2(compressed_x) mask_lr = self.kernel_normalizer(mask_lr, self.lowpass_kernel, hamming=self.hamming_lowpass) if self.semi_conv: lr_feat = carafe(lr_feat, mask_lr, self.lowpass_kernel, self.up_group, 2) else: lr_feat = resize( input=lr_feat, size=hr_feat.shape[2:], mode=self.upsample_mode, align_corners=None if self.upsample_mode == 'nearest' else self.align_corners) lr_feat = carafe(lr_feat, mask_lr, self.lowpass_kernel, self.up_group, 1) if self.use_high_pass: mask_hr = self.kernel_normalizer(mask_hr, self.highpass_kernel, hamming=self.hamming_highpass) hr_feat_hf = hr_feat - carafe(hr_feat, mask_hr, self.highpass_kernel, self.up_group, 1) if self.hr_residual: # print('using hr_residual') hr_feat = hr_feat_hf + hr_feat else: hr_feat = hr_feat_hf if self.feature_resample: # print(lr_feat.shape) lr_feat = self.dysampler(hr_x=compressed_hr_feat, lr_x=compressed_lr_feat, feat2sample=lr_feat) return hr_feat + lr_feat class LocalSimGuidedSampler(nn.Module): """ offset generator in FreqFusion """ def __init__(self, in_channels, scale=2, style='lp', groups=4, use_direct_scale=True, kernel_size=1, local_window=3, sim_type='cos', norm=True, direction_feat='sim_concat'): super().__init__() assert scale==2 assert style=='lp' self.scale = scale self.style = style self.groups = groups self.local_window = local_window self.sim_type = sim_type self.direction_feat = direction_feat if style == 'pl': assert in_channels >= scale ** 2 and in_channels % scale ** 2 == 0 assert in_channels >= groups and in_channels % groups == 0 if style == 'pl': in_channels = in_channels // scale ** 2 out_channels = 2 * groups else: out_channels = 2 * groups * scale ** 2 if self.direction_feat == 'sim': self.offset = nn.Conv2d(local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) elif self.direction_feat == 'sim_concat': self.offset = nn.Conv2d(in_channels + local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) else: raise NotImplementedError normal_init(self.offset, std=0.001) if use_direct_scale: if self.direction_feat == 'sim': self.direct_scale = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, padding=kernel_size//2) elif self.direction_feat == 'sim_concat': self.direct_scale = nn.Conv2d(in_channels + local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) else: raise NotImplementedError constant_init(self.direct_scale, val=0.) out_channels = 2 * groups if self.direction_feat == 'sim': self.hr_offset = nn.Conv2d(local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) elif self.direction_feat == 'sim_concat': self.hr_offset = nn.Conv2d(in_channels + local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) else: raise NotImplementedError normal_init(self.hr_offset, std=0.001) if use_direct_scale: if self.direction_feat == 'sim': self.hr_direct_scale = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, padding=kernel_size//2) elif self.direction_feat == 'sim_concat': self.hr_direct_scale = nn.Conv2d(in_channels + local_window**2 - 1, out_channels, kernel_size=kernel_size, padding=kernel_size//2) else: raise NotImplementedError constant_init(self.hr_direct_scale, val=0.) self.norm = norm if self.norm: self.norm_hr = nn.GroupNorm(in_channels // 8, in_channels) self.norm_lr = nn.GroupNorm(in_channels // 8, in_channels) else: self.norm_hr = nn.Identity() self.norm_lr = nn.Identity() self.register_buffer('init_pos', self._init_pos()) def _init_pos(self): h = torch.arange((-self.scale + 1) / 2, (self.scale - 1) / 2 + 1) / self.scale return torch.stack(torch.meshgrid([h, h])).transpose(1, 2).repeat(1, self.groups, 1).reshape(1, -1, 1, 1) def sample(self, x, offset, scale=None): if scale is None: scale = self.scale B, _, H, W = offset.shape offset = offset.view(B, 2, -1, H, W) coords_h = torch.arange(H) + 0.5 coords_w = torch.arange(W) + 0.5 coords = torch.stack(torch.meshgrid([coords_w, coords_h]) ).transpose(1, 2).unsqueeze(1).unsqueeze(0).type(x.dtype).to(x.device) normalizer = torch.tensor([W, H], dtype=x.dtype, device=x.device).view(1, 2, 1, 1, 1) coords = 2 * (coords + offset) / normalizer - 1 coords = F.pixel_shuffle(coords.view(B, -1, H, W), scale).view( B, 2, -1, scale * H, scale * W).permute(0, 2, 3, 4, 1).contiguous().flatten(0, 1) return F.grid_sample(x.reshape(B * self.groups, -1, x.size(-2), x.size(-1)), coords, mode='bilinear', align_corners=False, padding_mode="border").view(B, -1, scale * H, scale * W) def forward(self, hr_x, lr_x, feat2sample): hr_x = self.norm_hr(hr_x) lr_x = self.norm_lr(lr_x) if self.direction_feat == 'sim': hr_sim = compute_similarity(hr_x, self.local_window, dilation=2, sim='cos') lr_sim = compute_similarity(lr_x, self.local_window, dilation=2, sim='cos') elif self.direction_feat == 'sim_concat': hr_sim = torch.cat([hr_x, compute_similarity(hr_x, self.local_window, dilation=2, sim='cos')], dim=1) lr_sim = torch.cat([lr_x, compute_similarity(lr_x, self.local_window, dilation=2, sim='cos')], dim=1) hr_x, lr_x = hr_sim, lr_sim # offset = self.get_offset(hr_x, lr_x) offset = self.get_offset_lp(hr_x, lr_x, hr_sim, lr_sim) return self.sample(feat2sample, offset) # def get_offset_lp(self, hr_x, lr_x): def get_offset_lp(self, hr_x, lr_x, hr_sim, lr_sim): if hasattr(self, 'direct_scale'): # offset = (self.offset(lr_x) + F.pixel_unshuffle(self.hr_offset(hr_x), self.scale)) * (self.direct_scale(lr_x) + F.pixel_unshuffle(self.hr_direct_scale(hr_x), self.scale)).sigmoid() + self.init_pos offset = (self.offset(lr_sim) + F.pixel_unshuffle(self.hr_offset(hr_sim), self.scale)) * (self.direct_scale(lr_x) + F.pixel_unshuffle(self.hr_direct_scale(hr_x), self.scale)).sigmoid() + self.init_pos # offset = (self.offset(lr_sim) + F.pixel_unshuffle(self.hr_offset(hr_sim), self.scale)) * (self.direct_scale(lr_sim) + F.pixel_unshuffle(self.hr_direct_scale(hr_sim), self.scale)).sigmoid() + self.init_pos else: offset = (self.offset(lr_x) + F.pixel_unshuffle(self.hr_offset(hr_x), self.scale)) * 0.25 + self.init_pos return offset def get_offset(self, hr_x, lr_x): if self.style == 'pl': raise NotImplementedError return self.get_offset_lp(hr_x, lr_x) def compute_similarity(input_tensor, k=3, dilation=1, sim='cos'): """ 计算输入张量中每一点与周围KxK范围内的点的余弦相似度。 参数： - input_tensor: 输入张量，形状为[B, C, H, W] - k: 范围大小，表示周围KxK范围内的点 返回： - 输出张量，形状为[B, KxK-1, H, W] """ B, C, H, W = input_tensor.shape # 使用零填充来处理边界情况 # padded_input = F.pad(input_tensor, (k // 2, k // 2, k // 2, k // 2), mode='constant', value=0) # 展平输入张量中每个点及其周围KxK范围内的点 unfold_tensor = F.unfold(input_tensor, k, padding=(k // 2) * dilation, dilation=dilation) # B, CxKxK, HW # print(unfold_tensor.shape) unfold_tensor = unfold_tensor.reshape(B, C, k**2, H, W) # 计算余弦相似度 if sim == 'cos': similarity = F.cosine_similarity(unfold_tensor[:, :, k * k // 2:k * k // 2 + 1], unfold_tensor[:, :, :], dim=1) elif sim == 'dot': similarity = unfold_tensor[:, :, k * k // 2:k * k // 2 + 1] * unfold_tensor[:, :, :] similarity = similarity.sum(dim=1) else: raise NotImplementedError # 移除中心点的余弦相似度，得到[KxK-1]的结果 similarity = torch.cat((similarity[:, :k * k // 2], similarity[:, k * k // 2 + 1:]), dim=1) # 将结果重塑回[B, KxK-1, H, W]的形状 similarity = similarity.view(B, k * k - 1, H, W) return similarity四、添加方法
4.1 修改一
第一还是建立文件，我们找到如下ultralytics/nn文件夹下建立一个目录名字呢就是'Addmodules'文件夹(用群内的文件的话已经有了无需新建)！然后在其内部建立一个新的py文件将核心代码复制粘贴进去即可。
4.2 修改二
第二步我们在该目录下创建一个新的py文件名字为'__init__.py'(用群内的文件的话已经有了无需新建)，然后在其内部导入我们的检测头如下图所示。
4.3 修改三
第三步我门中到如下文件'ultralytics/nn/tasks.py'进行导入和注册我们的模块(用群内的文件的话已经有了无需重新导入直接开始第四步即可)！
4.4 修改四
按照我的添加在parse_model里添加即可。
elif m in {FreqFusion}: c2 = ch[f[0]] args = [[ch[x] for x in f], *args]
4.5 修改五
第五步我门中到如下文件'ultralytics/nn/tasks.py'进行修改，按照红框的位置进行定位，用我给的代码进行替换红框中的代码.
try: m.stride = torch.tensor([s / x.shape[-2] for x in _forward(torch.zeros(1, ch, s, s))]) # forward on CPU except RuntimeError: try: self.model.to(torch.device('cuda')) m.stride = torch.tensor([s / x.shape[-2] for x in _forward( torch.zeros(1, ch, s, s).to(torch.device('cuda')))]) # forward on CUDA except RuntimeError as error: raise error到此就修改完成了，大家可以复制下面的yaml文件运行。
五、正式训练
5.1 yaml文件
训练信息：YOLO26-Neck-BiFPN-FreqFusion summary: 291 layers, 2,353,512 parameters, 2,353,512 gradients, 6.3 GFLOPs
注意：本文的机制需要关闭AMP训练否则会报错.
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license # Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs # Model docs: https://docs.ultralytics.com/models/yolo26 # Task docs: https://docs.ultralytics.com/tasks/detect # Parameters nc: 80 # number of classes end2end: True # whether to use end-to-end mode reg_max: 1 # DFL bins scales: # model compound scaling constants, i.e. 'model=yolo26n.yaml' will call yolo26.yaml with scale 'n' # [depth, width, max_channels] n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs # YOLO26n backbone backbone: # [from, repeats, module, args] - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2 - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4 - [-1, 2, C3k2, [256, False, 0.25]] - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8 - [-1, 2, C3k2, [512, False, 0.25]] - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16 - [-1, 2, C3k2, [512, True]] - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32 - [-1, 2, C3k2, [1024, True]] - [-1, 1, SPPF, [1024, 5, 3, True]] # 9 - [-1, 2, C2PSA, [1024]] # 10 # YOLO26n head head: - [4, 1, Conv, [256]] # 11-P3/8 - [6, 1, Conv, [256]] # 12-P4/16 - [10, 1, Conv, [256]] # 13-P5/32 - [[12, -1], 1, FreqFusion, []] # 14 - [-1, 2, C3k2, [256, True]] # 15-P4/16 - [[11, -1], 1, FreqFusion, []] # 16 - [-1, 2, C3k2, [256, True]] # 17-P3/8 - [1, 1, Conv, [256, 3, 2]] # 18 P2->P3 - [[-1, 11, 17], 1, BiFPN, []] # 19 - [-1, 2, C3k2, [256, True]] # 20-P3/8 - [-1, 1, Conv, [256, 3, 2]] # 21 P3->P4 - [[-1, 12, 15], 1, BiFPN, []] # 22 - [-1, 2, C3k2, [512, True]] # 23-P4/16 - [-1, 1, Conv, [256, 3, 2]] # 24 P4->P5 - [[-1, 13], 1, BiFPN, []] # 25 - [-1, 2, C3k2, [1024, True, 0.5, True]] # 26-P5/32 - [[20, 23, 26], 1, Detect, [nc]] # Detect(P3, P4, P5)5.2 训练代码
大家可以创建一个py文件将我给的代码复制粘贴进去，配置好自己的文件路径即可运行。
import warnings warnings.filterwarnings('ignore') from ultralytics import YOLO if __name__ == '__main__': model = YOLO('yolov8-MLLA.yaml') # 如何切换模型版本, 上面的ymal文件可以改为 yolov8s.yaml就是使用的v8s, # 类似某个改进的yaml文件名称为yolov8-XXX.yaml那么如果想使用其它版本就把上面的名称改为yolov8l-XXX.yaml即可（改的是上面YOLO中间的名字不是配置文件的）！ # model.load('yolov8n.pt') # 是否加载预训练权重,科研不建议大家加载否则很难提升精度 model.train(data=r"C:\Users\Administrator\PycharmProjects\yolov5-master\yolov5-master\Construction Site Safety.v30-raw-images_latestversion.yolov8\data.yaml", # 如果大家任务是其它的'ultralytics/cfg/default.yaml'找到这里修改task可以改成detect, segment, classify, pose cache=False, imgsz=640, epochs=150, single_cls=False, # 是否是单类别检测 batch=16, close_mosaic=0, workers=0, device='0', optimizer='SGD', # using SGD # resume='runs/train/exp21/weights/last.pt', # 如过想续训就设置last.pt的地址 amp=False, # 如果出现训练损失为Nan可以关闭amp project='runs/train', name='exp', )5.3 训练过程截图
五、本文总结
到此本文的正式分享内容就结束了，在这里给大家推荐我的YOLOv26改进有效涨点专栏，本专栏目前为新开的平均质量分98分，后期我会根据各种最新的前沿顶会进行论文复现，也会对一些老的改进机制进行补充，如果大家觉得本文帮助到你了，订阅本专栏，关注后续更多的更新~
专栏链接：