Part8.第15章:Transformer

第15章:Transformer

动机与影响

Transformer与全连接前馈网络(FFN)、卷积神经网络(CNN)、循环神经网络(RNN)并称为深度学习的四大核心架构。
序列建模的挑战
在这里插入图片描述
Transformer的突破性贡献
在这里插入图片描述

注意力机制

在这里插入图片描述
在Transformer里,计算Q和K是直接计算Q和K的点乘。当然,计算点乘的前提是Q和K的维度是一致的。一个Q和一组K分别计算点乘,得到一组logits值,再通过softmax,就得到注意力权重

1.自注意力机制

在这里插入图片描述

2.多头自注意力

在这里插入图片描述

3.矩阵运算加速

在这里插入图片描述
以一个头的注意力机制为例,也就是q、k、v向量的维度和embedding向量的维度都一样,为hidden_size
我们将序列内所有token产生的q向量合并到一起,形成一个Q矩阵,形状为[seq_len,hidden_size]。将所有的k向量合并到一起,形成一个K矩阵,形状为[seq_len,hidden_size]。
同理生成V矩阵,形状也为[seq_len,hidden_size]。此时,将Q乘以K的转置相乘:QKTQK^{T}QKT,得到结果形状为[seq_len,seq_len]。然后按行应用softmax就得到了注意力权重矩阵。
在这里插入图片描述
在这里插入图片描述
在这里插入图片描述
比如上边就是4个token的Q,K,V矩阵,token的hidden_size为6,分了3个头,每个头的q,k,v向量维度为2。整个序列的4个token进行多头注意计算时就是Head1、Head2、Head3的Q,K,V矩阵分别进行计算,分别得到结果矩阵,再将结果矩阵进行拼接,就得到4个token更新后的Embedding。
在这里插入图片描述
一个完整的注意力机制计算后还会通过一个全连接层来整理token的embedding。这个全连接层不会改变token embedding的维度。

4.Pytorch实现代码

class MultiHeadAttentionBlock(nn.Module):

    def __init__(self, d_model: int, h: int, dropout: float) -> None:
        super().__init__()
        self.d_model = d_model  # embedding特征大小
        self.h = h  # 头的个数
        assert d_model % h == 0, "d_model 不能被 h整除" # 确保d_model可以被h整除,多头分割的必要条件

        self.d_k = d_model // h  # 每个头特征大小
        self.w_q = nn.Linear(d_model, d_model, bias=False)  # Wq,Q矩阵的线性变换层(无偏置)
        self.w_k = nn.Linear(d_model, d_model, bias=False)  # Wk,类似Wq
        self.w_v = nn.Linear(d_model, d_model, bias=False)  # Wv,类似Wq
        self.w_o = nn.Linear(d_model, d_model, bias=False)  # Wo,输出投影层(合并多头结果)
        self.dropout = nn.Dropout(dropout)

    @staticmethod
    def attention(query, key, value, mask, dropout: nn.Dropout):
        # 获取d_k的值。
        d_k = query.shape[-1]
        # Q乘以K的转置,除以根号下d_k。
        # (batch, h, seq_len, d_k) --> (batch, h, seq_len, seq_len)
        attention_scores = (query @ key.transpose(-2, -1)) / math.sqrt(d_k)
        if mask is not None:
            # 给mask为0的位置填入一个很大的负值,这样在进行softmax,注意力就为0。
            attention_scores.masked_fill_(mask == 0, -1e9)
        # 进行softmax,归一化。得到注意力权重
        # (batch, h, seq_len, seq_len)
        attention_scores = attention_scores.softmax(dim=-1)
        if dropout is not None:
            attention_scores = dropout(attention_scores)
        # 注意力权重乘以V,得到更新后的embedding。
        # (batch, h, seq_len, seq_len) --> (batch, h, seq_len, d_k)
        return (attention_scores @ value), attention_scores

    def forward(self, q, k, v, mask):
        # 通过3个全连接层,获取Q、K、V矩阵
        query = self.w_q(q)  # (batch, seq_len, d_model) --> (batch, seq_len, d_model)
        key = self.w_k(k)  # (batch, seq_len, d_model) --> (batch, seq_len, d_model)
        value = self.w_v(v)  # (batch, seq_len, d_model) --> (batch, seq_len, d_model)

        # 对多头进行拆分
        # (batch, seq_len, d_model) --> (batch, seq_len, h, d_k) --> (batch, h, seq_len, d_k)
        query = query.view(query.shape[0], query.shape[1], self.h, self.d_k).transpose(1, 2)
        key = key.view(key.shape[0], key.shape[1], self.h, self.d_k).transpose(1, 2)
        value = value.view(value.shape[0], value.shape[1], self.h, self.d_k).transpose(1, 2)

        # 计算注意力
        x, self.attention_scores = MultiHeadAttentionBlock.attention(query, key, value, mask, self.dropout)

        # 多个头合并
        # (batch, h, seq_len, d_k) --> (batch, seq_len, h, d_k) --> (batch, seq_len, d_model)
        x = x.transpose(1, 2).contiguous().view(x.shape[0], -1, self.h * self.d_k)

        # 乘以输出层
        return self.w_o(x)

forward函数传入的q,k,v是尚未经过全连接层的原始向量,在self-attention里,这里的q,k,v都是token的embedding。并且这里的embedding的形状为(batch,seq_len,d_model)。可见Transformer里的tensor将batch size放在第一个维度,因为Transformer里可以同时对所有token进行处理,并不需要按照序列顺序依次处理。
Attention计算时,可以传入一个mask矩阵,mask矩阵用0标记了哪些位置不参与注意力计算。比如对于<pad><pad><pad> token就不必参与注意力计算。对于mask标记了0的位置,在注意力logits值计算完成后,给赋值一个很大的负值,这样在进行softmax后,对于这个位置的注意力就为0。相当于不参加注意力计算。
关于代码的几个问题:
【1】Q矩阵的线性变换层(无偏置)为啥无偏置?
1.1.参数效率与过拟合风险控制
每个注意力头独立拥有d_model维的线性变换层(Q/K/V各一层)。若启用偏置,每个头需额外存储3个长度为d_model的偏置向量,总参数增加3×h×d_model(h为头数);
1.2.数学性质保持
注意力计算的核心是Q⋅KTQ·K^TQKT 的缩放点积。偏置项会引入额外的常数偏移,可能破坏点积的对称性或数值稳定性。例如,若Q的变换层包含偏置bqb_qbq,则实际计算变为(Q⋅Wq+bq)⋅KT(Q·W_q + b_q)·K^T(QWq+bq)KT,等价于Q⋅(Wq⋅KT)+bq⋅KTQ·(W_q·K^T) + b_q·K^TQ(WqKT)+bqKT。后一项bq⋅KTb_q·K^TbqKT是全局偏移,可能干扰注意力权重的分布,尤其在序列较长时累加效应明显;
1.3.表达冗余性消除
在深度学习中,层归一化(LayerNorm)和残差连接已能处理输入分布的偏移。若线性层再引入偏置,可能造成双重偏移补偿,导致表达冗余;
1.4.工程实践验证
原始论文《Attention Is All You Need》通过消融实验发现,无偏置设计在机器翻译、文本建模等任务上性能与含偏置版本相当,甚至略优。
【2】attention_scores = (query @ key.transpose(-2, -1)) / math.sqrt(d_k),语法解释?
2.1.矩阵乘法运算符 @
在Python的NumPy/PyTorch库中,@是矩阵乘法(Matmul) 的专用运算符,等价于 torch.matmul() 或 torch.einsum() 的高效实现。
2.2.维度交换方法 transpose(-2, -1)
参数含义:-2 和 -1 是 张量的最后两个维度索引(PyTorch采用负数索引从后往前计数)。
操作效果:交换张量最后两个维度的位置。
示例:若原始 key 形状为 (batch_size, h, seq_len, d_k),则 key.transpose(-2, -1) 变为 (batch_size, h, d_k, seq_len)。
2.3.缩放因子 math.sqrt(d_k)。数学动机:防止点积结果过大导致softmax梯度消失。
【3】attention_scores.masked_fill_(mask == 0, -1e9),这句代码含义?
在这里插入图片描述
【4】query = self.w_q(q) # (batch, seq_len, d_model) --> (batch, seq_len, d_model),代码含义?
多头注意力机制的核心初始化步骤,通过三个独立的线性全连接层将输入序列投影到查询(Q)、键(K)、值(V)三个不同的表示子空间。
线性变换

query = self.w_q(q)  # 等价于 q @ W_q^T + b_q(但b_q=0)

通过三个独立的全连接层,将相同的输入序列映射到三个不同的表示空间
Q(查询):捕捉“需要关注什么信息”的意图向量
K(键):存储“可被关注的信息”的索引向量
V(值):承载“实际信息内容”的载体向量
这种分离使模型能同时学习“如何提问”(Q)、“如何响应”(K)、“如何提取信息”(V)三种能力。
【5】 query = query.view(query.shape[0], query.shape[1], self.h, self.d_k).transpose(1, 2),代码含义?
多头拆分是多头注意力机制的核心步骤,通过维度重组将输入特征分割为多个并行的注意力头,使每个头能独立学习不同子空间的特征表示?
a.特征空间分割:将高维特征分解为多个低维子空间
b.计算效率优化:通过批量矩阵运算提升GPU利用率
c.表达容量扩展:通过并行子空间学习增强模型特征提取能力
.

在这里插入图片描述
【6】x = x.transpose(1, 2).contiguous().view(x.shape[0], -1, self.h * self.d_k),这句代码详细解释?
transpose(1, 2):交换张量的第1和第2维度(索引从0开始)
contiguous():确保张量在内存中连续存储(解决transpose导致的非连续问题)
view(…):重塑张量形状(等价于reshape,但要求张量连续)
在这里插入图片描述

层归一化

在这里插入图片描述
在这里插入图片描述

LN的Pytorch实现

class LayerNormalization(nn.Module):

    def __init__(self, features: int, eps: float = 10 ** -6) -> None:
        super().__init__()
        self.eps = eps # 数值稳定性常数
        # 可学习权重
        self.alpha = nn.Parameter(torch.ones(features))  # 缩放参数,初始化为1
        # 可学习偏差
        self.bias = nn.Parameter(torch.zeros(features))  # 偏移参数,初始化为0

    def forward(self, x):
        # x: (batch, seq_len, hidden_size)
        # 保留维度来进行广播
        mean = x.mean(dim=-1, keepdim=True)  # 计算均值:沿隐藏层维度(最后一维)求均值,保持维度为(batch, seq_len, 1)
        std = x.std(dim=-1, keepdim=True)  # 计算标准差:沿隐藏层维度求标准差,保持维度(batch, seq_len, 1)
        # eps 是为了防止除0设置的很小的值
        return self.alpha * (x - mean) / (std + self.eps) + self.bias

BN是在不同样本的同一个维度做归一, 因为样本的个数会发生变化就不太稳定;LN是在同一个样本的不同维度做归一, 这样就可以保证稳定性。
在推理阶段,层归一化(Layer Normalization)中的可学习参数 α(缩放) 和 β(偏移) 的值是直接使用训练过程中优化得到的固定值,无需重新计算或调整。
在这里插入图片描述
举个例子:
在这里插入图片描述
在这里插入图片描述

位置编码

给token加入位置信息。
在这里插入图片描述
在这里插入图片描述
pytorch实现位置编码

class PositionalEncoding(nn.Module):

    def __init__(self, d_model: int, seq_len: int, dropout: float) -> None:
        super().__init__()
        self.d_model = d_model
        self.seq_len = seq_len
        self.dropout = nn.Dropout(dropout)
        # 创建一个空的tensor
        pe = torch.zeros(seq_len, d_model)  # (seq_len, d_model)
        # 创建一个位置向量
        position = torch.arange(0, seq_len, dtype=torch.float).unsqueeze(1)  
        # 计算分母
        div_term = torch.pow(10000.0, -torch.arange(0, d_model, 2, dtype=torch.float) / d_model)  # (d_model / 2)
        # 偶数位调用sin
        pe[:, 0::2] = torch.sin(position * div_term) 
        # 奇数为调用cos
        pe[:, 1::2] = torch.cos(position * div_term) 
        # 增加batch维度
        pe = pe.unsqueeze(0)  # (1, seq_len, d_model)
        # 注册位置编码为一个buffer,这个tensor不会参与训练,但是会随同模型一起被保存或者迁移到GPU。
        self.register_buffer('pe', pe)

    def forward(self, x):
        x = x + (self.pe[:, :x.shape[1], :]).requires_grad_(False) 
        return self.dropout(x)

把这一切组装起来 TODO

我希望你能帮助我改进YOLOv12模型,一个识别水稻稻苗的模型,这个模型我是用来识别之后用来计数苗数的。现在我已经将其骨干网络换为了LSKNet结构,请你在这个基础上加入或者替换结构帮我进一步提升精度,记得后面加入的模块要适配LSKNet结构,下面是一些关于该结构的文件。 LSKNet.py import torch import torch.nn as nn from torch.nn.modules.utils import _pair as to_2tuple from timm.models.layers import DropPath, to_2tuple from functools import partial import warnings __all__ = ['LSKNET_T', 'LSKNET_S'] class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Conv2d(in_features, hidden_features, 1) self.dwconv = DWConv(hidden_features) self.act = act_layer() self.fc2 = nn.Conv2d(hidden_features, out_features, 1) self.drop = nn.Dropout(drop) def forward(self, x): x = self.fc1(x) x = self.dwconv(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class LSKblock(nn.Module): def __init__(self, dim): super().__init__() self.conv0 = nn.Conv2d(dim, dim, 5, padding=2, groups=dim) self.conv_spatial = nn.Conv2d(dim, dim, 7, stride=1, padding=9, groups=dim, dilation=3) self.conv1 = nn.Conv2d(dim, dim // 2, 1) self.conv2 = nn.Conv2d(dim, dim // 2, 1) self.conv_squeeze = nn.Conv2d(2, 2, 7, padding=3) self.conv = nn.Conv2d(dim // 2, dim, 1) def forward(self, x): attn1 = self.conv0(x) attn2 = self.conv_spatial(attn1) attn1 = self.conv1(attn1) attn2 = self.conv2(attn2) attn = torch.cat([attn1, attn2], dim=1) avg_attn = torch.mean(attn, dim=1, keepdim=True) max_attn, _ = torch.max(attn, dim=1, keepdim=True) agg = torch.cat([avg_attn, max_attn], dim=1) sig = self.conv_squeeze(agg).sigmoid() attn = attn1 * sig[:, 0, :, :].unsqueeze(1) + attn2 * sig[:, 1, :, :].unsqueeze(1) attn = self.conv(attn) return x * attn class Attention(nn.Module): def __init__(self, d_model): super().__init__() self.proj_1 = nn.Conv2d(d_model, d_model, 1) self.activation = nn.GELU() self.spatial_gating_unit = LSKblock(d_model) self.proj_2 = nn.Conv2d(d_model, d_model, 1) def forward(self, x): shorcut = x.clone() x = self.proj_1(x) x = self.activation(x) x = self.spatial_gating_unit(x) x = self.proj_2(x) x = x + shorcut return x class Block(nn.Module): def __init__(self, dim, mlp_ratio=4., drop=0., drop_path=0., act_layer=nn.GELU, norm_cfg=None): super().__init__() if norm_cfg: self.norm1 = nn.BatchNorm2d(norm_cfg, dim) self.norm2 = nn.BatchNorm2d(norm_cfg, dim) else: self.norm1 = nn.BatchNorm2d(dim) self.norm2 = nn.BatchNorm2d(dim) self.attn = Attention(dim) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) layer_scale_init_value = 1e-2 self.layer_scale_1 = nn.Parameter( layer_scale_init_value * torch.ones((dim)), requires_grad=True) self.layer_scale_2 = nn.Parameter( layer_scale_init_value * torch.ones((dim)), requires_grad=True) def forward(self, x): x = x + self.drop_path(self.layer_scale_1.unsqueeze(-1).unsqueeze(-1) * self.attn(self.norm1(x))) x = x + self.drop_path(self.layer_scale_2.unsqueeze(-1).unsqueeze(-1) * self.mlp(self.norm2(x))) return x class OverlapPatchEmbed(nn.Module): """ Image to Patch Embedding """ def __init__(self, img_size=224, patch_size=7, stride=4, in_chans=3, embed_dim=768, norm_cfg=None): super().__init__() patch_size = to_2tuple(patch_size) self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride, padding=(patch_size[0] // 2, patch_size[1] // 2)) if norm_cfg: self.norm = nn.BatchNorm2d(norm_cfg, embed_dim) else: self.norm = nn.BatchNorm2d(embed_dim) def forward(self, x): x = self.proj(x) _, _, H, W = x.shape x = self.norm(x) return x, H, W class LSKNet(nn.Module): def __init__(self, img_size=224, in_chans=3, dim=None, embed_dims=[64, 128, 256, 512], mlp_ratios=[8, 8, 4, 4], drop_rate=0., drop_path_rate=0., norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], num_stages=4, pretrained=None, init_cfg=None, norm_cfg=None): super().__init__() assert not (init_cfg and pretrained), \ 'init_cfg and pretrained cannot be set at the same time' if isinstance(pretrained, str): warnings.warn('DeprecationWarning: pretrained is deprecated, ' 'please use "init_cfg" instead') self.init_cfg = dict(type='Pretrained', checkpoint=pretrained) elif pretrained is not None: raise TypeError('pretrained must be a str or None') self.depths = depths self.num_stages = num_stages dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule cur = 0 for i in range(num_stages): patch_embed = OverlapPatchEmbed(img_size=img_size if i == 0 else img_size // (2 ** (i + 1)), patch_size=7 if i == 0 else 3, stride=4 if i == 0 else 2, in_chans=in_chans if i == 0 else embed_dims[i - 1], embed_dim=embed_dims[i], norm_cfg=norm_cfg) block = nn.ModuleList([Block( dim=embed_dims[i], mlp_ratio=mlp_ratios[i], drop=drop_rate, drop_path=dpr[cur + j], norm_cfg=norm_cfg) for j in range(depths[i])]) norm = norm_layer(embed_dims[i]) cur += depths[i] setattr(self, f"patch_embed{i + 1}", patch_embed) setattr(self, f"block{i + 1}", block) setattr(self, f"norm{i + 1}", norm) self.width_list = [i.size(1) for i in self.forward(torch.randn(1, 3, 640, 640))] def freeze_patch_emb(self): self.patch_embed1.requires_grad = False @torch.jit.ignore def no_weight_decay(self): return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better def get_classifier(self): return self.head def reset_classifier(self, num_classes, global_pool=''): self.num_classes = num_classes self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity() def forward_features(self, x): B = x.shape[0] outs = [] for i in range(self.num_stages): patch_embed = getattr(self, f"patch_embed{i + 1}") block = getattr(self, f"block{i + 1}") norm = getattr(self, f"norm{i + 1}") x, H, W = patch_embed(x) for blk in block: x = blk(x) x = x.flatten(2).transpose(1, 2) x = norm(x) x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous() outs.append(x) return outs def forward(self, x): x = self.forward_features(x) # x = self.head(x) return x class DWConv(nn.Module): def __init__(self, dim=768): super(DWConv, self).__init__() self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim) def forward(self, x): x = self.dwconv(x) return x def _conv_filter(state_dict, patch_size=16): """ convert patch embedding weight from manual patchify + linear proj to conv""" out_dict = {} for k, v in state_dict.items(): if 'patch_embed.proj.weight' in k: v = v.reshape((v.shape[0], 3, patch_size, patch_size)) out_dict[k] = v return out_dict def LSKNET_T(): model = LSKNet(depths=[2, 2, 2, 2]) return model def LSKNET_S(): model = LSKNet() return model if __name__ == '__main__': model = LSKNet() inputs = torch.randn((1, 3, 640, 640)) for i in model(inputs): print(i.size()) task.py # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license import contextlib import pickle import re import types from copy import deepcopy from pathlib import Path from .AddModules import * import thop import torch import torch.nn as nn from ultralytics.nn.modules import ( AIFI, C1, C2, C2PSA, C3, C3TR, ELAN1, OBB, PSA, SPP, SPPELAN, SPPF, AConv, ADown, Bottleneck, BottleneckCSP, C2f, C2fAttn, C2fCIB, C2fPSA, C3Ghost, C3k2, C3x, CBFuse, CBLinear, Classify, Concat, Conv, Conv2, ConvTranspose, Detect, DWConv, DWConvTranspose2d, Focus, GhostBottleneck, GhostConv, HGBlock, HGStem, ImagePoolingAttn, Index, Pose, RepC3, RepConv, RepNCSPELAN4, RepVGGDW, ResNetLayer, RTDETRDecoder, SCDown, Segment, TorchVision, WorldDetect, v10Detect, A2C2f, ) from ultralytics.utils import DEFAULT_CFG_DICT, DEFAULT_CFG_KEYS, LOGGER, colorstr, emojis, yaml_load from ultralytics.utils.checks import check_requirements, check_suffix, check_yaml from ultralytics.utils.loss import ( E2EDetectLoss, v8ClassificationLoss, v8DetectionLoss, v8OBBLoss, v8PoseLoss, v8SegmentationLoss, ) from ultralytics.utils.ops import make_divisible from ultralytics.utils.plotting import feature_visualization from ultralytics.utils.torch_utils import ( fuse_conv_and_bn, fuse_deconv_and_bn, initialize_weights, intersect_dicts, model_info, scale_img, time_sync, ) from .AddModules import * class BaseModel(nn.Module): """The BaseModel class serves as a base class for all the models in the Ultralytics YOLO family.""" def forward(self, x, *args, **kwargs): """ Perform forward pass of the model for either training or inference. If x is a dict, calculates and returns the loss for training. Otherwise, returns predictions for inference. Args: x (torch.Tensor | dict): Input tensor for inference, or dict with image tensor and labels for training. *args (Any): Variable length argument list. **kwargs (Any): Arbitrary keyword arguments. Returns: (torch.Tensor): Loss if x is a dict (training), or network predictions (inference). """ if isinstance(x, dict): # for cases of training and validating while training. return self.loss(x, *args, **kwargs) return self.predict(x, *args, **kwargs) def predict(self, x, profile=False, visualize=False, augment=False, embed=None): """ Perform a forward pass through the network. Args: x (torch.Tensor): The input tensor to the model. profile (bool): Print the computation time of each layer if True, defaults to False. visualize (bool): Save the feature maps of the model if True, defaults to False. augment (bool): Augment image during prediction, defaults to False. embed (list, optional): A list of feature vectors/embeddings to return. Returns: (torch.Tensor): The last output of the model. """ if augment: return self._predict_augment(x) return self._predict_once(x, profile, visualize, embed) def _predict_once(self, x, profile=False, visualize=False, embed=None): y, dt, embeddings = [], [], [] # outputs for m in self.model: if m.f != -1: # if not from previous layer x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers if profile: self._profile_one_layer(m, x, dt) if hasattr(m, 'backbone'): x = m(x) if len(x) != 5: # 0 - 5 x.insert(0, None) for index, i in enumerate(x): if index in self.save: y.append(i) else: y.append(None) x = x[-1] # 最后一个输出传给下一层 else: x = m(x) # run y.append(x if m.i in self.save else None) # save output if visualize: feature_visualization(x, m.type, m.i, save_dir=visualize) if embed and m.i in embed: embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1)) # flatten if m.i == max(embed): return torch.unbind(torch.cat(embeddings, 1), dim=0) return x def _predict_augment(self, x): """Perform augmentations on input image x and return augmented inference.""" LOGGER.warning( f"WARNING ⚠️ {self.__class__.__name__} does not support 'augment=True' prediction. " f"Reverting to single-scale prediction." ) return self._predict_once(x) def _profile_one_layer(self, m, x, dt): """ Profile the computation time and FLOPs of a single layer of the model on a given input. Appends the results to the provided list. Args: m (nn.Module): The layer to be profiled. x (torch.Tensor): The input data to the layer. dt (list): A list to store the computation time of the layer. Returns: None """ c = m == self.model[-1] and isinstance(x, list) # is final layer list, copy input as inplace fix flops = thop.profile(m, inputs=[x.copy() if c else x], verbose=False)[0] / 1e9 * 2 if thop else 0 # GFLOPs t = time_sync() for _ in range(10): m(x.copy() if c else x) dt.append((time_sync() - t) * 100) if m == self.model[0]: LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s} module") LOGGER.info(f"{dt[-1]:10.2f} {flops:10.2f} {m.np:10.0f} {m.type}") if c: LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s} Total") def fuse(self, verbose=True): """ Fuse the `Conv2d()` and `BatchNorm2d()` layers of the model into a single layer, in order to improve the computation efficiency. Returns: (nn.Module): The fused model is returned. """ if not self.is_fused(): for m in self.model.modules(): if isinstance(m, (Conv, Conv2, DWConv)) and hasattr(m, "bn"): if isinstance(m, Conv2): m.fuse_convs() m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv delattr(m, "bn") # remove batchnorm m.forward = m.forward_fuse # update forward if isinstance(m, ConvTranspose) and hasattr(m, "bn"): m.conv_transpose = fuse_deconv_and_bn(m.conv_transpose, m.bn) delattr(m, "bn") # remove batchnorm m.forward = m.forward_fuse # update forward if isinstance(m, RepConv): m.fuse_convs() m.forward = m.forward_fuse # update forward if isinstance(m, RepVGGDW): m.fuse() m.forward = m.forward_fuse self.info(verbose=verbose) return self def is_fused(self, thresh=10): """ Check if the model has less than a certain threshold of BatchNorm layers. Args: thresh (int, optional): The threshold number of BatchNorm layers. Default is 10. Returns: (bool): True if the number of BatchNorm layers in the model is less than the threshold, False otherwise. """ bn = tuple(v for k, v in nn.__dict__.items() if "Norm" in k) # normalization layers, i.e. BatchNorm2d() return sum(isinstance(v, bn) for v in self.modules()) < thresh # True if < 'thresh' BatchNorm layers in model def info(self, detailed=False, verbose=True, imgsz=640): """ Prints model information. Args: detailed (bool): if True, prints out detailed information about the model. Defaults to False verbose (bool): if True, prints out the model information. Defaults to False imgsz (int): the size of the image that the model will be trained on. Defaults to 640 """ return model_info(self, detailed=detailed, verbose=verbose, imgsz=imgsz) def _apply(self, fn): """ Applies a function to all the tensors in the model that are not parameters or registered buffers. Args: fn (function): the function to apply to the model Returns: (BaseModel): An updated BaseModel object. """ self = super()._apply(fn) m = self.model[-1] # Detect() if isinstance(m, Detect): # includes all Detect subclasses like Segment, Pose, OBB, WorldDetect m.stride = fn(m.stride) m.anchors = fn(m.anchors) m.strides = fn(m.strides) return self def load(self, weights, verbose=True): """ Load the weights into the model. Args: weights (dict | torch.nn.Module): The pre-trained weights to be loaded. verbose (bool, optional): Whether to log the transfer progress. Defaults to True. """ model = weights["model"] if isinstance(weights, dict) else weights # torchvision models are not dicts csd = model.float().state_dict() # checkpoint state_dict as FP32 csd = intersect_dicts(csd, self.state_dict()) # intersect self.load_state_dict(csd, strict=False) # load if verbose: LOGGER.info(f"Transferred {len(csd)}/{len(self.model.state_dict())} items from pretrained weights") def loss(self, batch, preds=None): """ Compute loss. Args: batch (dict): Batch to compute loss on preds (torch.Tensor | List[torch.Tensor]): Predictions. """ if getattr(self, "criterion", None) is None: self.criterion = self.init_criterion() preds = self.forward(batch["img"]) if preds is None else preds return self.criterion(preds, batch) def init_criterion(self): """Initialize the loss criterion for the BaseModel.""" raise NotImplementedError("compute_loss() needs to be implemented by task heads") class DetectionModel(BaseModel): """YOLOv8 detection model.""" def __init__(self, cfg="yolov8n.yaml", ch=3, nc=None, verbose=True): # model, input channels, number of classes """Initialize the YOLOv8 detection model with the given config and parameters.""" super().__init__() self.yaml = cfg if isinstance(cfg, dict) else yaml_model_load(cfg) # cfg dict if self.yaml["backbone"][0][2] == "Silence": LOGGER.warning( "WARNING ⚠️ YOLOv9 `Silence` module is deprecated in favor of nn.Identity. " "Please delete local *.pt file and re-download the latest model checkpoint." ) self.yaml["backbone"][0][2] = "nn.Identity" # Define model ch = self.yaml["ch"] = self.yaml.get("ch", ch) # input channels if nc and nc != self.yaml["nc"]: LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") self.yaml["nc"] = nc # override YAML value self.model, self.save = parse_model(deepcopy(self.yaml), ch=ch, verbose=verbose) # model, savelist self.names = {i: f"{i}" for i in range(self.yaml["nc"])} # default names dict self.inplace = self.yaml.get("inplace", True) self.end2end = getattr(self.model[-1], "end2end", False) # Build strides m = self.model[-1] # Detect() if isinstance(m, Detect): # includes all Detect subclasses like Segment, Pose, OBB, WorldDetect s = 256 # 2x min stride m.inplace = self.inplace def _forward(x): """Performs a forward pass through the model, handling different Detect subclass types accordingly.""" if self.end2end: return self.forward(x)["one2many"] return self.forward(x)[0] if isinstance(m, (Segment, Pose, OBB)) else self.forward(x) m.stride = torch.tensor([s / x.shape[-2] for x in _forward(torch.zeros(1, ch, s, s))]) # forward self.stride = m.stride m.bias_init() # only run once else: self.stride = torch.Tensor([32]) # default stride for i.e. RTDETR # Init weights, biases initialize_weights(self) if verbose: self.info() LOGGER.info("") def _predict_augment(self, x): """Perform augmentations on input image x and return augmented inference and train outputs.""" if getattr(self, "end2end", False) or self.__class__.__name__ != "DetectionModel": LOGGER.warning("WARNING ⚠️ Model does not support 'augment=True', reverting to single-scale prediction.") return self._predict_once(x) img_size = x.shape[-2:] # height, width s = [1, 0.83, 0.67] # scales f = [None, 3, None] # flips (2-ud, 3-lr) y = [] # outputs for si, fi in zip(s, f): xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) yi = super().predict(xi)[0] # forward yi = self._descale_pred(yi, fi, si, img_size) y.append(yi) y = self._clip_augmented(y) # clip augmented tails return torch.cat(y, -1), None # augmented inference, train @staticmethod def _descale_pred(p, flips, scale, img_size, dim=1): """De-scale predictions following augmented inference (inverse operation).""" p[:, :4] /= scale # de-scale x, y, wh, cls = p.split((1, 1, 2, p.shape[dim] - 4), dim) if flips == 2: y = img_size[0] - y # de-flip ud elif flips == 3: x = img_size[1] - x # de-flip lr return torch.cat((x, y, wh, cls), dim) def _clip_augmented(self, y): """Clip YOLO augmented inference tails.""" nl = self.model[-1].nl # number of detection layers (P3-P5) g = sum(4**x for x in range(nl)) # grid points e = 1 # exclude layer count i = (y[0].shape[-1] // g) * sum(4**x for x in range(e)) # indices y[0] = y[0][..., :-i] # large i = (y[-1].shape[-1] // g) * sum(4 ** (nl - 1 - x) for x in range(e)) # indices y[-1] = y[-1][..., i:] # small return y def init_criterion(self): """Initialize the loss criterion for the DetectionModel.""" return E2EDetectLoss(self) if getattr(self, "end2end", False) else v8DetectionLoss(self) class OBBModel(DetectionModel): """YOLOv8 Oriented Bounding Box (OBB) model.""" def __init__(self, cfg="yolov8n-obb.yaml", ch=3, nc=None, verbose=True): """Initialize YOLOv8 OBB model with given config and parameters.""" super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose) def init_criterion(self): """Initialize the loss criterion for the model.""" return v8OBBLoss(self) class SegmentationModel(DetectionModel): """YOLOv8 segmentation model.""" def __init__(self, cfg="yolov8n-seg.yaml", ch=3, nc=None, verbose=True): """Initialize YOLOv8 segmentation model with given config and parameters.""" super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose) def init_criterion(self): """Initialize the loss criterion for the SegmentationModel.""" return v8SegmentationLoss(self) class PoseModel(DetectionModel): """YOLOv8 pose model.""" def __init__(self, cfg="yolov8n-pose.yaml", ch=3, nc=None, data_kpt_shape=(None, None), verbose=True): """Initialize YOLOv8 Pose model.""" if not isinstance(cfg, dict): cfg = yaml_model_load(cfg) # load model YAML if any(data_kpt_shape) and list(data_kpt_shape) != list(cfg["kpt_shape"]): LOGGER.info(f"Overriding model.yaml kpt_shape={cfg['kpt_shape']} with kpt_shape={data_kpt_shape}") cfg["kpt_shape"] = data_kpt_shape super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose) def init_criterion(self): """Initialize the loss criterion for the PoseModel.""" return v8PoseLoss(self) class ClassificationModel(BaseModel): """YOLOv8 classification model.""" def __init__(self, cfg="yolov8n-cls.yaml", ch=3, nc=None, verbose=True): """Init ClassificationModel with YAML, channels, number of classes, verbose flag.""" super().__init__() self._from_yaml(cfg, ch, nc, verbose) def _from_yaml(self, cfg, ch, nc, verbose): """Set YOLOv8 model configurations and define the model architecture.""" self.yaml = cfg if isinstance(cfg, dict) else yaml_model_load(cfg) # cfg dict # Define model ch = self.yaml["ch"] = self.yaml.get("ch", ch) # input channels if nc and nc != self.yaml["nc"]: LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") self.yaml["nc"] = nc # override YAML value elif not nc and not self.yaml.get("nc", None): raise ValueError("nc not specified. Must specify nc in model.yaml or function arguments.") self.model, self.save = parse_model(deepcopy(self.yaml), ch=ch, verbose=verbose) # model, savelist self.stride = torch.Tensor([1]) # no stride constraints self.names = {i: f"{i}" for i in range(self.yaml["nc"])} # default names dict self.info() @staticmethod def reshape_outputs(model, nc): """Update a TorchVision classification model to class count 'n' if required.""" name, m = list((model.model if hasattr(model, "model") else model).named_children())[-1] # last module if isinstance(m, Classify): # YOLO Classify() head if m.linear.out_features != nc: m.linear = nn.Linear(m.linear.in_features, nc) elif isinstance(m, nn.Linear): # ResNet, EfficientNet if m.out_features != nc: setattr(model, name, nn.Linear(m.in_features, nc)) elif isinstance(m, nn.Sequential): types = [type(x) for x in m] if nn.Linear in types: i = len(types) - 1 - types[::-1].index(nn.Linear) # last nn.Linear index if m[i].out_features != nc: m[i] = nn.Linear(m[i].in_features, nc) elif nn.Conv2d in types: i = len(types) - 1 - types[::-1].index(nn.Conv2d) # last nn.Conv2d index if m[i].out_channels != nc: m[i] = nn.Conv2d(m[i].in_channels, nc, m[i].kernel_size, m[i].stride, bias=m[i].bias is not None) def init_criterion(self): """Initialize the loss criterion for the ClassificationModel.""" return v8ClassificationLoss() class RTDETRDetectionModel(DetectionModel): """ RTDETR (Real-time DEtection and Tracking using Transformers) Detection Model class. This class is responsible for constructing the RTDETR architecture, defining loss functions, and facilitating both the training and inference processes. RTDETR is an object detection and tracking model that extends from the DetectionModel base class. Attributes: cfg (str): The configuration file path or preset string. Default is 'rtdetr-l.yaml'. ch (int): Number of input channels. Default is 3 (RGB). nc (int, optional): Number of classes for object detection. Default is None. verbose (bool): Specifies if summary statistics are shown during initialization. Default is True. Methods: init_criterion: Initializes the criterion used for loss calculation. loss: Computes and returns the loss during training. predict: Performs a forward pass through the network and returns the output. """ def __init__(self, cfg="rtdetr-l.yaml", ch=3, nc=None, verbose=True): """ Initialize the RTDETRDetectionModel. Args: cfg (str): Configuration file name or path. ch (int): Number of input channels. nc (int, optional): Number of classes. Defaults to None. verbose (bool, optional): Print additional information during initialization. Defaults to True. """ super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose) def init_criterion(self): """Initialize the loss criterion for the RTDETRDetectionModel.""" from ultralytics.models.utils.loss import RTDETRDetectionLoss return RTDETRDetectionLoss(nc=self.nc, use_vfl=True) def loss(self, batch, preds=None): """ Compute the loss for the given batch of data. Args: batch (dict): Dictionary containing image and label data. preds (torch.Tensor, optional): Precomputed model predictions. Defaults to None. Returns: (tuple): A tuple containing the total loss and main three losses in a tensor. """ if not hasattr(self, "criterion"): self.criterion = self.init_criterion() img = batch["img"] # NOTE: preprocess gt_bbox and gt_labels to list. bs = len(img) batch_idx = batch["batch_idx"] gt_groups = [(batch_idx == i).sum().item() for i in range(bs)] targets = { "cls": batch["cls"].to(img.device, dtype=torch.long).view(-1), "bboxes": batch["bboxes"].to(device=img.device), "batch_idx": batch_idx.to(img.device, dtype=torch.long).view(-1), "gt_groups": gt_groups, } preds = self.predict(img, batch=targets) if preds is None else preds dec_bboxes, dec_scores, enc_bboxes, enc_scores, dn_meta = preds if self.training else preds[1] if dn_meta is None: dn_bboxes, dn_scores = None, None else: dn_bboxes, dec_bboxes = torch.split(dec_bboxes, dn_meta["dn_num_split"], dim=2) dn_scores, dec_scores = torch.split(dec_scores, dn_meta["dn_num_split"], dim=2) dec_bboxes = torch.cat([enc_bboxes.unsqueeze(0), dec_bboxes]) # (7, bs, 300, 4) dec_scores = torch.cat([enc_scores.unsqueeze(0), dec_scores]) loss = self.criterion( (dec_bboxes, dec_scores), targets, dn_bboxes=dn_bboxes, dn_scores=dn_scores, dn_meta=dn_meta ) # NOTE: There are like 12 losses in RTDETR, backward with all losses but only show the main three losses. return sum(loss.values()), torch.as_tensor( [loss[k].detach() for k in ["loss_giou", "loss_class", "loss_bbox"]], device=img.device ) def predict(self, x, profile=False, visualize=False, batch=None, augment=False, embed=None): """ Perform a forward pass through the model. Args: x (torch.Tensor): The input tensor. profile (bool, optional): If True, profile the computation time for each layer. Defaults to False. visualize (bool, optional): If True, save feature maps for visualization. Defaults to False. batch (dict, optional): Ground truth data for evaluation. Defaults to None. augment (bool, optional): If True, perform data augmentation during inference. Defaults to False. embed (list, optional): A list of feature vectors/embeddings to return. Returns: (torch.Tensor): Model's output tensor. """ y, dt, embeddings = [], [], [] # outputs for m in self.model[:-1]: # except the head part if m.f != -1: # if not from previous layer x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers if profile: self._profile_one_layer(m, x, dt) x = m(x) # run y.append(x if m.i in self.save else None) # save output if visualize: feature_visualization(x, m.type, m.i, save_dir=visualize) if embed and m.i in embed: embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1)) # flatten if m.i == max(embed): return torch.unbind(torch.cat(embeddings, 1), dim=0) head = self.model[-1] x = head([y[j] for j in head.f], batch) # head inference return x class WorldModel(DetectionModel): """YOLOv8 World Model.""" def __init__(self, cfg="yolov8s-world.yaml", ch=3, nc=None, verbose=True): """Initialize YOLOv8 world model with given config and parameters.""" self.txt_feats = torch.randn(1, nc or 80, 512) # features placeholder self.clip_model = None # CLIP model placeholder super().__init__(cfg=cfg, ch=ch, nc=nc, verbose=verbose) def set_classes(self, text, batch=80, cache_clip_model=True): """Set classes in advance so that model could do offline-inference without clip model.""" try: import clip except ImportError: check_requirements("git+https://github.com/ultralytics/CLIP.git") import clip if ( not getattr(self, "clip_model", None) and cache_clip_model ): # for backwards compatibility of models lacking clip_model attribute self.clip_model = clip.load("ViT-B/32")[0] model = self.clip_model if cache_clip_model else clip.load("ViT-B/32")[0] device = next(model.parameters()).device text_token = clip.tokenize(text).to(device) txt_feats = [model.encode_text(token).detach() for token in text_token.split(batch)] txt_feats = txt_feats[0] if len(txt_feats) == 1 else torch.cat(txt_feats, dim=0) txt_feats = txt_feats / txt_feats.norm(p=2, dim=-1, keepdim=True) self.txt_feats = txt_feats.reshape(-1, len(text), txt_feats.shape[-1]) self.model[-1].nc = len(text) def predict(self, x, profile=False, visualize=False, txt_feats=None, augment=False, embed=None): """ Perform a forward pass through the model. Args: x (torch.Tensor): The input tensor. profile (bool, optional): If True, profile the computation time for each layer. Defaults to False. visualize (bool, optional): If True, save feature maps for visualization. Defaults to False. txt_feats (torch.Tensor): The text features, use it if it's given. Defaults to None. augment (bool, optional): If True, perform data augmentation during inference. Defaults to False. embed (list, optional): A list of feature vectors/embeddings to return. Returns: (torch.Tensor): Model's output tensor. """ txt_feats = (self.txt_feats if txt_feats is None else txt_feats).to(device=x.device, dtype=x.dtype) if len(txt_feats) != len(x): txt_feats = txt_feats.repeat(len(x), 1, 1) ori_txt_feats = txt_feats.clone() y, dt, embeddings = [], [], [] # outputs for m in self.model: # except the head part if m.f != -1: # if not from previous layer x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers if profile: self._profile_one_layer(m, x, dt) if isinstance(m, C2fAttn): x = m(x, txt_feats) elif isinstance(m, WorldDetect): x = m(x, ori_txt_feats) elif isinstance(m, ImagePoolingAttn): txt_feats = m(x, txt_feats) else: x = m(x) # run y.append(x if m.i in self.save else None) # save output if visualize: feature_visualization(x, m.type, m.i, save_dir=visualize) if embed and m.i in embed: embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1)) # flatten if m.i == max(embed): return torch.unbind(torch.cat(embeddings, 1), dim=0) return x def loss(self, batch, preds=None): """ Compute loss. Args: batch (dict): Batch to compute loss on. preds (torch.Tensor | List[torch.Tensor]): Predictions. """ if not hasattr(self, "criterion"): self.criterion = self.init_criterion() if preds is None: preds = self.forward(batch["img"], txt_feats=batch["txt_feats"]) return self.criterion(preds, batch) class Ensemble(nn.ModuleList): """Ensemble of models.""" def __init__(self): """Initialize an ensemble of models.""" super().__init__() def forward(self, x, augment=False, profile=False, visualize=False): """Function generates the YOLO network's final layer.""" y = [module(x, augment, profile, visualize)[0] for module in self] # y = torch.stack(y).max(0)[0] # max ensemble # y = torch.stack(y).mean(0) # mean ensemble y = torch.cat(y, 2) # nms ensemble, y shape(B, HW, C) return y, None # inference, train output # Functions ------------------------------------------------------------------------------------------------------------ @contextlib.contextmanager def temporary_modules(modules=None, attributes=None): """ Context manager for temporarily adding or modifying modules in Python's module cache (`sys.modules`). This function can be used to change the module paths during runtime. It's useful when refactoring code, where you've moved a module from one location to another, but you still want to support the old import paths for backwards compatibility. Args: modules (dict, optional): A dictionary mapping old module paths to new module paths. attributes (dict, optional): A dictionary mapping old module attributes to new module attributes. Example: ```python with temporary_modules({"old.module": "new.module"}, {"old.module.attribute": "new.module.attribute"}): import old.module # this will now import new.module from old.module import attribute # this will now import new.module.attribute ``` Note: The changes are only in effect inside the context manager and are undone once the context manager exits. Be aware that directly manipulating `sys.modules` can lead to unpredictable results, especially in larger applications or libraries. Use this function with caution. """ if modules is None: modules = {} if attributes is None: attributes = {} import sys from importlib import import_module try: # Set attributes in sys.modules under their old name for old, new in attributes.items(): old_module, old_attr = old.rsplit(".", 1) new_module, new_attr = new.rsplit(".", 1) setattr(import_module(old_module), old_attr, getattr(import_module(new_module), new_attr)) # Set modules in sys.modules under their old name for old, new in modules.items(): sys.modules[old] = import_module(new) yield finally: # Remove the temporary module paths for old in modules: if old in sys.modules: del sys.modules[old] class SafeClass: """A placeholder class to replace unknown classes during unpickling.""" def __init__(self, *args, **kwargs): """Initialize SafeClass instance, ignoring all arguments.""" pass def __call__(self, *args, **kwargs): """Run SafeClass instance, ignoring all arguments.""" pass class SafeUnpickler(pickle.Unpickler): """Custom Unpickler that replaces unknown classes with SafeClass.""" def find_class(self, module, name): """Attempt to find a class, returning SafeClass if not among safe modules.""" safe_modules = ( "torch", "collections", "collections.abc", "builtins", "math", "numpy", # Add other modules considered safe ) if module in safe_modules: return super().find_class(module, name) else: return SafeClass def torch_safe_load(weight, safe_only=False): """ Attempts to load a PyTorch model with the torch.load() function. If a ModuleNotFoundError is raised, it catches the error, logs a warning message, and attempts to install the missing module via the check_requirements() function. After installation, the function again attempts to load the model using torch.load(). Args: weight (str): The file path of the PyTorch model. safe_only (bool): If True, replace unknown classes with SafeClass during loading. Example: ```python from ultralytics.nn.tasks import torch_safe_load ckpt, file = torch_safe_load("path/to/best.pt", safe_only=True) ``` Returns: ckpt (dict): The loaded model checkpoint. file (str): The loaded filename """ from ultralytics.utils.downloads import attempt_download_asset check_suffix(file=weight, suffix=".pt") file = attempt_download_asset(weight) # search online if missing locally try: with temporary_modules( modules={ "ultralytics.yolo.utils": "ultralytics.utils", "ultralytics.yolo.v8": "ultralytics.models.yolo", "ultralytics.yolo.data": "ultralytics.data", }, attributes={ "ultralytics.nn.modules.block.Silence": "torch.nn.Identity", # YOLOv9e "ultralytics.nn.tasks.YOLOv10DetectionModel": "ultralytics.nn.tasks.DetectionModel", # YOLOv10 "ultralytics.utils.loss.v10DetectLoss": "ultralytics.utils.loss.E2EDetectLoss", # YOLOv10 }, ): if safe_only: # Load via custom pickle module safe_pickle = types.ModuleType("safe_pickle") safe_pickle.Unpickler = SafeUnpickler safe_pickle.load = lambda file_obj: SafeUnpickler(file_obj).load() with open(file, "rb") as f: ckpt = torch.load(f, pickle_module=safe_pickle) else: ckpt = torch.load(file, map_location="cpu") except ModuleNotFoundError as e: # e.name is missing module name if e.name == "models": raise TypeError( emojis( f"ERROR ❌️ {weight} appears to be an Ultralytics YOLOv5 model originally trained " f"with https://github.com/ultralytics/yolov5.\nThis model is NOT forwards compatible with " f"YOLOv8 at https://github.com/ultralytics/ultralytics." f"\nRecommend fixes are to train a new model using the latest 'ultralytics' package or to " f"run a command with an official Ultralytics model, i.e. 'yolo predict model=yolov8n.pt'" ) ) from e LOGGER.warning( f"WARNING ⚠️ {weight} appears to require '{e.name}', which is not in Ultralytics requirements." f"\nAutoInstall will run now for '{e.name}' but this feature will be removed in the future." f"\nRecommend fixes are to train a new model using the latest 'ultralytics' package or to " f"run a command with an official Ultralytics model, i.e. 'yolo predict model=yolov8n.pt'" ) check_requirements(e.name) # install missing module ckpt = torch.load(file, map_location="cpu") if not isinstance(ckpt, dict): # File is likely a YOLO instance saved with i.e. torch.save(model, "saved_model.pt") LOGGER.warning( f"WARNING ⚠️ The file '{weight}' appears to be improperly saved or formatted. " f"For optimal results, use model.save('filename.pt') to correctly save YOLO models." ) ckpt = {"model": ckpt.model} return ckpt, file def attempt_load_weights(weights, device=None, inplace=True, fuse=False): """Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a.""" ensemble = Ensemble() for w in weights if isinstance(weights, list) else [weights]: ckpt, w = torch_safe_load(w) # load ckpt args = {**DEFAULT_CFG_DICT, **ckpt["train_args"]} if "train_args" in ckpt else None # combined args model = (ckpt.get("ema") or ckpt["model"]).to(device).float() # FP32 model # Model compatibility updates model.args = args # attach args to model model.pt_path = w # attach *.pt file path to model model.task = guess_model_task(model) if not hasattr(model, "stride"): model.stride = torch.tensor([32.0]) # Append ensemble.append(model.fuse().eval() if fuse and hasattr(model, "fuse") else model.eval()) # model in eval mode # Module updates for m in ensemble.modules(): if hasattr(m, "inplace"): m.inplace = inplace elif isinstance(m, nn.Upsample) and not hasattr(m, "recompute_scale_factor"): m.recompute_scale_factor = None # torch 1.11.0 compatibility # Return model if len(ensemble) == 1: return ensemble[-1] # Return ensemble LOGGER.info(f"Ensemble created with {weights}\n") for k in "names", "nc", "yaml": setattr(ensemble, k, getattr(ensemble[0], k)) ensemble.stride = ensemble[int(torch.argmax(torch.tensor([m.stride.max() for m in ensemble])))].stride assert all(ensemble[0].nc == m.nc for m in ensemble), f"Models differ in class counts {[m.nc for m in ensemble]}" return ensemble def attempt_load_one_weight(weight, device=None, inplace=True, fuse=False): """Loads a single model weights.""" ckpt, weight = torch_safe_load(weight) # load ckpt args = {**DEFAULT_CFG_DICT, **(ckpt.get("train_args", {}))} # combine model and default args, preferring model args model = (ckpt.get("ema") or ckpt["model"]).to(device).float() # FP32 model # Model compatibility updates model.args = {k: v for k, v in args.items() if k in DEFAULT_CFG_KEYS} # attach args to model model.pt_path = weight # attach *.pt file path to model model.task = guess_model_task(model) if not hasattr(model, "stride"): model.stride = torch.tensor([32.0]) model = model.fuse().eval() if fuse and hasattr(model, "fuse") else model.eval() # model in eval mode # Module updates for m in model.modules(): if hasattr(m, "inplace"): m.inplace = inplace elif isinstance(m, nn.Upsample) and not hasattr(m, "recompute_scale_factor"): m.recompute_scale_factor = None # torch 1.11.0 compatibility # Return model and ckpt return model, ckpt def parse_model(d, ch, verbose=True): # model_dict, input_channels(3) """Parse a YOLO model.yaml dictionary into a PyTorch model.""" import ast # Args legacy = True # backward compatibility for v3/v5/v8/v9 models max_channels = float("inf") nc, act, scales = (d.get(x) for x in ("nc", "activation", "scales")) depth, width, kpt_shape = (d.get(x, 1.0) for x in ("depth_multiple", "width_multiple", "kpt_shape")) if scales: scale = d.get("scale") if not scale: scale = tuple(scales.keys())[0] LOGGER.warning(f"WARNING ⚠️ no model scale passed. Assuming scale='{scale}'.") depth, width, max_channels = scales[scale] if act: Conv.default_act = eval(act) # redefine default activation, i.e. Conv.default_act = nn.SiLU() if verbose: LOGGER.info(f"{colorstr('activation:')} {act}") # print if verbose: LOGGER.info(f"\n{'':>3}{'from':>20}{'n':>3}{'params':>10} {'module':<45}{'arguments':<30}") ch = [ch] layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out backbone = False for i, (f, n, m, args) in enumerate(d["backbone"] + d["head"]): # from, number, module, args t=m m = getattr(torch.nn, m[3:]) if "nn." in m else globals()[m] # get module for j, a in enumerate(args): if isinstance(a, str): with contextlib.suppress(ValueError): args[j] = locals()[a] if a in locals() else ast.literal_eval(a) n = n_ = max(round(n * depth), 1) if n > 1 else n # depth gain if m in { Classify, Conv, ConvTranspose, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, C2fPSA, C2PSA, DWConv, Focus, BottleneckCSP, C1, C2, C2f, C3k2, RepNCSPELAN4, ELAN1, ADown, AConv, SPPELAN, C2fAttn, C3, C3TR, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x, RepC3, PSA, SCDown, C2fCIB, A2C2f, }: c1, c2 = ch[f], args[0] if c2 != nc: # if c2 not equal to number of classes (i.e. for Classify() output) c2 = make_divisible(min(c2, max_channels) * width, 8) if m is C2fAttn: args[1] = make_divisible(min(args[1], max_channels // 2) * width, 8) # embed channels args[2] = int( max(round(min(args[2], max_channels // 2 // 32)) * width, 1) if args[2] > 1 else args[2] ) # num heads args = [c1, c2, *args[1:]] if m in { BottleneckCSP, C1, C2, C2f, C3k2, C2fAttn, C3, C3TR, C3Ghost, C3x, RepC3, C2fPSA, C2fCIB, C2PSA, A2C2f, }: args.insert(2, n) # number of repeats n = 1 if m is C3k2: # for M/L/X sizes legacy = False if scale in "mlx": args[3] = True if m is A2C2f: legacy = False if scale in "lx": # for L/X sizes args.append(True) args.append(1.2) elif m is AIFI: args = [ch[f], *args] elif m in {HGStem, HGBlock}: c1, cm, c2 = ch[f], args[0], args[1] args = [c1, cm, c2, *args[2:]] if m is HGBlock: args.insert(4, n) # number of repeats n = 1 elif m is ResNetLayer: c2 = args[1] if args[3] else args[1] * 4 elif m is nn.BatchNorm2d: args = [ch[f]] #LSKNet elif m in {LSKNET_T,LSKNET_S}: m = m(*args) c2 = m.width_list backbone =True elif m is Concat: c2 = sum(ch[x] for x in f) elif m in {Detect, WorldDetect, Segment, Pose, OBB, ImagePoolingAttn, v10Detect}: args.append([ch[x] for x in f]) if m is Segment: args[2] = make_divisible(min(args[2], max_channels) * width, 8) if m in {Detect, Segment, Pose, OBB}: m.legacy = legacy elif m is RTDETRDecoder: # special case, channels arg must be passed in index 1 args.insert(1, [ch[x] for x in f]) elif m in {CBLinear, TorchVision, Index}: c2 = args[0] c1 = ch[f] args = [c1, c2, *args[1:]] elif m is CBFuse: c2 = ch[f[-1]] else: c2 = ch[f] # m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module # t = str(m)[8:-2].replace("__main__.", "") # module type # m_.np = sum(x.numel() for x in m_.parameters()) # number params # m_.i, m_.f, m_.type = i, f, t # attach index, 'from' index, type # if verbose: # LOGGER.info(f"{i:>3}{str(f):>20}{n_:>3}{m_.np:10.0f} {t:<45}{str(args):<30}") # print # save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist # layers.append(m_) # if i == 0: # ch = [] # ch.append(c2) if isinstance(c2, list): backbone = True m_ = m m_.backbone = True else: m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module t = str(m)[8:-2].replace('__main__.', '') # module type m.np = sum(x.numel() for x in m_.parameters()) # number params m_.i, m_.f, m_.type = i + 4 if backbone else i, f, t # attach index, 'from' index, type if verbose: LOGGER.info(f'{i:>3}{str(f):>20}{n_:>3}{m.np:10.0f} {t:<45}{str(args):<30}') # print save.extend(x % (i + 4 if backbone else i) for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist layers.append(m_) if i == 0: ch = [] if isinstance(c2, list): ch.extend(c2) for _ in range(5 - len(ch)): ch.insert(0, 0) else: ch.append(c2) return nn.Sequential(*layers), sorted(save) def yaml_model_load(path): """Load a YOLOv8 model from a YAML file.""" path = Path(path) if path.stem in (f"yolov{d}{x}6" for x in "nsmlx" for d in (5, 8)): new_stem = re.sub(r"(\d+)([nslmx])6(.+)?$", r"\1\2-p6\3", path.stem) LOGGER.warning(f"WARNING ⚠️ Ultralytics YOLO P6 models now use -p6 suffix. Renaming {path.stem} to {new_stem}.") path = path.with_name(new_stem + path.suffix) unified_path = re.sub(r"(\d+)([nslmx])(.+)?$", r"\1\3", str(path)) # i.e. yolov8x.yaml -> yolov8.yaml yaml_file = check_yaml(unified_path, hard=False) or check_yaml(path) d = yaml_load(yaml_file) # model dict d["scale"] = guess_model_scale(path) d["yaml_file"] = str(path) return d def guess_model_scale(model_path): """ Takes a path to a YOLO model's YAML file as input and extracts the size character of the model's scale. The function uses regular expression matching to find the pattern of the model scale in the YAML file name, which is denoted by n, s, m, l, or x. The function returns the size character of the model scale as a string. Args: model_path (str | Path): The path to the YOLO model's YAML file. Returns: (str): The size character of the model's scale, which can be n, s, m, l, or x. """ try: return re.search(r"yolo[v]?\d+([nslmx])", Path(model_path).stem).group(1) # noqa, returns n, s, m, l, or x except AttributeError: return "" def guess_model_task(model): """ Guess the task of a PyTorch model from its architecture or configuration. Args: model (nn.Module | dict): PyTorch model or model configuration in YAML format. Returns: (str): Task of the model ('detect', 'segment', 'classify', 'pose'). Raises: SyntaxError: If the task of the model could not be determined. """ def cfg2task(cfg): """Guess from YAML dictionary.""" m = cfg["head"][-1][-2].lower() # output module name if m in {"classify", "classifier", "cls", "fc"}: return "classify" if "detect" in m: return "detect" if m == "segment": return "segment" if m == "pose": return "pose" if m == "obb": return "obb" # Guess from model cfg if isinstance(model, dict): with contextlib.suppress(Exception): return cfg2task(model) # Guess from PyTorch model if isinstance(model, nn.Module): # PyTorch model for x in "model.args", "model.model.args", "model.model.model.args": with contextlib.suppress(Exception): return eval(x)["task"] for x in "model.yaml", "model.model.yaml", "model.model.model.yaml": with contextlib.suppress(Exception): return cfg2task(eval(x)) for m in model.modules(): if isinstance(m, Segment): return "segment" elif isinstance(m, Classify): return "classify" elif isinstance(m, Pose): return "pose" elif isinstance(m, OBB): return "obb" elif isinstance(m, (Detect, WorldDetect, v10Detect)): return "detect" # Guess from model filename if isinstance(model, (str, Path)): model = Path(model) if "-seg" in model.stem or "segment" in model.parts: return "segment" elif "-cls" in model.stem or "classify" in model.parts: return "classify" elif "-pose" in model.stem or "pose" in model.parts: return "pose" elif "-obb" in model.stem or "obb" in model.parts: return "obb" elif "detect" in model.parts: return "detect" # Unable to determine task from model LOGGER.warning( "WARNING ⚠️ Unable to automatically guess model task, assuming 'task=detect'. " "Explicitly define task for your model, i.e. 'task=detect', 'segment', 'classify','pose' or 'obb'." ) return "detect" # assume detect __init__.py from .LSKNet import * yolov12-LSKNet.yaml # YOLOv12 🚀, AGPL-3.0 license # YOLOv12 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect # Parameters nc: 1 # number of classes scales: # model compound scaling constants, i.e. 'model=yolov12n.yaml' will call yolov12.yaml with scale 'n' # [depth, width, max_channels] n: [0.50, 0.25, 1024] # summary: 465 layers, 2,603,056 parameters, 2,603,040 gradients, 6.7 GFLOPs s: [0.50, 0.50, 1024] # summary: 465 layers, 9,285,632 parameters, 9,285,616 gradients, 21.7 GFLOPs m: [0.50, 1.00, 512] # summary: 501 layers, 20,201,216 parameters, 20,201,200 gradients, 68.1 GFLOPs l: [1.00, 1.00, 512] # summary: 831 layers, 26,454,880 parameters, 26,454,864 gradients, 89.7 GFLOPs x: [1.00, 1.50, 512] # summary: 831 layers, 59,216,928 parameters, 59,216,912 gradients, 200.3 GFLOPs # YOLO12n backbone backbone: # [from, repeats, module, args] - [-1, 1, LSKNET_T, []] # YOLO12n head head: - [-1, 1, nn.Upsample, [None, 2, "nearest"]] - [[-1, 3], 1, Concat, [1]] # cat backbone P4 - [-1, 2, A2C2f, [512, False, -1]] # 8 - [-1, 1, nn.Upsample, [None, 2, "nearest"]] - [[-1, 2], 1, Concat, [1]] # cat backbone P3 - [-1, 2, A2C2f, [256, False, -1]] # 11 - [-1, 1, Conv, [256, 3, 2]] - [[-1, 7], 1, Concat, [1]] # cat head P4 - [-1, 2, A2C2f, [512, False, -1]] # 14 - [-1, 1, Conv, [512, 3, 2]] - [[-1, 4], 1, Concat, [1]] # cat head P5 - [-1, 2, C3k2, [1024, True]] # 17 (P5/32-large) - [[10, 13, 16], 1, Detect, [nc]] # Detect(P3, P4, P5) 下面是我训练时所使用到的代码 train.py from ultralytics import YOLO def main(): model = YOLO("yolov12-LSKNet.yaml") results = model.train( data='./data.yaml', epochs=200, imgsz=640, batch=16, amp=False, device=0, save=True, project='runs/train', name='rice_seedling-200L+NWD', verbose=True, workers=2, cache=True, ) if __name__ == '__main__': from multiprocessing import freeze_support freeze_support() main()
09-01
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值