yolov8载入预训练权重

问题中提到的方法可以通过在yolov8模型的代码中插入一段代码来实现载入预训练权重。具体插入位置如下图所示：

def _new(self, m):
    for i in m.names:
        if isinstance(m.names[i], nn.BatchNorm2d):
            m.names[i] = nn.SyncBatchNorm.convert_sync_batchnorm(m.names[i])
    if isinstance(m, nn.SyncBatchNorm):
        m = nn.SyncBatchNorm.convert_sync_batchnorm(m)
    if isinstance(m, nn.Conv2d):
        n = nn.Conv2d(m.in_channels, m.out_channels, m.kernel_size, m.stride, m.padding, m.dilation, m.groups)
        n.weight.data = m.weight.data.clone()
        n.bias.data = m.bias.data.clone()
        n.scale = m.scale.clone()
        return n
    elif isinstance(m, nn.BatchNorm2d):
        n = nn.BatchNorm2d(m.num_features, m.eps, m.momentum, m.affine, m.track_running_stats)
        n.weight.data = m.weight.data.clone()
        n.bias.data = m.bias.data.clone()
        n.running_mean.data = m.running_mean.data.clone()
        n.running_var.data = m.running_var.data.clone()
        return n
    elif isinstance(m, nn.Linear):
        n = nn.Linear(m.in_features, m.out_features, m.bias is not None)
        n.weight.data = m.weight.data.clone()
        n.bias.data = m.bias.data.clone()
        return n
    else:
        return m

def load_pretrained_weights(self, file, nclude=None):
        msg = f"Transferred {n}/{n_items} items from {file}\n"
        try:
            # load
            ckpt = torch.load(file, map_location=self.device)
            # exclude keys
            if include:
                ckpt['model'] = {k: v for k, v in ckpt['model'].items() if k in include}
            # load
            self.model.load_state_dict(ckpt['model'], strict=False)
            # report
            print(msg, end='')
        except KeyError as e:
            s = 'missing keys in pretrained weights: {e}\n'
            if allow_missing_keys:
                print(f"{msg}warning: {s}")
            else:
                raise KeyError(f"{msg}error: {s}")
        except Exception as e:
            raise type(e)(f"{msg}{e}")

# 在yolov8->ultralytics->yolo->engine->model.py的def _new函数中插入以下代码
ckpt = torch.load("yolov8.pt")
csd = ckpt['model'].float().state_dict()
csd = intersect_dicts(csd, self.model.state_dict())
self.model.load_state_dict(csd, strict=False)
print(f'Transferred {len(csd)}/{len(self.model.state_dict())} items')