### （方式一）从github下载并解压F3NET模型的源码，参考下述修改test.py/net.py/train.py文件，完成适配
git clone https://github.com/weijun-arc/F3Net.git

### （方式二）从gitcode下载并解压已适配后的F3NET模型的代码，即F3Net.tar压缩包
git clone https://atomgit.com/Ascend-SACT/F3NET.git

vim docker_start.sh
# 脚本内容如下
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 
CONTAINER_NAME=容器名称
IMAGE=镜像ID 
docker run -itd --privileged --name=$CONTAINER_NAME --ipc=host \
    --device=/dev/davinci_manager \
    --device=/dev/devmm_svm \
    --device=/dev/hisi_hdc \
    -v /usr/local/sbin:/usr/local/sbin \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /tmp:/tmp \
    -v /mnt:/mnt \
    -v /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime \
    -v /home:/home \
    -v /data:/data \
    -w /home \
    $IMAGE \
    /bin/bash
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 
# 进入容器
bash docker_start.sh
docker exec -it <容器名称> bash

### 添加torch_npu和torchair
import torch_npu
import torchair
from torchair.configs.compiler_config import CompilerConfig

### 修改设备选择和添加torchair配置
class Test(object):
    def __init__(self, Dataset, Network, path):
        ...
        # self.net.cuda()
        self.net.npu()

        config_torchair = CompilerConfig()
        npu_backend = torchair.get_npu_backend(compiler_config=config_torchair)
        self.net = torch.compile(self.net, backend=npu_backend)  
	
def show(self):
        with torch.no_grad():
            for image, mask, shape, name in self.loader:
                # image, mask = image.cuda().float(), mask.cuda().float()
                image, mask = image.npu().float(), mask.npu().float()

    def save(self):
        with torch.no_grad():

            for image, mask, shape, name in self.loader:
                # image = image.cuda().float()  
                image = image.npu().float()

###由于pytorch版本>2.6，需要进行适配
class ResNet(nn.Module):
    ...
    def initialize(self):
        # self.load_state_dict(torch.load('../res/resnet50-19c8e357.pth'), strict=False)
        self.load_state_dict(torch.load('../res/resnet50-19c8e357.pth',weights_only=False), strict=False)

class F3Net(nn.Module):
    ...
    def initialize(self):
        if self.cfg.snapshot:
            # self.load_state_dict(torch.load(self.cfg.snapshot))
            self.load_state_dict(torch.load(self.cfg.snapshot,weights_only=False))
        else:
            weight_init(self)

由于apex中amp调用和cuda直接相关，所有移除amp的调用，直接使用torch.npu.amp。
# from apex import amp
   
def train(Dataset, Network):
    ...
    # 移除apex.amp.initialize 初始化
    # net, optimizer = amp.initialize(net, optimizer, opt_level='O2')
    net = net.to("npu") 
    scaler = torch.npu.amp.GradScaler()

    sw = SummaryWriter(cfg.savepath)
    global_step = 0

    for epoch in range(cfg.epoch):
        optimizer.param_groups[0]['lr'] = (1-abs((epoch+1)/(cfg.epoch+1)*2-1))*cfg.lr*0.1
        optimizer.param_groups[1]['lr'] = (1-abs((epoch+1)/(cfg.epoch+1)*2-1))*cfg.lr

        for step, (image, mask) in enumerate(loader):
            image, mask = image.npu().float(), mask.npu().float()

            #NPU AMP 自动混合精度上下文
            with torch.npu.amp.autocast():
                out1u, out2u, out2r, out3r, out4r, out5r = net(image)

                loss1u = structure_loss(out1u, mask)
                loss2u = structure_loss(out2u, mask)

                loss2r = structure_loss(out2r, mask)
                loss3r = structure_loss(out3r, mask)
                loss4r = structure_loss(out4r, mask)
                loss5r = structure_loss(out5r, mask)
                loss   = (loss1u+loss2u)/2+loss2r/2+loss3r/4+loss4r/8+loss5r/16

            optimizer.zero_grad()
            # with amp.scale_loss(loss, optimizer) as scale_loss:
            #     scale_loss.backward()
            # optimizer.step()

            scaler.scale(loss).backward()
            scaler.step(optimizer)
            scaler.update()

进入./src目录后执行python test.py