Ascend/model-migration

Model Migration Skill

Clone open-source model repositories and apply NPU adaptation patches for Ascend NPU training.

When to Invoke

• User wants to migrate open-source model to Ascend NPU
• User needs to clone model repository
• User needs to apply DrivingSDK patches
• User asks about preparing model code for NPU training

Information to Collect

Ask user for the following:

1. Model name (e.g., BEVFormer)
2. Working directory path on server
3. Dataset path (if already prepared)
4. Pre-trained weights path (if available)

重要：DrivingSDK仓库地址为 https://gitcode.com/Ascend/DrivingSDK，包含所有必要的patch文件和训练脚本。

Workflow

Step 1: Clone Repository in DrivingSDK model_examples Directory

重要：必须在 DrivingSDK/model_examples/<model_name>/ 目录下克隆模型仓库，而不是在外部创建新目录。

# 进入DrivingSDK的model_examples目录
cd <DrivingSDK_path>/model_examples/BEVFormer

# 克隆模型仓库（会在当前目录下创建BEVFormer子目录）
git clone https://github.com/fundamentalvision/BEVFormer.git

Step 2: Copy Patch to Model Directory

# 将patch文件复制到克隆的模型目录中
cp bev_former_config.patch BEVFormer/

Step 3: Checkout Specific Commit

cd BEVFormer
git checkout 66b65f3a1f58caf0507cb2a971b9c0e7f842376c

Step 4: Apply NPU Patches

# 在模型目录内应用patch
git apply --reject --whitespace=fix bev_former_config.patch

Step 5: Setup Dataset and Weight Links

# 创建数据和权重目录
mkdir -p data ckpts

# 设置数据集软链接
ln -sf <dataset_path>/nuscenes data/nuscenes
ln -sf <dataset_path>/can_bus data/can_bus

# 设置预训练权重软链接
ln -sf <weight_path>/r101_dcn_fcos3d_pretrain.pth ckpts/

Step 6: Return to Parent Directory for Training

# 返回到model_examples/BEVFormer目录
cd ..

# 执行训练脚本
bash test/train_performance_8p_base_fp32.sh --batch-size=1 --num-npu=8

Supported Models

BEVFormer

Repository: https://github.com/fundamentalvision/BEVFormer.git

Commit: 66b65f3a1f58caf0507cb2a971b9c0e7f842376c

Patch: bev_former_config.patch

Dataset Required:

• nuscenes (processed)
• can_bus

Weights Required:

• r101_dcn_fcos3d_pretrain.pth

Commands:

# 进入DrivingSDK的model_examples/BEVFormer目录
cd <DrivingSDK_path>/model_examples/BEVFormer

# 克隆BEVFormer仓库
git clone https://github.com/fundamentalvision/BEVFormer.git

# 复制patch到BEVFormer目录
cp bev_former_config.patch BEVFormer/

# 进入BEVFormer目录
cd BEVFormer

# 切换到指定commit
git checkout 66b65f3a1f58caf0507cb2a971b9c0e7f842376c

# 应用patch
git apply --reject --whitespace=fix bev_former_config.patch

# 创建数据和权重目录
mkdir -p data ckpts

# 设置数据集软链接
ln -sf <dataset_path>/nuscenes data/nuscenes
ln -sf <dataset_path>/can_bus data/can_bus

# 设置预训练权重软链接
ln -sf <weight_path>/r101_dcn_fcos3d_pretrain.pth ckpts/

# 返回上级目录执行训练
cd ..
bash test/train_performance_8p_base_fp32.sh --batch-size=1 --num-npu=8

Network Strategy

| Situation | Strategy | |-----------|----------| | Server can access GitHub | Direct git clone on server | | Server network timeout | Configure proxy or clone locally → scp to server | | Proxy available | Configure proxy: export http_proxy=http://proxy:port |

重要：所有GitHub仓库克隆都应使用GitHub官方链接，不使用镜像站点。如果网络超时，优先配置代理。

Local Clone + SCP Transfer

If server cannot access GitHub:

# On local machine
git clone https://github.com/fundamentalvision/BEVFormer.git
cd BEVFormer
git checkout 66b65f3a1f58caf0507cb2a971b9c0e7f842376c

# Transfer to server
scp -r BEVFormer <user>@<server>:<working_directory>/

Verification

After migration, verify setup:

# Check directory structure
ls -la <model_directory>/

# Check patch applied
cd <model_directory> && git status

# Check data links
ls -la data/
ls -la ckpts/

Troubleshooting

Patch Application Fails

1. Check if correct commit is checked out
2. Check for .rej files for manual merge
3. Verify patch file is for correct model version

Dataset Not Found

1. Verify dataset path exists
2. Check symbolic link is correct
3. Ask user for correct dataset path

Weights Not Found

1. Verify weight file exists
2. Check weight file name matches config
3. Ask user for correct weight path

Reference

• DrivingSDK model examples: DrivingSDK/model_examples/
• Model-specific README: DrivingSDK/model_examples/<model>/README.md