安装conda
wget https://repo.anaconda.com/miniconda/Miniconda3-py310_25.1.1-2-Linux-x86_64.sh
chmod +x Miniconda3-py310_25.1.1-2-Linux-x86_64.sh
./Miniconda3-py310_25.1.1-2-Linux-x86_64.sh
echo 'export PATH=~/miniconda3/bin:$PATH' >> ~/.bashrc
source ~/.bashrc
# 非登录 shell 中手动 source conda
source /root/miniconda3/etc/profile.d/conda.sh
创建并激活环境
conda env create -f FlexCAD-main/environments.yaml
conda activate cadllm
conda env update -f environment.yml
克隆仓库
git clone https://github.com/microsoft/FlexCAD.git
cd FlexCAD-main
按照README，下载数据
gdown --id 1so_CCGLIhqGEDQxMoiR--A4CQk4MjuOp
or
https://drive.google.com/file/d/1so_CCGLIhqGEDQxMoiR--A4CQk4MjuOp/view
unzip cad_data.zip
按照README，解压后转换数据
python3 utils/convert.py --in_path ./cad_data/test.pkl --out_path ./cad_data/processed_data/test.pkl --circle_type ldru.pkl中存的是list数据，其中一条如下：
{'name': '0000/00009254', 'len_xy': np.int64(23), 'len_ext': np.int64(19), 'len_pix': np.int64(23), 'len_cmd': np.int64(11), 'num_se': 1, 'se_xy': [array([[13, 13],[ 3, 3],[13, 57],[ 3, 3],[57, 57],[ 3, 3],[57, 13],[ 3, 3],[ 2, 2],[35, 33],[35, 24],[39, 29],[31, 29],[ 3, 3],[ 2, 2],[35, 45],[35, 36],[39, 41],[31, 41],[ 3, 3],[ 2, 2],[ 1, 1],[ 0, 0]])], 'se_cmd': [array([3, 3, 3, 3, 2, 5, 2, 5, 2, 1, 0])], 'se_pix': [array([ 589, 3, 3405, 3, 3449, 3, 633, 3, 2, 1891, 1315,1639, 1631, 3, 2, 2659, 2083, 2407, 2399, 3, 2, 1,0])], 'se_ext': [array([32, 33, 32, 32, 32, 2, 3, 2, 2, 2, 3, 3, 2, 2, 1, 63, 32,32, 0])]}
转换后如下：
'line,9,9 <curve_end> line,9,53 <curve_end> line,53,53 <curve_end> line,53,9 <curve_end> <loop_end> circle,27,25,31,20,35,25,31,29 <curve_end> <loop_end> circle,27,37,31,32,35,37,31,41 <curve_end> <loop_end> <face_end> <sketch_end> add,31,32,31,31,31,0,1,0,0,0,1,1,0,0,62,31,31 <extrusion_end>'
下载模型(FlexCAD只存储了Lora层，还需要下载Llama-3-8B)
git clone https://huggingface.co/microsoft/FlexCAD
git clone https://www.modelscope.cn/LLM-Research/Meta-Llama-3-8B.git
模型存放目录：
./FlexCAD
./Meta-Llama-3-8B
编辑 sample.py
#model_id = "meta-llama/Meta-Llama-3-8B"
#print(f"Model size: {model_id}")
#pipeline = transformers.pipeline("text2text-generation",
# model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map='auto')local_model_path = "./Meta-Llama-3-8B"
pipeline = transformers.pipeline("text-generation",model=local_model_path,tokenizer=local_model_path, # 显式指定 tokenizer 路径（可选，但推荐）model_kwargs={"torch_dtype": torch.bfloat16},device_map="auto"
)tokenizer = pipeline.tokenizer
model = pipeline.model
推理：
CUDA_VISIBLE_DEVICES=<gpu_id> python3 sample.py --model_path <model_path> --num_samples <num_samples> --model_name <model_name> --data_path <data_path> --mask_type <mask_type>
CUDA_VISIBLE_DEVICES=0 python3 sample.py --model_path FlexCAD --num_samples 1 --data_path ./cad_data/processed_data/test.pkl --mask_type unconditional
可视化（LLM: string.json -> obj -> stl/step）
python调用occ可能失败，需要安装pythonocc-core需要的环境
apt install -y libgl1-mesa-glx libglib2.0-0 libsm6 libxrender1 libxext6
可视化
python3 utils/parser.py --in_path conditional_samples_1_unconditional_mask.json --out_path cad_data/output/
timeout 180 python3 utils/visual_obj.py --data_folder cad_data/output/