code网址
1 身体注意力采样需要安装工具
Install body-atttention sampler related tools (MobulaOP)
# Clone the project
git clone https://github.com/wkcn/MobulaOP
# Enter the directory
cd MobulaOP
# Install MobulaOP
pip install -v -e .
使用时如下 attention_sampler文件夹
主线测试逻辑
1 main_joint.py->
Solver_joint.py
2
1 训练 SGL-KRN
5.1 SGL-KRN
cd demo
python main_SGL_KRN.py
After training, the resulting model will be stored under results/sgl_krn/run-* folder.
2 训练PA-KRN
5.2 PA-KRN
The whole system can be trained in an end-to-end manner.
To get finer results, we first train CLM and FSM sequentially and then combine them to fine-tune.
整个系统可以端到端的方式进行训练。
为了得到更精确的结果,我们首先依次训练CLM和FSM,然后将它们结合起来进行微调。
cd demo
- Train CLM.
python main_clm.py
After training, the resulting model will be stored under results/clm/run-* folder.
- Train FSM.
python main_fsm.py --clm_model path/to/pretrained/clm/folder/
After training, the resulting model will be stored under results/fsm/run-* folder, and * changes accordingly. 'path/to/pretrained/clm/folder/' is the path to pretrained clm folder.
- Train PA-KRN.
python main_joint.py --clm_model path/to/pretrained/clm/folder/ --fsm_model path/to/pretrained/fsm/folder/
After training, the resulting model will be stored under results/joint/run-* folder. 'net_*.pth' is the parameter of CLM model and '.pth' is the parameter of FSM model.