The Jupyter notebook helps users to analyze the performance benefit from using Intel® Optimizations for PyTorch with the IPEX package. The notebooks will show the comparison between the performance of stock PyTorch and Intel's optimizations for PyTorch.
| Analysis Type | Notebook | Notes |
|---|---|---|
| stock vs. Intel PyTorch | benchmark_perf_comparison | Compare performance between Stock and Intel PyTorch among different image-recognition models |
- profiling:
- profile_utils.py: python classes that help plotting and data processing
- topo.ini : configurations for supported models
- Configurations include the model name and precision
- Users do not need to change those configurations. The just have to select from one of them using topo_index in the notebook.
To run this experiment please install the latest version of miniforge with python 3.8. The installer can be downloaded from following link
- Create conda env:
$conda create -n stock-torch python=3.7 matplotlib psutil pandas gitpython - Activate the created conda env:
$conda activate stock-torch - Install ipython:
(stock-torch) $conda install 'ipython<7.0.0' ipykernel - Install stock PyTorch with a specific version:
(stock-torch) $conda install pytorch==1.7.0 torchvision cpuonly -c pytorch - Install extra needed package:
(stock-torch) $pip install cxxfilt - Deactivate conda env:
(stock-torch) $conda deactivate - Register the kernel to Jupyter NB:
$~/miniforge/envs/stock-torch/bin/python -m ipykernel install --user --name=stock-torch
NOTE: Please change the python path if you have a different folder path for anaconda3/miniforge. After profiling, users can remove the kernel from Jupyter NB with
$jupyter kernelspec uninstall stock-torch.
- Create conda env:
$conda create -n intel-torch python=3.7 matplotlib psutil pandas gitpython - Activate the created conda env:
$conda activate intel-torch - Install ipython:
(intel-torch) $conda install 'ipython<7.0.0' ipykernel - Update the mkl from Intel's channel"
(intel-torch) $conda update -c https://software.repos.intel.com/python/conda/ mkl - Install Intel-optimized torch with a specific version:
(intel-torch) $conda install -c https://software.repos.intel.com/python/conda/ pytorch - Install IPEX from intel's conda channel:
(intel-torch) $conda install -c https://software.repos.intel.com/python/conda/ intel-extension-for-pytorch - Install Torchvision from PyTorch's channel:
(intel-torch) $conda install torchvision cpuonly -c pytorch - Install extra needed package:
(intel-torch) $pip install cxxfilt - Deactivate conda env:
(intel-torch) $conda deactivate - Register the kernel to Jupyter NB:
$~/miniforge/envs/intel-torch/bin/python -m ipykernel install --user --name=intel-torch
NOTE: Please change the python path if you have a different folder path for anaconda3/miniforge. After profiling, users can remove the kernel from Jupyter NB with
$jupyter kernelspec uninstall intel-torch.
- Clone the Intel Model Zoo:
$git clone https://github.com/IntelAI/models.git - Browse to the
models/docs/notebooks/perf_analysis/pytorchfolder - Launch Jupyter notebook:
$jupyter notebook --ip=0.0.0.0 - Follow the instructions to open the URL with the token in your browser
- Click the notebook file
benchmark_perf_comparison.ipynb.
Note: For "stock v.s. Intel PyTorch" analysis type, please change your Jupyter notebook kernel to either "stock-torch" or "intel-torch"
- Run through every cell of the notebook one by one
NOTE: For "stock vs. Intel PyTorch" analysis type, in order to compare between stock and Intel-optimized PyTorch results in Section 5 of the notebook, users need to run all cells before the comparison section with both stock-torch and intel-torch kernels.