DeepView.Profile is a tool to profile and debug the training performance of PyTorch neural networks.

• Installation
• Usage example
• Development Environment Setup
• Release History
• Meta
• Contributing

Installation

DeepView.Profile works with GPU-based neural networks that are implemented in PyTorch.

To run DeepView.Profile, you need:

• A system equipped with an NVIDIA GPU
• Python 3.9+
• PyTorch 1.1.0+ with CUDA
  • NOTE: We assume you have the correct version of PyTorch installed for their GPU. Default PyTorch installation on Linux distros might not have CUDA support. If you see error similar to below, your PyTorch version is incompatible with your version of CUDA. You can download the appropriate version from the PyTorch site

    The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_70.
    If you want to use the NVIDIA GeForce RTX 3060 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/
    

• For new RTX 4000 GPUs you need to install pytorch with cuda11.8 [pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118] otherwise CUPTI will not initialize correctly

Installation from PyPi

Installing with Pip

pip install deepview-profile

Installation from source

git clone https://github.com/CentML/DeepView.Profile
cd DeepView.Profile
poetry install
poetry run deepview --help

Usage example

To use DeepView.Profile in your project, you need to first write an entry point file, which is a regular Python file that describes how your model is created and trained. See the Entry Point for more information.

Once your entry point file is ready, there are two ways to profile interactive profiling and standalone profiling.

Interactive Profiling

Interactive profiling is done with VSCode with the DeepView.Explore plugin. Install the plugin in VSCode and run the profiling session to interactively profile your models.

python3 -m deepview_profile interactive

Standalone Profiling

Standalone profiling is useful when you just want access to DeepView.Profile's profiling functionality. DeepView.Profile will save the profiling results (called a "report") into a SQLite database file that you can then query yourself. We describe the database schema for DeepView.Profile's run time and memory reports in the Run Time Report Format and Memory Report Format pages respectively.

To have DeepView.Profile perform run time profiling, you use the deepview time subcommand. In addition to the entry point file, you also need to specify the file where you want DeepView.Profile to save the run time profiling report using the --output or -o flag.

python3 -m deepview_profile time entry_point.py --output my_output_file.sqlite

Launching memory profiling is almost the same as launching run time profiling. You just need to use deepview memory instead of deepview time.

python3 -m deepview_profile memory entry_point.py --output my_output_file.sqlite

To export various available analysis to json file, you may use deepview analysis --all command for exact entry point and output file. It is required to later view the analysis on the web viewer.

It is also possible to run several optional analysis. There are such analysis available: --measure-breakdown, --measure-throughput, --habitat-predict, --measure-utilization, --energy-compute, --exclude-source

python3 -m deepview_profile analysis entry_point.py --all --exclude-source --output=complete_analysis.json 

--exclude-source option allows not adding encodedFiles section to output, that is available for --measure-breakdown analysis

or various combinations of optional analysis

python3 -m deepview_profile analysis entry_point.py --measure-breakdown --measure-throughput --habitat-predict --measure-utilization --energy-compute --output=various_analysis.json

Development Environment Setup

From the project root, do

poetry install

Release History

See Releases

Meta

DeepView.Profile began as a research project at the University of Toronto in collaboration with Geofrey Yu, Tovi Grossman and Gennady Pekhimenko.

The accompanying research paper appears in the proceedings of UIST'20. If you are interested, you can read a preprint of the paper here.

If you use DeepView.Profile in your research, please consider citing our paper:

@inproceedings{skyline-yu20,
  title = {{Skyline: Interactive In-Editor Computational Performance Profiling
    for Deep Neural Network Training}},
  author = {Yu, Geoffrey X. and Grossman, Tovi and Pekhimenko, Gennady},
  booktitle = {{Proceedings of the 33rd ACM Symposium on User Interface
    Software and Technology (UIST'20)}},
  year = {2020},
}

It is distributed under Apache 2.0 license. See LICENSE and NOTICE for more information.

Contributing

Check out CONTRIBUTING.md for more information on how to help with DeepView.Profile.