Skip to content

Easily benchmark training of any model by (op type+parameters) to spot real bottlenecks. Catches more details then native tfprof, especially for RNN/LSTM models

Notifications You must be signed in to change notification settings

MikalaiDrabovich/TensorScope

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

72 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TensorScope

Alt text Alt text Alt text

Find bottlenecks in any complex model within minutes, with exact parameters, even for dynamically instantiated graphs.

  • Produces interactive pie charts and allows easy navigation through complex hierarchy of ops.
  • Compared to Tensorflow timeline, aggregates across op type and tensor dimensions to allow targeted tuning of top-k kernels.
  • Chart generation is based on Krona tools: https://github.com/marbl/Krona/wiki/KronaTools
  • For examples of generated interactive pie charts (html5) see /results

Usage

  1. Simply ./run_me.sh in reproduce_results/
  2. See results in pie_chart.html and data.tsv ./results_summary/
  3. To add to your existing training setup - see module comment section in tensorscope.py

Contact

nick.drabovich@amd.com

Update 09/2018 The project is in the process of moving to currently private https://github.com/ROCmSoftwarePlatform/TensorScope Coming soon - bug fixes and changes are to be published here as well

About

Easily benchmark training of any model by (op type+parameters) to spot real bottlenecks. Catches more details then native tfprof, especially for RNN/LSTM models

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published