tensorflow / tensorflow/privacy

mnist_dpsgd_tutorial_vectorized Training slower on GPU

Open
#116 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
2k
Forks
477
Avg merge
22h 12m
Merged PRs (30d)
1

Description

I'm running the vectorized tutorial (tf-gpu 1.14) and the per epoch training time on GPU (Nvidia Tesla M10) is slower than the non-vectorized implementation:

GPU
Vectorized : 203.55 seconds/epoch
Non-vectorized : 115 seconds/epoch

I was expecting to see better runtimes on GPU, but I am not sure why it's slower. Any help would be greatly appreciated.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the vectorized and non-vectorized MNIST DPSGD tutorial runs with tf-gpu 1.14 on the reported Nvidia Tesla M10, using the per-epoch timings in the issue as the baseline. Compare the two tutorial implementations and their GPU execution behavior; done means explaining the runtime difference or identifying a reproducible performance fix.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, tensorflow
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.