NVIDIA / NVIDIA/apex

Dose data_prefetcher() really speed up training?

Open
#304 21 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
9k
Forks
1.5k
Avg merge
2d 4h
Merged PRs (30d)
3

Description

I used your Python code https://github.com/NVIDIA/apex/blob/master/examples/imagenet/main_amp.py#L256

My code is

I think this data_prefetcher could speed up training, because there is another stream sending data to GPU memory when model is running in the GPU. So there is a very small gap between two iteration.

However, this trick does not work for me. So please help me

  • data_prefetcher is really for speed up?
  • why it dosen't work for my case

Contributor guide

No contributing guide indexed for this repository

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 with examples/imagenet/main_amp.py around the referenced data_prefetcher implementation, then compare it with maskrcnn_benchmark/engine/trainer.py and tools/train_net.py using DistributedDataParallel across 8 GPUs. Measure iteration timing with and without prefetching; done would be a documented explanation of whether it improves this setup and why the observed behavior differs.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, 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.