facebookresearch/maskrcnn-benchmark

Half/mixed-precision is not faster

Open

#1,062 opened on Aug 27, 2019

View on GitHub
 (3 comments) (0 reactions) (0 assignees)Python (9,161 stars) (2,574 forks)batch import
help wanted

Description

❓ Questions and Help

As written in https://github.com/facebookresearch/maskrcnn-benchmark/issues/807#issuecomment-500112612 , DTYPE "float16" does not make training faster( and takes the same amount of memory) when tested with e2e_mask_rcnn_R_50_FPN_1x.yaml. Is this normal or am I doing something wrong? I also want to know if I can make inference faster by FP16(I tried model.half() in test_net, but doesn't work).
My environment is RTX Titan, cudnn 7.6, cuda10.0, driver version Driver Version: 418.67, pytorch-1.2.0, ubuntu 18.04.

Contributor guide

Tech stack
pythonpytorch
Domain
machine learningperformance
Issue type
performance
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
PyTorch basicsmixed precision concepts
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
The issue is a user question about why half precision is not faster in maskrcnn benchmark. To help, one should investigate the use of automatic mixed precision (AMP) in PyTorch, check if the model and training script are properly configured for FP16 (e.g., using torch.cuda.amp or apex), and verify that the GPU supports fast FP16 operations (e.g., RTX Titan). Additionally, check for any known issues in the repository related to mixed precision, such as the linked comment in issue #807. The user also tried model.half() for inference without success, so explore proper FP16 inference workflows.