NVIDIA / NVIDIA/TensorRT

how to use polygraphy debug precision?

Open
#4,616 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Module:Polygraphy
Dominant language
C++
Stars
13.4k
Forks
2.4k
Avg merge
5d 3h
Merged PRs (30d)
2

Description

When using Polygraphy to debug mixed precision, it suggests which layers should be set to high precision (such as float32), but I cannot determine exactly which layers to configure or map them to the corresponding nodes in the ONNX model. How can I accurately identify the layers that need to be set to float32 to ensure precision is not compromised?

commond is : polygraphy debug precision model.onnx --mode bisect --fp16 --no-remove-intermediate --check polygraphy run polygraphy_debug.engine --trt --load-inputs net_input.json --load-outputs onnx_res.json --atol 1e-3 --rtol 1e-3

Final info :
[I] To achieve acceptable accuracy, try running the first 1301 layer(s) in higher precision
[I] Finished 12 iteration(s) | Passed: 5/12 | Pass Rate: 41.666666666666664%

tensorrt version is 8.5.1
Polygraphy version is 0.49

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 with the supplied polygraphy debug precision command and its bisect output. Read the Polygraphy precision-debug documentation and inspect how the reported layer count relates to the ONNX model nodes. Done means documenting a reliable way to identify the affected nodes and configure them for higher precision.

Written by the indexing model from the issue text.

Assessment

Domain
cli, machine-learning, performance
Issue type
Documentation
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.