[Accuracy] MoGe selected convolutions replace Native replicate padding with zeros

Cerrado
  • ¿Algo no está bien?
#1,727 0 comentarios 0 reacciones 1 asignado Ver en GitHub

Los mantenedores suelen responder en 1 día

@chaofengw-nv ya está trabajando en esto.

Desde el 15/10/2026.

  • #1734 de @chaofengw-nv — cerrado sin fusionar
  • #1822 de @chaofengw-nv — abierto

Evaluación

Este issue todavía no se ha evaluado.

Descripción

bug

Affected area

Inference output or task execution

TensorRT-Model-Connect version

Diagnostic snapshot 174df6571e2152ad95edeb1a5c1c30b1257ecb6f; affected family production files also present at cef761444d2525ad00a86cfe0413157801d175ab.

Installation method

Development container / source build.

Model and revision

Ruicheng/moge-2-vitl @ 39c4d5e957afe587e04eec59dc2bcc3be5ecd968

Environment

  • NVIDIA GB300, CUDA 13.0, TensorRT 11.2.0.113 for the prior model controls.
  • CPU regression uses the installed TensorRT Python API without CUDA inference when applicable.

Minimal reproducer

A hardcoded set of 24 neck/points/mask convolutions skips the official replicate-padding operation and substitutes convolution zero padding. These are different boundary conditions and change valid border predictions.

With an AIPerf environment configured as described in apps/aiperf_qual/README.md, run the family profile before and after the production change:

trtmc-aiperf-qual run --profile moge-2-vitl --environment "$E" --out out/moge-2-vitl

The exact prior replay outcomes are listed below. This submission does not claim that this public CLI command was rerun on the final PR head.

For local CPU contract coverage (not a model-parity reproducer):

python -m pytest -q families/moge/tests/test_selective_zero_padding.py

Observed behavior

HUMAN REVIEW REQUIRED: the three existing tests explicitly demand the conflicting zero-padding behavior and still fail after this correction. No tests, gates or thresholds were changed.

Earlier GB300 full 50-image geometry control: restoring replicate padding improved 47/50 -> 49/50 with original mask IoU/depth criteria. One depth-error case remains above the original 1% limit.

Expected behavior

Remove the selected zero-padding override and retain each convolution caller's replicate-padding contract.

Relevant logs

Specific affected source: families/moge/model.py. No private logs or model assets are attached.

Impact and scope

Restore Native padding only; do not relax the 1% depth criterion or claim complete MoGe parity.

Draft only: existing test expectations require a maintainer decision; no passing-CI claim. Remaining depth error and performance are not fixed.

Submission checks

  • Searched open and closed issues for this root cause; no duplicate found.
  • This is not a security vulnerability.
  • Removed credentials, private URLs, internal paths and restricted artifacts.
Lenguaje dominante
Python
Estrellas
283
Forks
75
Merge medio
1 d 17 h
PR fusionados (30 d)
218

Preparar el entorno

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de NVIDIA/TensorRT-Model-Connect

Todos los issues de NVIDIA/TensorRT-Model-Connect

Issues similares

Más issues de Python

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.