Skip to content

Releases: arm/neural-graphics-model-gym

v0.2.0

28 Nov 14:56
@Burton2000 Burton2000
v0.2.0
8c980ee
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Choose a tag to compare

View all tags
v0.2.0 Latest
Latest

New Features

Bring Your Own Model (BYOM) Support

You can now register custom neural graphics models and datasets that plug into the ng-model-gym’s training, QAT, evaluation and export flows.

Added

  • Abstract base classes: BaseNGModel and BaseNGModelWrapper for creating neural graphics models
  • Model and dataset registries with decorators for auto-discovery and loading during ng-model-gym startup
  • Ability to export custom metadata alongside the VGF file
  • Configurable schedulers, optimizers, metrics, and losses specified in the config file
  • Safetensors cropper script
  • Best performing checkpoint during training is now highlighted and saved

Changed

  • Refactored repository into /core (shared logic) and /usecase folders
  • Dependency updates:
    • ExecuTorch 1.0.0
    • PyTorch 2.9.0
    • TorchAO 0.14.0
  • BaseModelEvaluator renamed to NGModelEvaluator
  • User-defined static and dynamic model export shape instead of using dataset shape
  • Dockerfile base now uses: nvidia/cuda:12.8.0-devel-ubuntu22.04
  • History buffers moved from the NSSModel to FeedbackModel
  • FeedbackModel attribute renamed from nss_model → ng_model
    • Pre v0.2.0 NSS checkpoints can still be loaded.

Breaking Changes

  • Configuration file schema changed. To generate an updated config file run: 
    ng-model-gym init

Removed

  • The bundled model-converter binary is no longer included.
  • Wheels now depend on the ai-ml-sdk-model-converter package from PyPI instead.

Documentation

  • Added a CONTRIBUTING guide
  • General updates and improvements to the README
  • Added data capture documentation
Assets 2
Loading

v0.1.0

08 Aug 15:12
@Burton2000 Burton2000
v0.1.0
ad37e03

Choose a tag to compare

View all tags
v0.1.0

Features

  • Neural Super Sampling (NSS)

    • A trainable upscaling model for real-time graphics on mobile devices with Neural Accelerators.

  • Training & Evaluation

    • FP32 and Quantization-Aware Training (QAT INT8) modes

    • Train from scratch or finetune from pre-trained weights

    • Model quality evaluation across a range of metrics

  • Export to VGF

    • Uses ExecuTorch with the Arm backend to export models to VGF file

  • CLI and Python API

    • Choose between the ng-model-gym command-line interface or import the Python package in your own code

  • Profiling & Visualization

    • PyTorch profiler instrumentation

    • TensorBoard integration for monitoring training 

Dependencies & Sources

Note: We rely on a nightly build of ExecuTorch and non-PyPI wheels for PyTorch and TorchVision


Package Version / Channel Source URL
Python 3.10, 3.11, 3.12 —
torch 2.7.1+cu118 https://download.pytorch.org/whl/cu118/torch-2.7.1%2Bcu118-cp311-cp311-manylinux_2_28_x86_64.whl
torchvision 0.22.1+cu118 https://download.pytorch.org/whl/cu118/torchvision-0.22.1%2Bcu118-cp311-cp311-manylinux_2_28_x86_64.whl
executorch 0.8.0.dev20250702+cpu (nightly, unstable) https://download.pytorch.org/whl/nightly/cpu/executorch-0.8.0.dev20250702%2Bcpu-cp311-cp311-manylinux_2_28_x86_64.whl
tosa_serialization_lib commit 6454bc0f (v1.0-branch) git+https://git.gitlab.arm.com/tosa/tosa-serialization.git@6454bc0fhttps://git.gitlab.arm.com/tosa/tosa-serialization/-/tree/6454bc0fef8404a58cbfc2eaa6bcad4b17910795
Loading