ae9is / ezsam Goto Github PK

Create binary executable of gui (for ex. via pyinstaller)

GitHub Actions workflow for building binaries via Nuitka works for nuitka-test-simple but fails for nuitka-test-advanced and ezsam/gui/app.py throwing: ModuleNotFoundError: No module named 'ezsam'

Safety tool isn't catching vulnerabilities and currently its dependency chain introduces a couple vulnerabilities by pinning the cryptography library to an old version. Easiest to just remove it.

Dependabot only scans requirements.txt for vulnerabilities, not pdm.lock files.

Need to create pre-commit hooks for /docs project so that docs/requirements.txt is kept in sync.

Need to add CNAME file to /docs/CNAME for MkDocs so custom domain config is persistent.

Could possibly add support for OWL-ViT v2 (as a replacement to GroundingDINO) for object detections, which allows adding an image to prompting (1-shot prompting).

See colab of interactive segmenting here: https://github.com/google-research/scenic/blob/main/scenic/projects/owl_vit/README.md#inference-playground

GroundingDINO groundingdino.util.inference.Model() constructor requires a path to a config file, and dev installs of the app come with the config file at the default location (config/GroundingDINO_SwinT_OGC.py). However, wheel installs don't ship with the config and this requires the end user to grab the config file themselves, which isn't a great new user experience.

Could rework app to automatically create and pass a temporary config file to Model()

With the decision to bundle ezsam-gui with ezsam (cli), might as well restructure the monorepo to something simpler like pdm project's format: https://github.com/pdm-project/pdm

This means eliminating the packages/ directory and moving packages/app into the project root.

Make workflow binary build works fine, but some other issues remain with the make workflow itself (caching, size limitations)

ref: #42 #43

Documentation site for github pages

Could rework video file input code to support GIF input, or at least document. Current state is unexpected since GIF output is supported.

Make workflow's actions/upload-artifact fails to zip and upload standalone executable folder due to number of objects

Need to configure model correctly when torch is in CPU only mode

throws

Torch not compiled with CUDA enabled

Add support for processing video files directly

Setup a release.yml to build and publish tags on demand to PyPI

Updating dependencies in #25 (specifically 3a6f3a7) broke debug mode with the update of supervision-0.17.1 to supervision-0.18.0 in pdm.lock.

There was a breaking change in the library here:
roboflow/supervision@a10b984

Affected line is:

Could check whether ffmpeg and convert (ImageMagick's convert tool) are installed, when required for video output modes, and terminate early with descriptive user error messages

Add ability to inverse prompt selection, and/or add negative prompting at the same time as positive prompting

Could create GUI and binary executable via pyinstaller

Sometimes jobs can occasionally fail even when previously run successfully on the same image with the same GPU, throwing an exception like Error processing file examples/car-1.jpg: CUDA out of memory. Tried to allocate...

This is a PyTorch issue, but could explicitly delete objects referencing PyTorch and try to get PyTorch to free up allocated GPU memory via torch.cuda.empty_cache()

ref: https://github.com/ae9is/ezsam/tree/main/packages/cli#gpu-memory-troubleshooting

Could move ./models and ./config folders to a new cache folder ~/.cache/ezsam

Test pdm builds and publish to PyPI

When prompts fail to pass threshold to match any objects, the error message (`Error processing file [input]: list indices must be integers or slices, not NoneType) is not intuitive right now. Need to handle exception and fail gracefully