Skip to content

Release 0.5.0 — review hardening, progress bar, crash fix, export refinements - #14

Merged
kouya-marino merged 5 commits into
mainfrom
release/0.5.0
May 31, 2026
Merged

kouya-marino merged 5 commits into
mainfrom
release/0.5.0

Conversation

@kouya-marino

Copy link
Copy Markdown
Owner

Code-review hardening pass + small features, released as 0.5.0. Five focused commits:

  1. Deduplicate workers — SpawnProcess base (train/export, with terminate()), ImageInferenceWorker base (inference/deploy); workers emit finished_all in a finally; removed dead run_inference_single.
  2. GUI dedup + features — ResultsViewerMixin shared by Inference/Deploy tabs; training progress bar; graceful shutdown (closeEvent stops processes / awaits workers); crash-safe worker teardown (wait() before release).
  3. Core hardening — TensorRT keeps the intermediate ONNX fp32 (fp16 only at engine build) + records engine settings; build_cfg no longer registers datasets in the GUI process; stronger validate_coco_json; session-log retention.
  4. Entry/docs — main.py→app.main (single freeze_support); .pth/torch.load security note; RetinaNet TensorRT de-risk script.
  5. Bump to 0.5.0 + CHANGELOG.

Notable fixes

  • App crash QThread: Destroyed while thread is still running after a Deploy/Inference run.
  • App close no longer orphans the training GPU process.
  • fp16 ONNX export is fail-safe; TensorRT path uses fp32 ONNX.

Verification

Full suite green locally (60 passed, 2 skipped), ruff + format clean, MainWindow constructs + closes cleanly. Tag v0.5.0 will publish to PyPI after merge.

prashantrawat11 added 5 commits May 31, 2026 17:47
Add SpawnProcess (train/export spawn managers) with a terminate() for shutdown, and ImageInferenceWorker (inference/deploy QThread workers). Workers now emit finished_all in a finally so the UI always re-enables. Removed the now-unused run_inference_single from inferencer.py.
Extract ResultsViewerMixin (shared image viewer + detections table + nav) used by the Inference and Deploy tabs. Add a training progress bar to the Train tab. MainWindow.closeEvent now shuts down each tab's worker/process cleanly; tabs wait() for worker threads before releasing them.
TensorRT export keeps the intermediate ONNX fp32 (fp16 only at engine build) and records engine fp16/workspace in metadata; export metadata records the baked score threshold. build_cfg no longer registers datasets in the GUI process. validate_coco_json checks category id/name, category-id references, and bbox shape (bounded scan). Session logs are capped to the newest 20.
main.py now delegates to app.main; freeze_support() lives in app.main() (covers both entry points). README documents the torch.load (.pth) security caveat. test_scripts/retinanet_trt_derisk.py checks whether a RetinaNet ONNX is TensorRT-convertible.
Release 0.5.0: training progress bar, graceful shutdown, QThread crash fix, TensorRT/ONNX export hardening, stronger COCO validation, worker/tab deduplication, and the onnxruntime-gpu + RoiAlign TensorRT docs.
@kouya-marino
kouya-marino merged commit 3e1de1a into main May 31, 2026
4 checks passed
@kouya-marino
kouya-marino deleted the release/0.5.0 branch May 31, 2026 08:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant