Skip to content

test_report.md

Latest commit

 

History

History
88 lines (66 loc) · 3.63 KB

test_report.md

File metadata and controls

88 lines (66 loc) · 3.63 KB
Raw

Radeon Cloud release-candidate test report

Environment

Item Verified value
GPU AMD Radeon Graphics
VRAM 47.98 GiB
PyTorch 2.13.0+rocm7.2
HIP 7.2.53211
torch.cuda.is_available() True
VLM Qwen2.5-VL-3B-Instruct, official local checkpoint
Person segmentation torchvision Mask R-CNN, local checkpoint

No CUDA PyTorch package was installed. Core inference, segmentation, planning, criticism, and card rendering ran on AMD Radeon Cloud. No external AI API or image-generation model was used.

Automated tests

Command:

/workspace/venv/bin/python -m unittest discover -s tests -p "test_*.py" -v

Result for the final public non-generative product path on Radeon Cloud: 71/71 passed.

The final demo also shows 76/76 for the complete development workspace. That workspace still contained five tests for retired generation experiments. Those tests and modules are intentionally omitted from this public release, whose product flow does not generate replacement images.

The suite covers the fixed Planner schema, Critic contract, image orientation, physical camera-geometry calibration, malformed Qwen JSON repair, permission policy, HTTP privacy behavior, person-mask matching, deterministic rendering, and single/group/no-prop guidance.

Real-GPU regression matrix

Each scenario ran three times through the local HTTP product path.

Scenario Runs Critic pass People detected / matched Mean agent time Peak VRAM
Single portrait with phone 3 3/3 1 / 1 113.401 s including first-run warm-up; warm runs ≈52 s 9.567 GiB
Single portrait, no handheld prop 3 3/3 1 / 1 116.459 s 10.684 GiB
Three-person group 3 3/3 3 / 3 226.018 s 10.677 GiB

The three-person case intentionally reports the slower path: malformed-comment repair, local grounding, and a second visual-field completion pass. Results were not replaced with synthetic estimates.

Release fixes proven by the matrix

  • A no-prop image now renders Both hands with a concrete action instead of an empty or invented object label.
  • Clothing, ties, belts, shoes, and watches remain visible evidence but are not presented as interactive photography props.
  • Qwen responses containing trailing commas, one missing property comma, or // line comments are handled by narrow deterministic repairs.
  • Model-output parse failures are not misreported as corrupt user images.
  • Full-body low-angle instructions require compatible geometric evidence. A half-body waist + level label no longer forces an incorrect low camera.
  • All final cards use the real local person mask and deterministic PIL drawing.

Evidence files

  • submission/evidence/SAMPLE_SOURCES.md — licenses and source URLs
  • submission/evidence/RELEASE_CANDIDATE_REPORT.md — historical nine-run summary
  • submission/final/evidence/*.json — fresh final English results
  • submission/final/evidence/*.png — fresh final English cards

The final English files are complete Agent results produced after the final copy and calibration fixes.

Fresh final English UI cases

These three complete Agent paths were rerun after the final English UI and camera-geometry calibration changes:

Scenario End-to-end Agent time Peak allocated VRAM Notes
Single portrait with phone 82.376 s 7.558 GiB One bounded JSON repair
Single portrait without a prop 57.867 s 10.403 GiB No invented handheld object
Three-person graduation group 59.569 s 10.385 GiB Three foreground people

The corresponding JSON and cards are under submission/final/evidence/.