Q and A

Common questions and what the messages mean. When something fails, look at the logs first.

Where are the logs?

~/.yoru/logs/yoru_explain.log (on Windows C:\Users\<you>\.yoru\logs\), next to YORU’s own log; the YORU_HOME environment variable moves both. Each run writes logs/run.log in its run folder, with tracebacks. Help → Open log folder and Settings → Open log folder open the folder. For a bug report, add the text from Settings → Versions, adapters and the loaded model.

Starting

python -m yoru_explain does nothing, or opens the Microsoft Store.

On some Windows machines python is the Store stub, not the environment’s Python. Use uv run python -m yoru_explain, or .venv\Scripts\python.exe -m yoru_explain.

uv sync cannot find YORU.

YORU-Explain installs YORU from the folder next to it, ../YORU. Check that YORU-dev/YORU and YORU-dev/YORU-Explain are siblings (Install).

A YOLOv5 model does not load.

YORU-Explain runs YOLOv5 models with YORU’s bundled YOLOv5 code, yoru/libs/yolov5, which not every YORU 2.0 version contains. Use a YORU checkout that has it, such as the yoru-tracker-lite branch.

Grad-CAM

Grad-CAM is not in the method list.

It is offered only when the loaded model passed the gradient probe; the reason is shown in amber under the methods on the Analyze tab. ONNX models and black-box models have no gradients, so they never offer it.

A Grad-CAM map is bright on another animal too.

With YOLOv5 at its default layer, the score’s gradient reaches exactly one cell of the feature map; the map then shows where the channels of that cell are active anywhere in the frame, on other animals of the same class as well. That is the measurement, not a failure (Where the gradient reaches).

Identity modes and statuses

fixed_candidate is not offered for my model.

RT-DETR’s queries and Faster / Mask R-CNN’s proposals are chosen anew for every input, so “the same candidate” on a changed frame does not exist. These models are followed by re-detection only (What is measured).

Many patches are disappeared, and the map has holes.

In re-detection mode a patch after which the detection is not found again has no score, and pixels covered only by such patches have no value. That is reported, not hidden: the counts are in Statistics, lost_fraction in the map file shows where. Fixed-candidate mode, where the model offers it, gives a score for every patch.

Runs that stop or fail

identity_check_failed: the run stopped at once.

On its first frame, the adapter’s detections were compared with YORU’s own detector and differed. Check that YORU-Explain’s environment uses the YORU next to it (uv sync), and report the case with the log: no result is written when the check fails.

frame_mismatch.

The decoded frame’s pixels differ from those the event was made from — the video file changed, or a different file now sits at its path. Verify files on the Project tab shows which.

target_not_reproduced.

The event’s detection was not found again on its frame with the loaded model (the same candidate and class with IoU ≥ 0.9, or the same class with IoU ≥ 0.9). Typical causes: another model or other thresholds than the event was made with, or events imported from a CSV of another model.

cuda_oom.

The GPU ran out of memory. Lower batch size on the Analyze tab; batch size 1 is also the exact one.

Seeking

“Seeking is not exact for this video”.

Check seeking found frames for which seeking lands on another frame than decoding in order (some files state a frame rate their frames do not have). Results are not affected: analysis runs always read frames in order and compare each frame’s hash, and Explore seeks only through a presentation-time index and verifies where each seek landed.

Other

Paths in the window show ¥ instead of \.

The window uses YORU’s Japanese-capable font, in which the backslash is drawn as a yen sign. The paths themselves are unchanged.

Results differ slightly between the GPU and the CPU.

CUDA runs with fp16 autocast, as YORU does, and the CPU in fp32, so scores differ in the last digits (e.g. 0.857316 and 0.857245). Every run records its device; compare runs made on the same one.