Environment and FAQ¶

What to keep in the complete package¶

Distribute the complete application folder, including the executable, _internal, classic runtimes and anomaly runtime. Customers need working GPU drivers and appropriate system runtime libraries; bundled framework libraries do not replace NVIDIA drivers.

MkDocs is used only to build documentation during development. Customers do not need MkDocs, Python or a separate web server. The application starts a lightweight service bound only to the local machine; all five language editions and search resources are bundled.

Bundled help lives in _internal/help_site; keep it for the Help action. The root docs technical notes and package-manifest.json are still not training/inference dependencies. The manifest supports file-integrity diagnosis.

Why mAP is missing¶

First check for an independent validation set, then whether an evaluation epoch has been reached. Classic COCO training without validation does not split images or evaluate. If evaluation ran but yielded no valid metrics, inspect its raw logs; loss is not an evaluation substitute.

Low GPU utilization¶

GPU utilization depends on the task, batch size, image size, data loading, CPU preprocessing and sampling interval. GPU utilization and GPU memory usage are different measurements. Check that the logs show GPU execution and whether data_cost is large; for classic Windows training, adjust workers moderately and compare actual throughput.

Insufficient RAM or GPU memory¶

Reduce batch first, then image size as appropriate for the task. For anomaly detection, also reduce candidate-feature limits, coreset limits or the retention ratio. Parameters differ across tasks: a coreset limit is not a batch size.

Classic models are still downloading pretrained weights¶

Using pretrained settings from the configuration follows its URL; a customer machine may have no cached copy. Choose local compatible weights or random initialization offline. Having architecture files does not mean pretrained weights are cached.

Mismatched weights, classes or preprocessing¶

Use the same task, configuration, class order and preprocessing as the weights. Classic models can import their experiment settings; modern classification must retain preprocessing metadata. Do not use a horizontal detector as OBB or treat an exported classic static model as a .pdparams checkpoint.

Why ONNX runs on CPU¶

The main inference page currently forces modern ONNX models to use CPU and reports this in the raw log. Selecting GPU does not override this existing compatibility policy.

Help pages or search cannot be opened¶

Keep the application running, reopen Help from the sidebar, and check the default browser and local permissions. Internet is unnecessary. If files are missing, use a complete new package containing help_site; adding documentation to source code does not update an old package automatically.