artelnics/opennn/run-opennn-examples

Run the OpenNN example matrix

Exercise every target registered through opennn_example(...) in examples/CMakeLists.txt. Treat that file as the source of truth; do not keep a separate hard-coded target list in this skill.

Matrix

Use the current two-argument configuration API from opennn/core/configuration.h:

Do not pass a third inference-precision argument; it no longer exists. Do not add Auto rows to the comparison: Device::Auto and Type::Auto resolve to one of the explicit cells according to available hardware. CPU with BF16 or INT8 is invalid by design. CUDA BF16 and INT8 require compute capability 8.0 or newer.

Prepare

1. Read AGENTS.md, README.md, examples/README.md, examples/CMakeLists.txt

and the current Configuration declaration and implementation. The code overrides this document if the API changes again.

1. Record git status --short. Never overwrite or restore changes that were

already present.

1. Discover the example targets from CMake and inspect each entry point for

required arguments, model downloads, interactive input and an existing Configuration::instance().set(...) call. Resolve current default filenames, labels and sample counts from the entry point and catalog; do not reuse filenames or expected counts from an earlier dataset.

1. Use separate Release build directories outside the checkout for CPU and

CUDA. Configure with examples enabled and benchmarks/tests disabled unless the user requested otherwise.

1. Confirm required datasets and model assets before starting a long cell.

Missing data or model assets is BLOCKED, not a code failure.

Apply a matrix cell

Set the configuration before constructing a dataset, network, model or session. For an example that already calls Configuration::set, temporarily replace that complete call. For an example without a call, temporarily add the direct opennn/core/configuration.h include and one call at the start of its try block. The blank target is an empty template that only prints two lines; do not edit it, and report it as N/A in every cell.

Before editing, save the exact original contents outside the repository. Restore them in a finally-style cleanup path after every example, including build failure, timeout or interruption. Verify restoration by comparing the file with the saved original; do not use git checkout, git restore or another command that could discard the user's changes.

Build only the target being exercised. Run it from the build's runtime-output directory so paths such as ../data/<example> resolve to the data copied or unpacked by CMake. Use the current bundled data and check the loaded sample count and classes against the catalog when the example prints them. Supply deterministic, minimal input to interactive programs; for Qwen, pipe a prompt followed by exit and use a user-provided/prepared model directory. Apply a finite timeout suited to the workload.

Use supported command-line flags for focused checks. emotion_analysis accepts --device cpu|cuda|auto, --seed N, --epochs E and --data FILE; verify these against its current entry point. --epochs 1 checks loading, training and evaluation quickly, but report it as a one-epoch smoke run. Its CLI selects FP32; the CUDA BF16 and INT8 cells still require the temporary precision edit above.

electrocardiogram_anomaly_detection takes no arguments and trains for 20 epochs. Use its bundled waveform CSV and testing indices from the current entry point. A successful run prints the reconstruction threshold, a confusion matrix, binary classification metrics and ROC AUC, then exits with code zero.

melanoma_cancer takes no arguments and loads 102 RGB images of 300 x 300 from ../data/melanoma_cancer/isic2016, in the benign and malignant class folders. Check that the loader sees 102 samples and two classes, then that training and binary evaluation finish with finite metrics. For a focused smoke run, a temporary set_maximum_epochs(1) on its optimizer limits training; save and restore the entry point as above, and report the one-epoch limit.

translation takes no arguments, loads 1,000 Spanish-English pairs from ../data/translation/tatoeba_es_en.tsv, and trains for 50 epochs in batches of 16 before translating Vete.. Autoregressive generation requires CUDA; a CPU check limited to loading or training does not pass the complete example. For a focused CUDA smoke run, temporarily limit the optimizer to one epoch, restore the entry point afterwards, and report that limit. Require finite training results, a non-empty generated translation and exit code zero.

bert accepts optional review-file and model-cache arguments, defaulting to ../data/bert/amazon_cells_labelled.txt and ../data/bert. It loads 1,000 reviews with Bad/Good labels and a WordPiece vocabulary of 30,522 tokens, uses sequences of length 64, and trains for three epochs in batches of 32. The first run downloads about 437 MB from Artelnics/bert-base-uncased-opennn on Hugging Face; record download time separately from execution with a populated cache. Require finite training and binary classification metrics, the final Good bye! message and exit code zero. For a focused smoke run, temporarily set one epoch, restore the entry point afterwards, and report that limit.

gpt2 accepts one optional quoted prompt argument, defaults to Artificial intelligence, and generates at most 40 new tokens. Its model cache is ../data/gpt2; the first run downloads about 650 MB from Artelnics/gpt2-small-opennn on Hugging Face. Autoregressive generation requires CUDA. Require a non-empty continuation after the GPT-2 generated text: label and exit code zero; record the prompt and any download time separately from a run using the populated cache.

qwen3 accepts one optional model-directory argument and selects the 4B model. Its default cache is ../data/qwen3; pass an existing cache explicitly to reuse it. The initial BF16 weight download is about 8.83 GB, plus tokenizer files. Downloading a 0.6B cache does not change the example's selected model. Pipe a short prompt such as Say hello in one short sentence. /no_think followed by exit. Require a non-empty Response: and exit code zero; opening the chat and immediately exiting checks loading only. Record the actual resolved device/precision and separate download time from execution. The CUDA binary loader explicitly requires BF16 or INT8; report a rejected CUDA FP32 cell as UNSUPPORTED, retaining the diagnostic. Do not substitute a BF16 run for a CUDA FP32 result.

yolo [experiment] currently selects v3-pretrained when called without arguments. It needs an external synthetic dataset at SYNTHETIC_YOLO_IMAGES and SYNTHETIC_YOLO_LABELS, falling back to synthetic_yolo/images and synthetic_yolo/labels. It expects 416 x 416 images and detection labels; record the actual loaded sample count and classes. The internal coloured-block generator is not called by this entry point. A missing dataset is BLOCKED. v3-scratch skips pretrained-backbone loading, but an existing experiment checkpoint can still be resumed; run in a new output directory when checking training from scratch. A missing pretrained file causes scratch initialization, so record that fallback instead of counting it as a pretrained run.

An experiment name containing voc selects VOC_ROOT and optional VOC12_ROOT; one containing bccd selects BCCD_IMAGES and BCCD_LABELS. The use_coco and use_raccoon source switches are false by default. v8-pretrained can import a supplied yolov8s.onnx; v8-scratch skips the pretrained import. There is no --epochs CLI option. For a focused training check, temporarily shorten the actual lr_schedule and any applicable backbone-freeze warmup, restore the source afterwards, and report the limits. Do not rely on the quick_test comment to cap training: inspect its uses. Require finite training results, saved checkpoints, completed detection/mAP output and exit code zero; report loading-only or shortened runs separately.

Classify results

Every target must have a result for every matrix column:

• PASS: exit code zero and the expected final output or metric is present.
• FAIL: it built and ran with its prerequisites, but compilation, execution,

numeric validity or output validation failed.

• UNSUPPORTED: the code explicitly rejects that device/type combination.
• BLOCKED: required hardware, data, model assets or external tooling is

unavailable.

• N/A: the target does not instantiate OpenNN. Currently only blank; every

other example target requires a result for each requested cell.

Do not silently convert failures to unsupported results. Record the command, exit code, elapsed time and a short diagnostic for every non-pass cell. Capture the example's meaningful final metric when it has one, and reject NaN/Inf or an unexpectedly empty result.

Finish

Restore every temporarily edited source and verify that git status --short differs from the initial snapshot only by artifacts explicitly requested by the user. Delete temporary Python runners created for the task after using them. Report a table with one row per CMake example and these columns:

| Example | CPU FP32 | CUDA FP32 | CUDA BF16 | CUDA INT8 |

Follow it with prerequisites, failures, unsupported paths, per-example elapsed time, run arguments and the exact build configuration. A partial matrix is useful, but never describe it as complete.