From labeled drawings to a cloud endpoint.

Azure references: Compute clusters · Command jobs

hvac-symbols-v1/
├── data.yaml
├── split-manifest.csv
├── images/
│   ├── train/  drawing01_tile001.png
│   ├── val/    drawing08_tile001.png
│   └── test/   drawing12_tile001.png
└── labels/
    ├── train/  drawing01_tile001.txt
    ├── val/    drawing08_tile001.txt
    └── test/   drawing12_tile001.txt

path: /data/hvac-symbols-v1
train: images/train
val: images/val
test: images/test

names:
  0: supply_diffuser
  1: return_grille
  2: vav_box

0 0.500000 0.250000 0.125000 0.125000

• Annotate every target symbol; keep reviewed background tiles with empty label files.
• Start with 1,024 px tiles and 256 px overlap (768 px stride). Remap boxes into tile coordinates and use a consistent clipping/filtering policy for partial objects.
• Store the source image, project ID, tile offset, and split in the manifest so full-sheet predictions can be reconstructed.
• The app’s 80 reference crops are a symbol library, not a labeled detection dataset. Add real drawings, boxes, and representative style variations.
• Leader paths, OCR text, and label-to-symbol links need separate annotations and evaluation; YOLO box labels do not encode those relationships.

Official YOLO detection dataset format

1. Prepare and review labeled data

   Draw boxes, agree on class names, split by project, and generate tiles. Check missing labels, invalid coordinates, duplicates, and class counts. Upload and version the data asset.

   OUTPUT · versioned dataset + manifest

2. Define the Azure pipeline

   Create Prepare, Train, and Evaluate/Report command components using Azure ML SDK v2 or CLI v2. Declare data/model folders as inputs and outputs, connect them, and assign CPU or GPU compute to each step. Log parameters and metrics to MLflow from the scripts.

   OUTPUT · reusable pipeline definition + versioned environment

3. Set the first experiment’s hyperparameters

   Begin with the settings below. Save the dataset version, seed, package versions, and all arguments with the run. Keep the split fixed while comparing models.

   Parameter	Starting value	How to adjust
   Pretrained weights	yolo26s.pt	Try Medium after establishing a Small baseline.
   Image size	1024	Increase only if small details remain unresolved and memory allows.
   Batch size	-1 (auto)	Log the resolved size; reduce it if memory is exhausted.
   Epochs / patience	100 / 20	Maximum epochs / early stopping patience. Inspect validation trends.
   Optimizer / learning rate	Baseline: optimizer=auto	Record resolved settings. To test a specific learning rate, select an explicit optimizer; auto may override manual learning-rate settings.
   Augmentation	Review rotations, flips, scaling	Use only transformations that preserve symbol meaning; disable invalid defaults.
   Compute	One GPU; 24 GB VRAM target	Verify actual SKU memory and availability. This is not a measured requirement.

4. Run a 2–5 epoch smoke test

   Confirm CUDA is active, labels load correctly, checkpoints reach persistent storage, and MLflow receives metrics. Measure seconds per epoch and peak memory before estimating the full run’s duration and cost.

   OUTPUT · verified environment + runtime estimate

5. Train the full baseline on the GPU

   Submit the job and inspect Azure ML Studio → Jobs → the run. Log training losses and validation metrics each epoch. Save best.pt, last.pt, configuration, and plots. Use the best checkpoint rather than assuming the final epoch is best.

   OUTPUT · checkpoints + learning curves

6. Compare fine-tuned candidates

   Change one major setting per experiment, such as Small versus Medium or image resolution. Compare full-sheet validation results after tile merging. Choose confidence thresholds on validation data and inspect missed rare symbols, not only the overall average.

   OUTPUT · chosen run + frozen inference settings

7. Test once and register the selected model

   Run the selected configuration on untouched test projects. Save the final report, failure examples, and latency measurements. Register the weights with class mapping, preprocessing, tile overlap, merge settings, thresholds, and environment version.

   OUTPUT · final test report + registered model version

Training options · Azure ML experiment tracking

Illustrative validation comparison Invented numbers · not project results

Example decision rule: require at least 95% precision, then prefer higher recall within an agreed latency budget. Tune each run’s threshold on validation data only.

Run B meets the example precision target and finds more symbols than A. Run C’s higher recall does not satisfy the precision target at the shown threshold; inspect its precision–recall curve before rejecting the model entirely.

What to publish in the cloud report

• Azure ML metrics: validation loss, precision, recall, mAP, dataset version, and epoch number.
• Downloadable artifacts: precision–recall curves, confusion matrix, per-class CSV, and annotated missed/false detections.
• Final full-sheet report: deduplicated precision/recall, count errors, class support, and latency on fixed test projects.
• Separate semantic report later: OCR accuracy and label-to-symbol association accuracy. Detector metrics do not measure leader understanding.

YOLO validation · Metric interpretation

1. Register the model package. Include weights, class names, preprocessing and inference settings, and the environment version.
2. Write a scoring adapter. In score.py, load the model once in init(). In run(), validate the image input, tile it, infer, merge duplicates, and return class names, scores, and full-image boxes. This adapter is future implementation, not included in this guide.
3. Create a managed online endpoint and deployment. Configure authentication, the registered model, scoring script, environment, and serving instance type/count. Training’s 24 GB recommendation does not determine serving requirements.
4. Test before routing traffic. Check health, response schema, known images, latency, invalid inputs, and memory. Assign traffic to the tested deployment; retain a previous version for rollback.
5. Connect through an application backend. The backend authenticates to Azure; do not embed endpoint credentials in browser JavaScript. Monitor errors, latency, and new drawing styles.

Deploy and test a managed online endpoint