Fix the crashes in Prediction.save

[iurisilvio]

What does this PR do?

Prediction.save() draws a single prediction on its image and writes it to a file. It raises before writing anything for object-detection, keypoint, classification and semantic-segmentation predictions, and for hosted https:// images. Commit db50306 ("fixed save method + predict for hosted images") fixed most of these in PredictionGroup.save(), but they were never applied to Prediction.save(). This PR applies them:

• Object detection and keypoint (roboflow/util/prediction.py): the label background's corners were floats (x - width / 2, ...). cv2.rectangle rejects float points (cv2.error: ... Can't parse 'pt1'. Sequence item with index 0 has a wrong type) on every supported OpenCV: checked on 4.10, the minimum in requirements.txt, and on 5.0. The corners are now ints. The label text now sits on that background at the top of the box, the same position PredictionGroup.save() uses; before, it was drawn at the box's vertical center.
• Classification: "Confidence: " + self["confidence"] raised TypeError on the float confidence the API returns. It is now str(self["confidence"]), as in PredictionGroup.save().
• Hosted images: Prediction.__load_image downloaded only http:// URLs. An https:// image went to cv2.imread, came back empty, and failed in cv2.imwrite (!_img.empty()). It now downloads both, as PredictionGroup.__load_image does.
• Semantic segmentation (roboflow/util/image_utils.py): mask_image decoded the mask with np.fromstring. NumPy 2.3 removed its binary mode (ValueError: The binary mode of fromstring is removed, use frombuffer instead). It now uses np.frombuffer, which also fixes PredictionGroup.save() for semantic segmentation.

Instance segmentation already worked and is unchanged.

Related Issue(s): none.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:

• The new tests/util/test_prediction.py saves a Prediction for tests/images/rabbit.JPG to a PNG and reads it back. It covers object detection, keypoint, classification, semantic segmentation, and a hosted https:// image (with the download mocked). The object-detection test also checks that the label text is drawn on its background. Each test fails on main; the semantic-segmentation one fails only on NumPy 2.3 or later.
• python -m unittest gives 1238 tests OK, 1 skipped, on Python 3.12 (NumPy 2.5.4, OpenCV 5.0.0) and on Python 3.10 (NumPy 2.2.6, OpenCV 5.0.0). make check_code_quality passes. cv2.imdecode also decodes the read-only array from np.frombuffer on OpenCV 4.10 with NumPy 1.26.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable): not applicable

Additional Context

• Send a roboflow-python User-Agent on every SDK request #559 (User-Agent) changes the urlopen call right below the https:// check in Prediction.__load_image, so whichever of the two merges second gets a small conflict there. Keep both changes.

🤖 Generated with Claude Code