ImagePropertyDrift compares image-property distributions between reference and current vision datasets, helping detect production changes that may affect an object-detection system. section 11.2.1
Apply it as an operational monitoring workflow:
- Load training and current data as deepchecks VisionData objects. section 11.2.1
- Run ImagePropertyDrift on the reference and current datasets. section 11.2.1
- Save the HTML report and inspect raw drift values and property distributions. section 11.2.1
- Store inference images and predicted labels, then calculate drift periodically against training data. section 11.2.1
Brightness-shifted ID-card images
A training dataset contains unmodified images, while the current dataset contains brightness-adjusted images. Brightness shows substantial drift (0.619), RMS Contrast also shifts (0.310), and Area and Aspect Ratio remain unchanged at 0.0. section 11.2.1
Listing 11.6 Running deepchecks ImagePropertyDrift
from Dataset import load_dataset
from deepchecks.vision.checks import ImagePropertyDrift
train_dataset = load_dataset(train=True, object_type="VisionData")
test_dataset = load_dataset(train=False, object_type="VisionData")
check_result = ImagePropertyDrift().run( #1
train_dataset, test_dataset #1
) #1
check_result.save_as_html( #2
"deepcheck_vision_drift_check.html" #2
) #2
print(check_result.value) #3
Use the runnable deepchecks listing to compare datasets, save the report, and print raw property-level results. section 11.2.1
