Skip to main content

Inference

trolo provides both a CLI interface and Python API for inference.

CLI Usage​

Run inference on different inputs:

Basic inference with automatic model download​

trolo predict --model dfine-n

Inference on specific input​

trolo predict --model dfine-n.pth --input img.jpg

Support for multiple input types​

trolo predict --model dfine-n.pth --input folder/ # image folder
trolo predict --model dfine-n.pth --input video.mp4 # video file
trolo predict --model dfine-n.pth --input 0 # webcam

Smart Video Stream Inference​

trolo implements streaming inference for videos to handle memory efficiently. This means you can process large videos without worrying about memory constraints.

Python API​

from trolo.inference import DetectionPredictor

# Initialize predictor
predictor = DetectionPredictor(model="dfine-n")

# Get raw predictions
predictions = predictor.predict()

# Visualize results
plotted_preds = predictor.visualize(show=True, save=True)

Prediction results​

The predict method returns a list of dictionaries, where each dictionary contains the prediction results for an image.

# Example prediction
preds = predictor.predict()

for pred in preds:
    boxes = pred['boxes']
    scores = pred['scores']
    labels = pred['labels']
    # ... #

Useful model information like class names​

Models trained with trolo are saved with useful information like class names. These can be accessed by:

class_names = predictor.config.yaml_cfg['class_names']

You can then implement your own custom visualization function using these class names.

# Example
predictor = DetectionPredictor(model="dfine-n")

preds = predictor.predict()
class_names = predictor.config.yaml_cfg['class_names']

for pred in preds:
    boxes = pred['boxes']
    scores = pred['scores']
    labels = pred['labels']
    # ... #
    for box, score, label in zip(boxes, scores, labels):
        print(f"Label: {class_names[label]}, Score: {score:.2f}")