app.py

import gradio as gr
from ultralytics import YOLO
import cv2
import tempfile
import time
import numpy as np

# Load a custom YOLO model from the uploaded file.
def load_model(model_file):
    try:
        model = YOLO(model_file.name)
        return model
    except Exception as e:
        return f"Error loading model: {e}"

# Run inference on an image and return a processed image as an np.ndarray.
def predict_image(model, image, conf):
    try:
        start_time = time.time()
        # Run inference with confidence threshold.
        results = model(image, conf=conf)
        process_time = time.time() - start_time

        # Get the annotated image using the model's built-in plotting.
        annotated_frame = results[0].plot()
        # Optional: Convert BGR (OpenCV default) to RGB if needed.
        annotated_frame = cv2.cvtColor(annotated_frame, cv2.COLOR_BGR2RGB)

        # Count detections if available (assumes results[0].boxes exists).
        num_detections = len(results[0].boxes) if hasattr(results[0], "boxes") else "N/A"
        return annotated_frame, process_time, num_detections
    except Exception as e:
        return f"Error during image inference: {e}", None, None

# Run inference on a video by processing selected frames and return a processed video file.
def predict_video(model, video_file, conf, frame_step):
    try:
        cap = cv2.VideoCapture(video_file.name)
        frames = []
        frame_count = 0
        start_time = time.time()

        while True:
            success, frame = cap.read()
            if not success:
                break

            # Only process every nth frame determined by frame_step.
            if frame_count % frame_step == 0:
                results = model(frame, conf=conf)
                annotated_frame = results[0].plot()
                frames.append(annotated_frame)
            else:
                # Optionally, append the original frame, or skip entirely.
                frames.append(frame)
            frame_count += 1

        process_time = time.time() - start_time
        cap.release()

        if not frames:
            return "Error: No frames processed", None, None

        height, width, _ = frames[0].shape
        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
        tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
        out = cv2.VideoWriter(tmp.name, fourcc, 20.0, (width, height))
        for frame in frames:
            out.write(frame)
        out.release()

        # For video, we return a placeholder for number of detections. (More logic can be added to aggregate detections.)
        num_detections = "See individual frames"
        return tmp.name, process_time, num_detections
    except Exception as e:
        return f"Error during video inference: {e}", None, None

# Main inference function.
# Returns a tuple: (annotated_image, annotated_video, metadata)
# For image inputs, the video output is None; for video inputs, the image output is None.
def inference(model_file, input_media, media_type, conf, frame_step):
    model = load_model(model_file)
    if isinstance(model, str):  # This indicates an error during model loading.
        return model, None, {"processing_time": None, "detections": None}

    if media_type == "Image":
        out_img, process_time, detections = predict_image(model, input_media, conf)
        metadata = {"processing_time": process_time, "detections": detections}
        return out_img, None, metadata

    elif media_type == "Video":
        out_vid, process_time, detections = predict_video(model, input_media, conf, frame_step)
        metadata = {"processing_time": process_time, "detections": detections}
        return None, out_vid, metadata
    else:
        return "Unsupported media type", None, {"processing_time": None, "detections": None}

# Define Gradio interface components.
# Component for uploading a custom YOLO model (.pt file).
model_file_input = gr.File(label="Upload Custom YOLO Model (.pt file)")

# Component for uploading an image or video.
media_file_input = gr.File(label="Upload Image/Video File")

# Radio button to choose media type.
media_type_dropdown = gr.Radio(choices=["Image", "Video"], label="Select Media Type", value="Image")

# Detection confidence slider.
confidence_slider = gr.Slider(minimum=0.1, maximum=1.0, step=0.05, value=0.5, label="Detection Confidence Threshold")

# Frame step slider (for video processing).
frame_step_slider = gr.Slider(minimum=1, maximum=10, step=1, value=1, label="Frame Step (for Video Processing)")

# For display on the site:
# - Use gr.Image to display the processed image.
# - Use gr.Video to display the processed video.
# - Use gr.JSON to display the metadata.
output_image = gr.Image(label="Annotated Image")
output_video = gr.Video(label="Annotated Video")
output_metadata = gr.JSON(label="Metadata")

# Create the Gradio interface.
# Note: The function returns a triple: (processed image, processed video, metadata).
iface = gr.Interface(
    fn=inference,
    inputs=[model_file_input, media_file_input, media_type_dropdown, confidence_slider, frame_step_slider],
    outputs=[output_image, output_video, output_metadata],
    title="Custom YOLO Model Inference for Real-Time Detection",
    description=(
        "Upload your custom YOLO model (detection, segmentation, or OBB) along with an image or video file "
        "to run inference. Adjust the detection confidence and frame step (for video) as needed. "
        "The app shows the processed image/video and returns metadata for real-time API integration. "
        "This is optimized for users who wish to host a YOLO model on Hugging Face and use it for real-time "
        "object detection via the Gradio API."
    )
)

if __name__ == "__main__":
    iface.launch()