app.py

import torch
import gradio as gr
import cv2
import numpy as np
from sahi.utils.yolov5 import download_yolov5s6_model

# import required functions, classes
from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction, visualize_object_predictions

# Autodetect GPU
if torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")

# Load the model
yolov5_model_path = "best.pt"
download_yolov5s6_model(destination_path=yolov5_model_path)
detection_model = AutoDetectionModel.from_pretrained(
    model_type="yolov5",
    model_path=yolov5_model_path,
    confidence_threshold=0.01,
    device=device,
)


def do_detection(image_path, hide_labels, confidence_scores):

    # Obtain detection results
    result = get_sliced_prediction(
        image_path,
        detection_model,
        slice_height=512,
        slice_width=512,
        overlap_height_ratio=0.12,
        overlap_width_ratio=0.12,
    )

    # Filter detections according to the slider and count the number of classes
    # for visualization
    predictions = []
    class_counts = {}
    for i in result.object_prediction_list:
        score = i.score
        value = score.value
        category = i.category
        category_name = category.name
        if value > confidence_scores[category_name]:
            predictions.append(i)
            if i.category.name not in class_counts:
                class_counts[i.category.name] = 1
            else:
                class_counts[i.category.name] += 1

    # Draw the boxes and labels on top of the image
    img_rgb = visualize_object_predictions(
        image_path,
        object_prediction_list=predictions,
        text_size=1,
        text_th=1,
        hide_labels=hide_labels,
        rect_th=3,
    )["image"]

    # Construct a legend
    legend_text = "Symbols Counted:"
    for class_name, count in class_counts.items():
        legend_text += f" {class_name}: {count} |"

    font = cv2.FONT_HERSHEY_SIMPLEX
    if hide_labels:
        font_scale = 1.5
    else:
        font_scale = 1
    font_color = (255, 255, 255)
    font_thickness = 2
    legend_bg_color = (131, 79, 0)
    legend_padding = 10

    legend_size, _ = cv2.getTextSize(legend_text, font, font_scale, font_thickness)
    legend_bg_height = legend_size[1] + 2 * legend_padding
    legend_bg_width = legend_size[0] + 2 * legend_padding

    legend_bg = np.zeros((legend_bg_height, legend_bg_width, 3), dtype=np.uint8)
    legend_bg[:] = legend_bg_color
    cv2.putText(
        legend_bg,
        legend_text,
        (legend_padding, legend_padding + legend_size[1]),
        font,
        font_scale,
        font_color,
        font_thickness,
    )

    img_height, img_width, _ = img_rgb.shape
    legend_x = img_width - legend_bg_width
    legend_y = img_height - legend_bg_height

    img_rgb[legend_y:, legend_x:, :] = legend_bg

    return (
        img_rgb,
        result.to_coco_predictions(),
    )


def call_func(
    image_path,
    hide_labels,
    singleplex_value,
    duplex_value,
    triplex_value,
    quadruplex_value,
    gfci_value,
    gfci_wp_value,
):
    confidence_scores = {
        "Singleplex - Standard": singleplex_value,
        "Duplex - Standard": duplex_value,
        "Triplex - Standard": triplex_value,
        "Quadruplex - Standard": quadruplex_value,
        "Duplex - GFCI": gfci_value,
        "Duplex - Weatherproof-GFCI": gfci_wp_value,
    }
    return do_detection(image_path, hide_labels, confidence_scores)


demo = gr.Blocks()
theme = gr.themes.Soft()

with gr.Blocks(theme=theme) as demo:
    gr.Markdown(
        """
        <h1 align="center">Receptacle Detector for Takeoff Automation</h1>
        """
    )

    with gr.Row():
        input_image = gr.Image(
            label="Upload an image here.",
            source="upload",
            interactive=True,
        )
        examples = gr.Examples(
            examples=[
                ["test1.jpg"],
                ["test2.jpg"],
                ["test3.jpg"],
                ["test4.jpg"],
            ],
            inputs=[input_image],
            examples_per_page=4,
            label="Examples to use.",
        )

    hide_labels = gr.Checkbox(label="Hide labels")
    with gr.Accordion("Visualization Confidence Thresholds", open=False):
        singleplex_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.53,
            interactive=True,
            label="Singleplex",
        )
        duplex_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.66,
            interactive=True,
            label="Duplex",
        )
        triplex_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.65,
            interactive=True,
            label="Triplex",
        )
        quadruplex_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.63,
            interactive=True,
            label="Quadruplex",
        )
        gfci_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.31,
            interactive=True,
            label="GFCI",
        )
        gfci_wp_slider = gr.Slider(
            minimum=0.1,
            maximum=1,
            value=0.33,
            interactive=True,
            label="GFCI/WP",
        )

    results_button = gr.Button("Submit")
    results_button.click(
        call_func,
        inputs=[
            input_image,
            hide_labels,
            singleplex_slider,
            duplex_slider,
            triplex_slider,
            quadruplex_slider,
            gfci_slider,
            gfci_wp_slider,
        ],
        outputs=[
            gr.Image(type="numpy", label="Output Image"),
            gr.Json(),
        ],
    )

demo.launch()