app.py

import gradio as gr
import torch
from torchvision import transforms
from PIL import Image
import numpy as np
from trainer import CustomSemanticSegmentationTask

# Load a pre-trained semantic segmentation model
task = CustomSemanticSegmentationTask.load_from_checkpoint("maui_demo_model.ckpt", map_location="cpu")
task.freeze()
model = task.model
model = model.eval()


# Define the image transformations
preprocess = transforms.Compose([
    transforms.ToTensor(),
])

# Function to perform semantic segmentation
def segment_image(image):
    input_tensor = preprocess(image).unsqueeze(0)
    with torch.inference_mode():
        output = model(input_tensor)
    output_predictions = output.argmax(1).squeeze().numpy()
    return output_predictions


# Preexisting images
preexisting_images = ["crop1.png", "crop2.png", "crop3.png"]

# Function to handle user input and run the model
def handle_image(image):
    image = Image.open(image)
    mask = segment_image(image)

    # Decode the segmentation output
    colormap = np.array([
        [0, 0, 0],  # nodata
        [0, 0, 0],  # background
        [0, 255, 0],  # building
        [255, 0, 0],  # damage
    ])
    output = colormap[mask].astype('uint8')

    segmented_image = np.array(image)
    segmented_image[mask > 1] = (0.5 * output[mask > 1]) + (0.5 * segmented_image[mask > 1])
    segmented_image = Image.fromarray(segmented_image)
    return segmented_image

# Create the Gradio interface
image_input = gr.Image(type="filepath", label="Upload an Image", sources=["upload"])
image_output = gr.Image(type="pil", label="Output")

css_content = """
.legend {
    list-style: none;
    padding: 0;
}
.legend li {
    line-height: 20px; /* Match the height of the color-box */
}
.legend .color-box {
    display: inline-block;
    width: 20px;
    height: 20px;
    margin-right: 5px;
    border: 1px solid #000; /* Optional: adds a border around the color box */
    vertical-align: middle; /* Centers the box vertically relative to the text */
}
.background { background-color: #FFFFFF; } /* White */
.building { background-color: #00FF00; } /* Green */
.damage { background-color: #FF0000; } /* Red */
"""

html_content = """
<div style="font-size:large;">
    <p>
        This application demonstrates the input and output of the building damage assessment model trained through the tutorial of the
        <a href="https://github.com/microsoft/building-damage-assessment/" target="_blank">Microsoft AI for Good Building Damage
        Assessment Toolkit</a>. This particular model was trained on 
        <a href="https://radiantearth.github.io/stac-browser/#/external/maxar-opendata.s3.amazonaws.com/events/Maui-Hawaii-fires-Aug-23/collection.json?.language=en">Maxar Open Data imagery</a>
        captured over Lahaina during the Maui Wildfires in August, 2023 and 106 polygon annotations
        created over the same imagery. The "Building Damage Assessment Toolkit" details a workflow for quickly
        modeling <i>any</i> new post disaster imagery by:
        <ul style='padding-left:20px'>
            <li>Setting up a web labeling tool instance with the post disaster imagery to facilitate rapid annotations of the imagery</li>
            <li>Fine-tuning a building damage assessment model using the imagery and annotations</li>
            <li>Running inference with the model over potentially large scenes</li>
            <li>Merging and summarizing the output of the model using different building footprints layers</li>
        </ul>
        This workflow allows a user to consistently and rapidly create a good model for a particular event,
        but that is overfit to that event (i.e. it will not generalize to other events).
    </p>
    <p>
        The model outputs per-pixel predictions of building damage, with the following classes:
        <ul class="legend">
            <li><span class="color-box background"></span>Background (transparent)</li>
            <li><span class="color-box building"></span>Building (green)</li>
            <li><span class="color-box damage"></span>Damage (red)</li>
        </ul>
    </p>
</div>
"""

iface = gr.Interface(
    fn=handle_image,
    inputs=image_input,
    outputs=image_output,
    title="Building damage assessment model demo -- Maui Wildfires 2023",
    examples=preexisting_images,
    css=css_content,
    description=html_content,
    allow_flagging="never"
)

# Launch the app
iface.launch(share=True)