app.py

import gradio
import numpy

from matplotlib import cm
from pathlib import Path
from PIL import Image

from fastai.vision.all import load_learner, PILImage, PILMask


MODEL_PATH = Path('.') / 'models'
TEST_IMAGES_PATH = Path('.') / 'test'


def preprocess_mask(file_name):
    """Ensures masks are in grayscale format and removes transparency."""
    mask_path = Path(
        '/kaggle/inumpyut/car-segmentation/car-segmentation/masks') / file_name.name
    mask = Image.open(mask_path)

    if mask.mode == 'P':
        mask = mask.convert('RGBA')

    if mask.mode != 'RGBA':
        mask = mask.convert('RGBA')

    mask_data = mask.getdata()

    new_mask_data = [
        (r, g, b, 255) if a > 0 else (0, 0, 0, 255)
        for r, g, b, a in mask_data
    ]

    mask.putdata(new_mask_data)

    return PILMask.create(mask.convert('L'))


LEARNER = load_learner(MODEL_PATH / 'car-segmentation_v1.pkl')


def segment_image(image):
    # Store original size
    original_size = image.size  # (width, height)
    
    # Resize the input image to 400x400 for the model
    resized_image = image.resize((400, 400))
    resized_image = PILImage.create(resized_image)
    
    # Get the prediction from the model
    prediction, _, _ = LEARNER.predict(resized_image)
    
    # Convert prediction to a NumPy array
    prediction_array = numpy.asarray(prediction, dtype=numpy.uint8)
    
    # Resize the mask back to the original image size
    prediction_resized = Image.fromarray(prediction_array).resize(original_size, Image.NEAREST)
    prediction_resized = numpy.array(prediction_resized)
    
    # Apply a colormap for visualization (using the public API)
    colormap = cm._colormaps['jet']
    # Normalize the mask and apply the colormap (result is in float [0,1])
    colored_mask = colormap(prediction_resized / numpy.max(prediction_resized))[:, :, :3]
    
    # Convert the original image to a NumPy array and normalize to [0,1]
    image_array = numpy.array(image).astype(numpy.float32) / 255.0
    
    # Blend the original image and the colored mask
    overlay = (image_array * 0.7) + (colored_mask * 0.3)
    
    # Convert the blended image back to uint8
    overlay = (overlay * 255).astype(numpy.uint8)
    
    return overlay


demo = gradio.Interface(
    segment_image,
    inputs=gradio.Image(type='pil'),
    outputs=gradio.Image(type='numpy'),
    examples=[str(image) for image in TEST_IMAGES_PATH.iterdir()]
)
demo.launch()