build_image_classifier.md

Build an Image Classifier

In this tutorial, we will collaboratively build an Image Classifier that can detect wind turbines in images. As a team, we will support the following functionalities:

• Uploading an image file

• Using an example image file

• Viewing the image file

• Viewing the predicted result

We will be using a random classifier, but you can later replace it with your own custom classifier if you want.

:::{note} When we ask you to run the code in the sections below, you may either execute the code directly in the Panel docs via the green run button, in a cell in a notebook, or in a file app.py that is served with panel serve app.py --dev. :::

Install the Dependencies

Please ensure that hvPlot is installed.

::::{tab-set}

:::{tab-item} pip :sync: pip

pip install hvplot panel

:::

:::{tab-item} conda :sync: conda

conda install -y -c conda-forge hvplot panel

:::

::::

Create the App

Run the code below.

import random
from io import BytesIO
from time import sleep

import hvplot.pandas
import pandas as pd
import requests
from PIL import Image

import panel as pn

IMAGE_DIM = 350

pn.extension(design="material", sizing_mode="stretch_width")

## Transformations


@pn.cache
def get_pil_image(url):
    response = requests.get(url)
    return Image.open(BytesIO(response.content))


@pn.cache
def get_plot(label):
    data = pd.Series(label).sort_values()
    return data.hvplot.barh(title="Prediction", ylim=(0, 100)).opts(default_tools=[])


## Custom Components


def get_label_view(fn, image: Image):
    return get_plot(fn(image))


def get_image_button(url, image_pane):
    button = pn.widgets.Button(
        width=100,
        height=100,
        stylesheets=[
            f"button {{background-image: url({url});background-size: cover;}}"
        ],
    )
    pn.bind(handle_example_click, button, url, image_pane, watch=True)
    return button


## Event Handlers


def handle_example_click(event, url, image_pane):
    image_pane.object = get_pil_image(url)


def handle_file_upload(value, image_pane):
    file = BytesIO(value)
    image_pane.object = Image.open(file)


def image_classification_interface(fn, examples):

    ## State

    image_view = pn.pane.Image(
        get_pil_image(examples[0]),
        height=IMAGE_DIM,
        width=IMAGE_DIM,
        fixed_aspect=False,
        margin=0,
    )

    label_view = pn.pane.JSON()

    ## Inputs

    file_input = pn.widgets.FileInput(
        accept=".png,.jpeg",
    )
    pn.bind(handle_file_upload, file_input, image_view, watch=True)
    file_input_component = pn.Column("### Upload Image", file_input)

    examples_input_component = pn.Column(
        "### Examples", pn.Row(*(get_image_button(url, image_view) for url in examples))
    )

    ## Views

    label_view = pn.Row(
        pn.panel(
            pn.bind(get_label_view, fn=fn, image=image_view.param.object),
            defer_load=True,
            loading_indicator=True,
            height=IMAGE_DIM,
            width=IMAGE_DIM,
        )
    )

    ## Layouts

    input_component = pn.Column(
        "# Input",
        image_view,
        file_input_component,
        examples_input_component,
        width=IMAGE_DIM,
        margin=10,
    )

    output_component = pn.Column(
        "# Output",
        label_view,
        width=IMAGE_DIM,
        margin=10,
    )
    return pn.FlexBox(input_component, output_component)


def predict(image: Image):
    # Replace with your own image classification model
    sleep(1.5)
    a = random.uniform(0, 100)
    b = random.uniform(0, 100 - a)
    c = 100 - a - b
    return {
        "Wind Turbine": a,
        "Solar Panel": b,
        "Battery Storage": c,
    }


EXAMPLES = [
    # Replace with your own examples
    "https://assets.holoviz.org/panel/tutorials/wind_turbine.png",
    "https://assets.holoviz.org/panel/tutorials/solar_panel.png",
    "https://assets.holoviz.org/panel/tutorials/battery_storage.png",
]

image_classification_interface(fn=predict, examples=EXAMPLES).servable()

Try to

• Click an example image

• Upload an .png or .jpg image file

Recap

In this tutorial we have built an Image Classifier that can detect wind turbines. We will support

• Uploading an image file

• Using an example image file

• Viewing the image file

• Viewing the predicted result