app.py

import os
from io import BytesIO
from multiprocessing import Pool, cpu_count
from datasets import load_dataset
from PIL import Image
import gradio as gr
import pandas as pd

imagenet_hard_dataset = load_dataset("taesiri/imagenet-hard", split="validation")
THUMBNAIL_PATH = "dataset/thumbnails"
os.makedirs(THUMBNAIL_PATH, exist_ok=True)

max_size = (480, 480)

all_origins = set()
all_labels = set()
dataset_df = None


def process_image(i):
    global all_origins
    image = imagenet_hard_dataset[i]["image"].convert("RGB")
    url_prefix = "https://imagenet-hard.taesiri.ai/"

    origin = imagenet_hard_dataset[i]["origin"]
    label = imagenet_hard_dataset[i]["english_label"]

    save_path = os.path.join(THUMBNAIL_PATH, origin)
    # make sure the folder exists
    os.makedirs(save_path, exist_ok=True)
    image_path = os.path.join(save_path, f"{i}.jpg")

    image.thumbnail(max_size, Image.LANCZOS)

    image.save(image_path, "JPEG", quality=100)

    url = url_prefix + image_path

    return {
        "preview": url,
        "filepath": image_path,
        "origin": imagenet_hard_dataset[i]["origin"],
        "labels": imagenet_hard_dataset[i]["english_label"],
    }


# PREPROCESSING
if os.path.exists("dataset.pkl"):
    dataset_df = pd.read_pickle("dataset.pkl")
    all_origins = set(dataset_df["origin"])
    all_labels = set().union(*dataset_df["labels"])
else:
    with Pool(cpu_count()) as pool:
        samples_data = pool.map(process_image, range(len(imagenet_hard_dataset)))
        dataset_df = pd.DataFrame(samples_data)
        print(dataset_df)
        all_origins = set(dataset_df["origin"])
        all_labels = set().union(*dataset_df["labels"])
        # save dataframe on disk
        dataset_df.to_csv("dataset.csv")
        dataset_df.to_pickle("dataset.pkl")


def get_slice(origin, label):
    global dataset_df

    if not origin and not label:
        filtered_df = dataset_df
    else:
        filtered_df = dataset_df[
            (dataset_df["origin"] == origin if origin else True)
            & (dataset_df["labels"].apply(lambda x: label in x) if label else True)
        ]

    max_value = len(filtered_df) // 16

    returned_values = []

    start_index = 0
    end_index = start_index + 16

    slice_df = filtered_df.iloc[start_index:end_index]

    for row in slice_df.itertuples():
        returned_values.append(gr.update(value=row.preview))
        returned_values.append(gr.update(value=row.origin))
        returned_values.append(gr.update(value=row.labels))

    if len(returned_values) < 48:
        returned_values.extend([None] * (48 - len(returned_values)))

    filtered_df = gr.Dataframe(filtered_df, datatype="markdown")
    return filtered_df, gr.update(maximum=max_value, value=0), *returned_values


def reset_filters_fn():
    return gr.update(value=None), gr.update(value=None)


def make_grid(grid_size):
    list_of_components = []

    with gr.Row():
        for row_counter in range(grid_size[0]):
            with gr.Column():
                for col_counter in range(grid_size[1]):
                    item_image = gr.Image()
                    with gr.Accordion("Click for details", open=False):
                        item_source = gr.Textbox(label="Source Dataset")
                        item_labels = gr.Textbox(label="Labels")

                    list_of_components.append(item_image)
                    list_of_components.append(item_source)
                    list_of_components.append(item_labels)

    return list_of_components


def slider_upadte(slider, df):
    returned_values = []

    start_index = (slider) * 16
    end_index = start_index + 16

    slice_df = df.iloc[start_index:end_index]

    for row in slice_df.itertuples():
        returned_values.append(gr.update(value=row.preview))
        returned_values.append(gr.update(value=row.origin))
        returned_values.append(gr.update(value=row.labels))

    if len(returned_values) < 48:
        returned_values.extend([None] * (48 - len(returned_values)))

    return returned_values


with gr.Blocks() as demo:
    gr.Markdown("# ImageNet-Hard Browser")
    # add link to home page and dataset
    gr.HTML("")
    gr.HTML()
    gr.HTML(
        """
    <center>
        <span style="font-size: 14px; vertical-align: middle;">
            <a href='https://zoom.taesiri.ai/'>Project Home Page</a> &nbsp;|&nbsp;
            <a href='https://huggingface.co/datasets/taesiri/imagenet-hard'>Dataset</a>
        </span>
    </center>
    """
    )

    with gr.Row():
        origin_dropdown = gr.Dropdown(all_origins, label="Origin")
        label_dropdown = gr.Dropdown(all_labels, label="Label")
    with gr.Row():
        show_btn = gr.Button("Show")
        reset_filters = gr.Button("Reset Filters")

    preview_dataframe = gr.Dataframe(height=500, visible=False)

    gr.Markdown("## Preview")

    maximum_vale = len(dataset_df) // 16

    preview_slider = gr.Slider(minimum=1, maximum=maximum_vale, step=1, value=1)
    all_components = make_grid((4, 4))

    show_btn.click(
        fn=get_slice,
        inputs=[origin_dropdown, label_dropdown],
        outputs=[preview_dataframe, preview_slider, *all_components],
    )

    reset_filters.click(
        fn=reset_filters_fn,
        inputs=[],
        outputs=[origin_dropdown, label_dropdown],
    )

    preview_slider.change(
        fn=slider_upadte,
        inputs=[preview_slider, preview_dataframe],
        outputs=[*all_components],
    )


demo.launch()