app.py

from datasets import load_dataset
import gradio as gr
import os
import random

wmtis = load_dataset("nlphuji/wmtis-identify")['test']
print(f"Loaded WMTIS identify, first example:")
print(wmtis[0])
dataset_size = len(wmtis) - 1

NATURAL_IMAGE = 'natural_image'
NORMAL_IMAGE = 'normal_image'
STRANGE_IMAGE = 'strange_image'

def func(index):
    example = wmtis[index]
    outputs = []
    target_size = example['normal_image'].size
    add_outputs_for_key(example, outputs, target_size, 'natural')
    add_outputs_for_key(example, outputs, target_size, 'normal')
    add_outputs_for_key(example, outputs, target_size, 'strange')

    return outputs


def add_outputs_for_key(example, outputs, target_size, item):
    for item_key in [f'{item}_image', f'{item}_image_caption', f'rating_{item}', f'comments_{item}', f'{item}_hash']:
        if item_key == f'comments_{item}':
            outputs.append(get_empty_comment_if_needed(example[item_key]))
        elif item_key == f'{item}_image':
            outputs.append(example[item_key].resize(target_size))
        else:
            outputs.append(example[item_key])


demo = gr.Blocks()

def get_empty_comment_if_needed(item):
    if item == 'nan':
        return '-'
    return item


def add_column_by_key(item, target_size):
    with gr.Column():
        img = wmtis[index][f"{item}_image"]
        img_resized = img.resize(target_size)
        i1 = gr.Image(value=img_resized, label=f'{item.capitalize()} Image')
        p1 = gr.Textbox(value=wmtis[index][f"{item}_image_caption"], label='BLIP2 Predicted Caption')
        r1 = gr.Textbox(value=wmtis[index][f"rating_{item}"], label='Rating')
        c1 = gr.Textbox(value=get_empty_comment_if_needed(wmtis[index][f"comments_{item}"]), label='Comments')
        t1 = gr.Textbox(value=wmtis[index][f"{item}_hash"], label='Image ID')
        item_outputs = [i1, p1, r1, c1, t1]
        return item_outputs


with demo:
    gr.Markdown("# Main Challenge: Weirdness, not Synthesis")

    with gr.Column():
        slider = gr.Slider(minimum=0, maximum=dataset_size)
        with gr.Row():
            index = slider.value
            if index >= dataset_size:
                index = 0
            target_size = wmtis[index]['normal_image'].size
            natural_outputs = add_column_by_key('natural', target_size)
            normal_outputs = add_column_by_key('normal', target_size)
            strange_outputs = add_column_by_key('strange', target_size)

    slider.change(func, inputs=[slider], outputs=natural_outputs + normal_outputs + strange_outputs)

demo.launch()