app.py

import math
from datasets import load_dataset
import gradio as gr
import os
import ast

auth_token = os.environ.get("auth_token")
whoops = load_dataset("nlphuji/whoops", token=auth_token, trust_remote_code=True)['test'].shuffle()
# print(f"Loaded WHOOPS!, first example:")
# print(whoops[0])
dataset_size = len(whoops)

IMAGE = 'image'
IMAGE_DESIGNER = 'image_designer'
DESIGNER_EXPLANATION = 'designer_explanation'
CROWD_CAPTIONS = 'crowd_captions'
CROWD_EXPLANATIONS = 'crowd_explanations'
CROWD_UNDERSPECIFIED_CAPTIONS = 'crowd_underspecified_captions'
QA = 'question_answering_pairs'
IMAGE_ID = 'image_id'
SELECTED_CAPTION = 'selected_caption'
COMMONSENSE_CATEGORY = 'commonsense_category'


left_side_columns = [IMAGE]
right_side_columns = [x for x in whoops.features.keys() if x not in left_side_columns]
enumerate_cols = [CROWD_CAPTIONS, CROWD_EXPLANATIONS, CROWD_UNDERSPECIFIED_CAPTIONS]


right_side_columns.remove('image_url')

emoji_to_label = {IMAGE_DESIGNER: '🎨, 🧑‍🎨, 💻', DESIGNER_EXPLANATION: '💡, 🤔, 🧑‍🎨',
                  CROWD_CAPTIONS: '👥, 💬, 📝', CROWD_EXPLANATIONS: '👥, 💡, 🤔', CROWD_UNDERSPECIFIED_CAPTIONS: '👥, 💬, 👎',
                  QA: '❓, 🤔, 💡', IMAGE_ID: '🔍, 📄, 💾', COMMONSENSE_CATEGORY: '🤔, 📚, 💡', SELECTED_CAPTION: '📝, 👌, 💬'}
# batch_size = 16
batch_size = 8
target_size = (1024, 1024)


def func(index):
    start_index = index * batch_size
    end_index = start_index + batch_size
    all_examples = [whoops[index] for index in list(range(start_index, end_index))]
    values_lst = []
    for example_idx, example in enumerate(all_examples):
        values = get_instance_values(example)
        values_lst += values
    return values_lst


def get_instance_values(example):
    values = []
    for k in left_side_columns + right_side_columns:
        if k == IMAGE:
            value = example["image"].resize(target_size)
        elif k in enumerate_cols:
            value = list_to_string(ast.literal_eval(example[k]))
        elif k == QA:
            qa_list = [f"Q: {x[0]} A: {x[1]}" for x in ast.literal_eval(example[k])]
            value = list_to_string(qa_list)
        else:
            value = example[k]
        values.append(value)
    return values
    
def list_to_string(lst):
    return '\n'.join(['{}. {}'.format(i+1, item) for i, item in enumerate(lst)])

demo = gr.Blocks()


def get_col(example):
    instance_values = get_instance_values(example)
    with gr.Column():
        inputs_left = []
        assert len(left_side_columns) == len(
            instance_values[:len(left_side_columns)])  # excluding the image & designer
        for key, value in zip(left_side_columns, instance_values[:len(left_side_columns)]):
            if key == IMAGE:
                img_resized = example["image"].resize(target_size)
                # input_k = gr.Image(value=img_resized, label=example['commonsense_category'])
                input_k = gr.Image(value=img_resized)
            else:
                label = key.capitalize().replace("_", " ")
                input_k = gr.Textbox(value=value, label=f"{label} {emoji_to_label[key]}")
            inputs_left.append(input_k)
            
        with gr.Accordion("Click for details", open=False):
            text_inputs_right = []
            assert len(right_side_columns) == len(
                instance_values[len(left_side_columns):])  # excluding the image & designer
            for key, value in zip(right_side_columns, instance_values[len(left_side_columns):]):
                label = key.capitalize().replace("_", " ")
                
                if type(value) != str:
                    num_lines = 1
                else:
                    num_lines = max(1, len(value) // 50 + (len(value) % 45 > 0))  # Assuming ~50 chars per line
                    
                text_input_k = gr.Textbox(value=value, label=f"{label} {emoji_to_label[key]}", lines=num_lines)
                # text_input_k = gr.Textbox(value=value, label=f"{label} {emoji_to_label[key]}")
                text_inputs_right.append(text_input_k)
    return inputs_left, text_inputs_right


with demo:
    gr.Markdown("# Slide to iterate WHOOPS!")

    with gr.Column():
        num_batches = math.ceil(dataset_size / batch_size)
        slider = gr.Slider(minimum=0, maximum=num_batches, step=1, label=f'Page (out of {num_batches})')
        with gr.Row():
            index = slider.value
            start_index = 0 * batch_size
            end_index = start_index + batch_size
            all_examples = [whoops[index] for index in list(range(start_index, end_index))]
            all_inputs_left_right = []
            for example_idx, example in enumerate(all_examples):
                inputs_left, text_inputs_right = get_col(example)
                inputs_left_right = inputs_left + text_inputs_right
                all_inputs_left_right += inputs_left_right

    slider.change(func, inputs=[slider], outputs=all_inputs_left_right)

demo.launch()