pages/2_new_Demo_1.py

import streamlit as st
import pandas as pd
from datasets import load_dataset, Dataset
from random import sample
from utils.metric import Regard
from utils.model import gpt2
import matplotlib.pyplot as plt
import os

# Set up the Streamlit interface
st.title('Gender Bias Analysis in Text Generation')


def check_password():
    def password_entered():
        if password_input == os.getenv('PASSWORD'):
        # if password_input == "  ":
            st.session_state['password_correct'] = True
        else:
            st.error("Incorrect Password, please try again.")

    password_input = st.text_input("Enter Password:", type="password")
    submit_button = st.button("Submit", on_click=password_entered)

    if submit_button and not st.session_state.get('password_correct', False):
        st.error("Please enter a valid password to access the demo.")


if not st.session_state.get('password_correct', False):
    check_password()
else:
    st.sidebar.success("Password Verified. Proceed with the demo.")

    if 'data_size' not in st.session_state:
        st.session_state['data_size'] = 10
    if 'bold' not in st.session_state:
        bold = pd.DataFrame({})
        bold_raw = pd.DataFrame(load_dataset("AlexaAI/bold", split="train"))
        for index, row in bold_raw.iterrows():
            bold_raw_prompts = list(row['prompts'])
            bold_raw_wikipedia = list(row['wikipedia'])
            bold_expansion = zip(bold_raw_prompts, bold_raw_wikipedia)
            for bold_prompt, bold_wikipedia in bold_expansion:
                bold = bold._append(
                    {'domain': row['domain'], 'name': row['name'], 'category': row['category'], 'prompts': bold_prompt,
                     'wikipedia': bold_wikipedia}, ignore_index=True)
        st.session_state['bold'] = Dataset.from_pandas(bold)
    if 'female_bold' not in st.session_state:
        st.session_state['female_bold'] = []
    if 'male_bold' not in st.session_state:
        st.session_state['male_bold'] = []

    st.subheader('Step 1: Set Data Size')
    data_size = st.slider('Select number of samples per category:', min_value=1, max_value=50,
                          value=st.session_state['data_size'])
    st.session_state['data_size'] = data_size

    if st.button('Show Data'):
        st.session_state['female_bold'] = sample(
            [p for p in st.session_state['bold'] if p['category'] == 'American_actresses'], data_size)
        st.session_state['male_bold'] = sample(
            [p for p in st.session_state['bold'] if p['category'] == 'American_actors'], data_size)

        st.write(f'Sampled {data_size} female and male American actors.')
        st.write('**Female Samples:**', pd.DataFrame(st.session_state['female_bold']))
        st.write('**Male Samples:**', pd.DataFrame(st.session_state['male_bold']))

    if st.session_state['female_bold'] and st.session_state['male_bold']:
        st.subheader('Step 2: Generate Text')

        if st.button('Generate Text'):
            GPT2 = gpt2()
            st.session_state['male_prompts'] = [p['prompts'] for p in st.session_state['male_bold']]
            st.session_state['female_prompts'] = [p['prompts'] for p in st.session_state['female_bold']]
            st.session_state['male_wiki_continuation'] = [p['wikipedia'].replace(p['prompts'], '') for p in
                                                          st.session_state['male_bold']]
            st.session_state['female_wiki_continuation'] = [p['wikipedia'].replace(p['prompts'], '') for p in
                                                            st.session_state['female_bold']]

            progress_bar = st.progress(0)

            st.write('Generating text for male prompts...')
            male_generation = GPT2.text_generation(st.session_state['male_prompts'], pad_token_id=50256, max_length=50,
                                                   do_sample=False, truncation=True)
            st.session_state['male_continuations'] = [gen[0]['generated_text'].replace(prompt, '') for gen, prompt in
                                                      zip(male_generation, st.session_state['male_prompts'])]

            progress_bar.progress(50)

            st.write('Generating text for female prompts...')
            female_generation = GPT2.text_generation(st.session_state['female_prompts'], pad_token_id=50256,
                                                     max_length=50, do_sample=False, truncation=True)
            st.session_state['female_continuations'] = [gen[0]['generated_text'].replace(prompt, '') for gen, prompt in
                                                        zip(female_generation, st.session_state['female_prompts'])]

            progress_bar.progress(100)
            st.write('Text generation completed.')

    if st.session_state.get('male_continuations') and st.session_state.get('female_continuations'):
        st.subheader('Step 3: Sample Generated Texts')

        st.write("Male Data Samples:")
        samples_df = pd.DataFrame({
            'Male Prompt': st.session_state['male_prompts'],
            'Male Continuation': st.session_state['male_continuations'],
            'Male Wiki Continuation': st.session_state['male_wiki_continuation'],
        })
        st.write(samples_df)

        st.write("Female Data Samples:")
        samples_df = pd.DataFrame({
            'Female Prompt': st.session_state['female_prompts'],
            'Female Continuation': st.session_state['female_continuations'],
            'Female Wiki Continuation': st.session_state['female_wiki_continuation'],
        })
        st.write(samples_df)

        if st.button('Evaluate'):
            st.subheader('Step 4: Regard Results')
            regard = Regard("inner_compare")
            st.write('Computing regard results to compare male and female continuations...')

            with st.spinner('Computing regard results...'):
                regard_male_results = regard.compute(data=st.session_state['male_continuations'],
                                                     references=st.session_state['male_wiki_continuation'])
                st.write('**Raw Regard Results:**')
                st.json(regard_male_results)
                st.session_state['rmr'] = regard_male_results

                regard_female_results = regard.compute(data=st.session_state['female_continuations'],
                                                       references=st.session_state['female_wiki_continuation'])
                st.write('**Average Regard Results:**')
                st.json(regard_female_results)
                st.session_state['rfr'] = regard_female_results

        if st.button('Plot'):
            st.subheader('Step 5: Regard Results Plotting')
            categories = ['GPT2', 'Wiki']

            mp_gpt = st.session_state['rmr']['no_ref_diff_mean']['positive']
            mn_gpt = st.session_state['rmr']['no_ref_diff_mean']['negative']
            mo_gpt = 1 - (mp_gpt + mn_gpt)

            mp_wiki = mp_gpt - st.session_state['rmr']['ref_diff_mean']['positive']
            mn_wiki = mn_gpt -st.session_state['rmr']['ref_diff_mean']['negative']
            mo_wiki = 1 - (mn_wiki + mp_wiki)

            fp_gpt = st.session_state['rfr']['no_ref_diff_mean']['positive']
            fn_gpt = st.session_state['rfr']['no_ref_diff_mean']['negative']
            fo_gpt = 1 - (fp_gpt + fn_gpt)

            fp_wiki = fp_gpt - st.session_state['rfr']['ref_diff_mean']['positive']
            fn_wiki = fn_gpt - st.session_state['rfr']['ref_diff_mean']['negative']
            fo_wiki = 1 - (fn_wiki + fp_wiki)

            positive_m = [mp_gpt, mp_wiki]
            other_m = [mo_gpt, mo_wiki]
            negative_m = [mn_gpt, mn_wiki]

            positive_f = [fp_gpt, fp_wiki]
            other_f = [fo_gpt, fo_wiki]
            negative_f = [fn_gpt, fn_wiki]

            # Plotting
            fig_a, ax_a = plt.subplots()
            ax_a.bar(categories, negative_m, label='Negative', color='blue')
            ax_a.bar(categories, other_m, bottom=negative_m, label='Other', color='orange')
            ax_a.bar(categories, positive_m, bottom=[negative_m[i] + other_m[i] for i in range(len(negative_m))],
                     label='Positive', color='green')

            plt.xlabel('Categories')
            plt.ylabel('Proportion')
            plt.title('GPT vs Wiki on male regard')
            plt.legend()

            st.pyplot(fig_a)

            fig_b, ax_b = plt.subplots()
            ax_b.bar(categories, negative_f, label='Negative', color='blue')
            ax_b.bar(categories, other_f, bottom=negative_f, label='Other', color='orange')
            ax_b.bar(categories, positive_f, bottom=[negative_f[i] + other_f[i] for i in range(len(negative_f))],
                     label='Positive', color='green')

            plt.xlabel('Categories')
            plt.ylabel('Proportion')
            plt.title('GPT vs Wiki on female regard')
            plt.legend()
            st.pyplot(fig_b)

            m_increase = mp_gpt - mn_gpt
            m_relative_increase = mp_gpt - mp_wiki - (mn_gpt - mn_wiki)
            f_increase = fp_gpt - fn_gpt
            f_relative_increase = fp_gpt - fp_wiki - (fn_gpt - fn_wiki)

            absolute_difference = [m_increase, f_increase]
            relative_difference = [m_relative_increase, f_relative_increase]

            new_categories = ['Male', 'Female']

            fig_c, ax_c = plt.subplots()
            ax_c.bar(new_categories, absolute_difference, label='Positive - Negative', color='#40E0D0')

            plt.xlabel('Categories')
            plt.ylabel('Proportion')
            plt.title('Difference of positive and negative: Male vs Female')
            plt.legend()
            st.pyplot(fig_c)

            fig_d, ax_d = plt.subplots()
            ax_d.bar(new_categories, relative_difference, label='Positive - Negative', color='#40E0D0')

            plt.xlabel('Categories')
            plt.ylabel('Proportion')
            plt.title('Difference of positive and negative (relative to Wiki): Male vs Female')
            plt.legend()
            st.pyplot(fig_d)