|
import streamlit as st |
|
import pandas as pd |
|
from datasets import load_dataset |
|
from random import sample |
|
from utils.metric import Regard |
|
from utils.model import gpt2 |
|
import os |
|
|
|
|
|
st.title('Gender Bias Analysis in Text Generation') |
|
|
|
|
|
def check_password(): |
|
def password_entered(): |
|
if password_input == os.getenv('PASSWORD'): |
|
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: |
|
st.session_state['bold'] = load_dataset("AlexaAI/bold", split="train") |
|
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'][0] for p in st.session_state['male_bold']] |
|
st.session_state['female_prompts'] = [p['prompts'][0] 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'], |
|
}) |
|
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'] |
|
}) |
|
st.write(samples_df) |
|
|
|
if st.button('Evaluate'): |
|
st.subheader('Step 4: Regard Results') |
|
regard = Regard("compare") |
|
st.write('Computing regard results to compare male and female continuations...') |
|
|
|
with st.spinner('Computing regard results...'): |
|
regard_results = regard.compute(data=st.session_state['male_continuations'], |
|
references=st.session_state['female_continuations']) |
|
st.write('**Raw Regard Results:**') |
|
st.json(regard_results) |
|
|
|
regard_results_avg = regard.compute(data=st.session_state['male_continuations'], |
|
references=st.session_state['female_continuations'], |
|
aggregation='average') |
|
st.write('**Average Regard Results:**') |
|
st.json(regard_results_avg) |