emilylearning's picture
Added adult example. Tighten up presentation.
fdbad51
raw
history blame
14.1 kB
# %%
import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import random
from matplotlib.ticker import MaxNLocator
from transformers import pipeline
MODEL_NAMES = ["bert-base-uncased", "roberta-base", "bert-large-uncased", "roberta-large"]
OWN_MODEL_NAME = 'add-a-model'
DECIMAL_PLACES = 1
EPS = 1e-5 # to avoid /0 errors
# Example date conts
DATE_SPLIT_KEY = "DATE"
START_YEAR = 1801
STOP_YEAR = 1999
NUM_PTS = 20
DATES = np.linspace(START_YEAR, STOP_YEAR, NUM_PTS).astype(int).tolist()
DATES = [f'{d}' for d in DATES]
# Example place conts
# https://www3.weforum.org/docs/WEF_GGGR_2021.pdf
# Bottom 10 and top 10 Global Gender Gap ranked countries.
PLACE_SPLIT_KEY = "PLACE"
PLACES = [
"Afghanistan",
"Yemen",
"Iraq",
"Pakistan",
"Syria",
"Democratic Republic of Congo",
"Iran",
"Mali",
"Chad",
"Saudi Arabia",
"Switzerland",
"Ireland",
"Lithuania",
"Rwanda",
"Namibia",
"Sweden",
"New Zealand",
"Norway",
"Finland",
"Iceland"]
# Example Reddit interest consts
# in order of increasing self-identified female participation.
# See http://bburky.com/subredditgenderratios/ , Minimum subreddit size: 400000
SUBREDDITS = [
"GlobalOffensive",
"pcmasterrace",
"nfl",
"sports",
"The_Donald",
"leagueoflegends",
"Overwatch",
"gonewild",
"Futurology",
"space",
"technology",
"gaming",
"Jokes",
"dataisbeautiful",
"woahdude",
"askscience",
"wow",
"anime",
"BlackPeopleTwitter",
"politics",
"pokemon",
"worldnews",
"reddit.com",
"interestingasfuck",
"videos",
"nottheonion",
"television",
"science",
"atheism",
"movies",
"gifs",
"Music",
"trees",
"EarthPorn",
"GetMotivated",
"pokemongo",
"news",
# removing below subreddit as most of the tokens are taken up by it:
# ['ff', '##ff', '##ff', '##fu', '##u', '##u', '##u', '##u', '##u', '##u', '##u', '##u', '##u', '##u', '##u', ...]
# "fffffffuuuuuuuuuuuu",
"Fitness",
"Showerthoughts",
"OldSchoolCool",
"explainlikeimfive",
"todayilearned",
"gameofthrones",
"AdviceAnimals",
"DIY",
"WTF",
"IAmA",
"cringepics",
"tifu",
"mildlyinteresting",
"funny",
"pics",
"LifeProTips",
"creepy",
"personalfinance",
"food",
"AskReddit",
"books",
"aww",
"sex",
"relationships",
]
GENDERED_LIST = [
['he', 'she'],
['him', 'her'],
['his', 'hers'],
["himself", "herself"],
['male', 'female'],
['man', 'woman'],
['men', 'women'],
["husband", "wife"],
['father', 'mother'],
['boyfriend', 'girlfriend'],
['brother', 'sister'],
["actor", "actress"],
]
# %%
# Fire up the models
models = dict()
for bert_like in MODEL_NAMES:
models[bert_like] = pipeline("fill-mask", model=bert_like)
# %%
def get_gendered_token_ids():
male_gendered_tokens = [list[0] for list in GENDERED_LIST]
female_gendered_tokens = [list[1] for list in GENDERED_LIST]
return male_gendered_tokens, female_gendered_tokens
def prepare_text_for_masking(input_text, mask_token, gendered_tokens, split_key):
text_w_masks_list = [
mask_token if word.lower() in gendered_tokens else word for word in input_text.split()]
num_masks = len([m for m in text_w_masks_list if m == mask_token])
text_portions = ' '.join(text_w_masks_list).split(split_key)
return text_portions, num_masks
def get_avg_prob_from_pipeline_outputs(mask_filled_text, gendered_token, num_preds):
pronoun_preds = [sum([
pronoun["score"] if pronoun["token_str"].strip().lower() in gendered_token else 0.0
for pronoun in top_preds])
for top_preds in mask_filled_text
]
return round(sum(pronoun_preds) / (EPS + num_preds) * 100, DECIMAL_PLACES)
# %%
def get_figure(df, gender, n_fit=1):
df = df.set_index('x-axis')
cols = df.columns
xs = list(range(len(df)))
ys = df[cols[0]]
fig, ax = plt.subplots()
# Trying small fig due to rendering issues on HF, not on VS Code
fig.set_figheight(3)
fig.set_figwidth(9)
# find stackoverflow reference
p, C_p = np.polyfit(xs, ys, n_fit, cov=1)
t = np.linspace(min(xs)-1, max(xs)+1, 10*len(xs))
TT = np.vstack([t**(n_fit-i) for i in range(n_fit+1)]).T
# matrix multiplication calculates the polynomial values
yi = np.dot(TT, p)
C_yi = np.dot(TT, np.dot(C_p, TT.T)) # C_y = TT*C_z*TT.T
sig_yi = np.sqrt(np.diag(C_yi)) # Standard deviations are sqrt of diagonal
ax.fill_between(t, yi+sig_yi, yi-sig_yi, alpha=.25)
ax.plot(t, yi, '-')
ax.plot(df, 'ro')
ax.legend(list(df.columns))
ax.axis('tight')
ax.set_xlabel("Value injected into input text")
ax.set_title(
f"Probability of predicting {gender} pronouns.")
ax.set_ylabel(f"Softmax prob for pronouns")
ax.xaxis.set_major_locator(MaxNLocator(6))
ax.tick_params(axis='x', labelrotation=5)
return fig
# %%
def predict_gender_pronouns(
model_name,
own_model_name,
indie_vars,
split_key,
normalizing,
n_fit,
input_text,
):
"""Run inference on input_text for each model type, returning df and plots of percentage
of gender pronouns predicted as female and male in each target text.
"""
if model_name not in MODEL_NAMES:
model = pipeline("fill-mask", model=own_model_name)
else:
model = models[model_name]
mask_token = model.tokenizer.mask_token
indie_vars_list = indie_vars.split(',')
male_gendered_tokens, female_gendered_tokens = get_gendered_token_ids()
text_segments, num_preds = prepare_text_for_masking(
input_text, mask_token, male_gendered_tokens + female_gendered_tokens, split_key)
male_pronoun_preds = []
female_pronoun_preds = []
for indie_var in indie_vars_list:
target_text = f"{indie_var}".join(text_segments)
mask_filled_text = model(target_text)
# Quick hack as realized return type based on how many MASKs in text.
if type(mask_filled_text[0]) is not list:
mask_filled_text = [mask_filled_text]
female_pronoun_preds.append(get_avg_prob_from_pipeline_outputs(
mask_filled_text,
female_gendered_tokens,
num_preds
))
male_pronoun_preds.append(get_avg_prob_from_pipeline_outputs(
mask_filled_text,
male_gendered_tokens,
num_preds
))
if normalizing:
total_gendered_probs = np.add(
female_pronoun_preds, male_pronoun_preds)
female_pronoun_preds = np.around(
np.divide(female_pronoun_preds, total_gendered_probs+EPS)*100,
decimals=DECIMAL_PLACES
)
male_pronoun_preds = np.around(
np.divide(male_pronoun_preds, total_gendered_probs+EPS)*100,
decimals=DECIMAL_PLACES
)
results_df = pd.DataFrame({'x-axis': indie_vars_list})
results_df['female_pronouns'] = female_pronoun_preds
results_df['male_pronouns'] = male_pronoun_preds
female_fig = get_figure(results_df.drop(
'male_pronouns', axis=1), 'female', n_fit,)
male_fig = get_figure(results_df.drop(
'female_pronouns', axis=1), 'male', n_fit,)
display_text = f"{random.choice(indie_vars_list)}".join(text_segments)
return (
display_text,
female_fig,
male_fig,
results_df,
)
# %%
title = "Causing Gender Pronouns"
description = """
## Intro
"""
date_example = [
MODEL_NAMES[1],
'',
', '.join(DATES),
'DATE',
"False",
1,
'She was a teenager in DATE.'
]
place_example = [
MODEL_NAMES[0],
'',
', '.join(PLACES),
'PLACE',
"False",
1,
'She became an adult in PLACE.'
]
subreddit_example = [
MODEL_NAMES[3],
'',
', '.join(SUBREDDITS),
'SUBREDDIT',
"False",
1,
'She was a kid. SUBREDDIT.'
]
own_model_example = [
OWN_MODEL_NAME,
'emilyalsentzer/Bio_ClinicalBERT',
', '.join(DATES),
'DATE',
"False",
1,
'She was exposed to the virus in DATE.'
]
def date_fn():
return date_example
def place_fn():
return place_example
def reddit_fn():
return subreddit_example
def your_fn():
return own_model_example
# %%
demo = gr.Blocks()
with demo:
gr.Markdown("# Spurious Correlation Evaluation for Pre-trained LLMs")
gr.Markdown("Find spurious correlations between seemingly independent variables (for example between `gender` and `time`) in almost any BERT-like LLM on Hugging Face, below.")
gr.Markdown("See why this happens how in our paper, [Selection Bias Induced Spurious Correlations in Large Language Models](https://arxiv.org/pdf/2207.08982.pdf), presented at [ICML 2022 Workshop on Spurious Correlations, Invariance, and Stability](https://sites.google.com/view/scis-workshop/home).")
gr.Markdown("## Instructions for this Demo")
gr.Markdown("1) Click on one of the examples below (where we sweep through a spectrum of `places`, `dates` and `subreddits`) to pre-populate the input fields.")
gr.Markdown("2) Check out the pre-populated fields as you scroll down to the ['Hit Submit...'] button!")
gr.Markdown("3) Repeat steps (1) and (2) with more pre-populated inputs or with your own values in the input fields!")
gr.Markdown("## Example inputs")
gr.Markdown("Click a button below to pre-populate input fields with example values. Then scroll down to Hit Submit to generate predictions.")
with gr.Row():
date_gen = gr.Button('Click for date example inputs')
gr.Markdown("<-- x-axis sorted by older to more recent dates:")
place_gen = gr.Button('Click for country example inputs')
gr.Markdown(
"<-- x-axis sorted by bottom 10 and top 10 [Global Gender Gap](https://www3.weforum.org/docs/WEF_GGGR_2021.pdf) ranked countries:")
subreddit_gen = gr.Button('Click for Subreddit example inputs')
gr.Markdown(
"<-- x-axis sorted in order of increasing self-identified female participation (see [bburky](http://bburky.com/subredditgenderratios/)): ")
your_gen = gr.Button('Add-a-model example inputs')
gr.Markdown("<-- x-axis dates, with your own model loaded! (If first time, try another example, it can take a while to load new model.)")
gr.Markdown("## Input fields")
gr.Markdown(
f"A) Pick a spectrum of comma separated values for text injection and x-axis.")
with gr.Row():
x_axis = gr.Textbox(
lines=3,
label="A) Comma separated values for text injection and x-axis",
)
gr.Markdown("B) Pick a pre-loaded BERT-family model of interest on the right.")
gr.Markdown(f"Or C) select `{OWN_MODEL_NAME}`, then add the mame of any other Hugging Face model that supports the [fill-mask](https://huggingface.co/models?pipeline_tag=fill-mask) task on the right (note: this may take some time to load).")
with gr.Row():
model_name = gr.Radio(
MODEL_NAMES + [OWN_MODEL_NAME],
type="value",
label="B) BERT-like model.",
)
own_model_name = gr.Textbox(
label="C) If you selected an 'add-a-model' model, put any Hugging Face pipeline model name (that supports the fill-mask task) here.",
)
gr.Markdown("D) Pick if you want to the predictions normalied to these gendered terms only.")
gr.Markdown("E) Also tell the demo what special token you will use in your input text, that you would like replaced with the spectrum of values you listed above.")
gr.Markdown("And F) the degree of polynomial fit used for high-lighting potential spurious association.")
with gr.Row():
to_normalize = gr.Dropdown(
["False", "True"],
label="D) Normalize model's predictions to only the gendered ones?",
type="index",
)
place_holder = gr.Textbox(
label="E) Special token place-holder",
)
n_fit = gr.Dropdown(
list(range(1, 5)),
label="F) Degree of polynomial fit",
type="value",
)
gr.Markdown(
"G) Finally, add input text that includes at least one gendered pronouns and one place-holder token specified above.")
with gr.Row():
input_text = gr.Textbox(
lines=2,
label="G) Input text with pronouns and place-holder token",
)
gr.Markdown("## Outputs!")
#gr.Markdown("Scroll down and 'Hit Submit'!")
with gr.Row():
btn = gr.Button("Hit submit to generate predictions!")
with gr.Row():
sample_text = gr.Textbox(
type="auto", label="Output text: Sample of text fed to model")
with gr.Row():
female_fig = gr.Plot(type="auto")
male_fig = gr.Plot(type="auto")
with gr.Row():
df = gr.Dataframe(
show_label=True,
overflow_row_behaviour="show_ends",
label="Table of softmax probability for pronouns predictions",
)
with gr.Row():
date_gen.click(date_fn, inputs=[], outputs=[model_name, own_model_name,
x_axis, place_holder, to_normalize, n_fit, input_text])
place_gen.click(place_fn, inputs=[], outputs=[
model_name, own_model_name, x_axis, place_holder, to_normalize, n_fit, input_text])
subreddit_gen.click(reddit_fn, inputs=[], outputs=[
model_name, own_model_name, x_axis, place_holder, to_normalize, n_fit, input_text])
your_gen.click(your_fn, inputs=[], outputs=[
model_name, own_model_name, x_axis, place_holder, to_normalize, n_fit, input_text])
btn.click(
predict_gender_pronouns,
inputs=[model_name, own_model_name, x_axis, place_holder,
to_normalize, n_fit, input_text],
outputs=[sample_text, female_fig, male_fig, df])
demo.launch(debug=True)
# %%