Anonymous Authors
Upload 10 files
6257003
raw
history blame
No virus
15.4 kB
import gradio as gr
import json
import numpy as np
import pandas as pd
from datasets import load_from_disk
from itertools import chain
import operator
pd.options.plotting.backend = "plotly"
TITLE = "Identity Biases in Diffusion Models: Professions"
_INTRO = """
# Identity Biases in Diffusion Models: Professions
Explore profession-level social biases in the data from [DiffusionBiasExplorer](https://hf.co/spaces/tti-bias/DiffusionBiasExplorer)!
This demo leverages the gender and ethnicity representation clusters described in the [companion app](https://hf.co/spaces/tti-bias/DiffusionFaceClustering)
to analyze social trends in machine-generated visual representations of professions.
The **Professions Overview** tab lets you compare the distribution over
[identity clusters](https://hf.co/spaces/tti-bias/DiffusionFaceClustering "Identity clusters identify visual features in the systems' output space correlated with variation of gender and ethnicity in input prompts.")
across professions for Stable Diffusion and Dalle-2 systems (or aggregated for `All Models`).
The **Professions Focus** tab provides more details for each of the individual professions, including direct system comparisons and examples of profession images for each cluster.
This work was done in the scope of the [Stable Bias Project](https://hf.co/spaces/tti-bias/StableBias).
As you use this demo, please share findings and comments [in the discussions tab](https://hf.co/spaces/tti-bias/DiffusionClustering/discussions)!
"""
_ = """
For example, you can use this tool to investigate:
- How do each model's representation of professions correlate with the gender ratios reported by the [U.S. Bureau of Labor
Statistics](https://www.bls.gov/cps/cpsaat11.htm "The reported percentage of women in each profession in the US is indicated in the `Labor Women` column in the Professions Overview tab.")?
Are social trends reflected, are they exaggerated?
- Which professions have the starkest differences in how different models represent them?
"""
professions_dset = load_from_disk("professions")
professions_df = professions_dset.to_pandas()
clusters_dicts = dict(
(num_cl, json.load(open(f"clusters/professions_to_clusters_{num_cl}.json")))
for num_cl in [12, 24, 48]
)
cluster_summaries_by_size = json.load(open("clusters/cluster_summaries_by_size.json"))
prompts = pd.read_csv("promptsadjectives.csv")
professions = ["all professions"] + list(
# sorted([p.lower() for p in prompts["Occupation-Noun"].tolist()])
sorted([p for p in prompts["Occupation-Noun"].tolist()])
)
models = {
"All": "All Models",
"SD_14": "Stable Diffusion 1.4",
"SD_2": "Stable Diffusion 2",
"DallE": "Dall-E 2",
}
df_models = {
"All Models": "All",
"Stable Diffusion 1.4": "SD_14",
"Stable Diffusion 2": "SD_2",
"Dall-E 2": "DallE",
}
def describe_cluster(num_clusters, block="label"):
cl_dict = clusters_dicts[num_clusters]
labels_values = sorted(cl_dict.items(), key=operator.itemgetter(1))
labels_values.reverse()
total = float(sum(cl_dict.values()))
lv_prcnt = list(
(item[0], round(item[1] * 100 / total, 0)) for item in labels_values
)
top_label = lv_prcnt[0][0]
description_string = (
"<span>The most represented %s is <b>%s</b>, making up about <b>%d%%</b> of the cluster.</span>"
% (to_string(block), to_string(top_label), lv_prcnt[0][1])
)
description_string += "<p>This is followed by: "
for lv in lv_prcnt[1:]:
description_string += "<BR/><b>%s:</b> %d%%" % (to_string(lv[0]), lv[1])
description_string += "</p>"
return description_string
def make_profession_plot(num_clusters, prof_name):
sorted_cl_scores = [
(k, v)
for k, v in sorted(
clusters_dicts[num_clusters]["All"][prof_name][
"cluster_proportions"
].items(),
key=lambda x: x[1],
reverse=True,
)
if v > 0
]
pre_pandas = dict(
[
(
models[mod_name],
dict(
(
f"Cluster {k}",
clusters_dicts[num_clusters][mod_name][prof_name][
"cluster_proportions"
][k],
)
for k, _ in sorted_cl_scores
),
)
for mod_name in models
]
)
df = pd.DataFrame.from_dict(pre_pandas)
prof_plot = df.plot(kind="bar", barmode="group")
cl_summary_text = f"Profession '{prof_name}':\n"
for cl_id, _ in sorted_cl_scores:
cl_summary_text += f"- {cluster_summaries_by_size[str(num_clusters)][int(cl_id)].replace(' gender terms', '').replace('; ethnicity terms:', ',')} \n"
return (
prof_plot,
gr.update(
choices=[k for k, _ in sorted_cl_scores], value=sorted_cl_scores[0][0]
),
gr.update(value=cl_summary_text),
)
def make_profession_table(num_clusters, prof_names, mod_name, max_cols=8):
professions_list_clusters = [
(
prof_name,
clusters_dicts[num_clusters][df_models[mod_name]][prof_name][
"cluster_proportions"
],
)
for prof_name in prof_names
]
totals = sorted(
[
(
k,
sum(
prof_clusters[str(k)]
for _, prof_clusters in professions_list_clusters
),
)
for k in range(num_clusters)
],
key=lambda x: x[1],
reverse=True,
)[:max_cols]
prof_list_pre_pandas = [
dict(
[
("Profession", prof_name),
(
"Entropy",
clusters_dicts[num_clusters][df_models[mod_name]][prof_name][
"entropy"
],
),
(
"Labor Women",
clusters_dicts[num_clusters][df_models[mod_name]][prof_name][
"labor_fm"
][0],
),
("", ""),
]
+ [(f"Cluster {k}", prof_clusters[str(k)]) for k, v in totals if v > 0]
)
for prof_name, prof_clusters in professions_list_clusters
]
clusters_df = pd.DataFrame.from_dict(prof_list_pre_pandas)
cl_summary_text = ""
for cl_id, _ in totals[:max_cols]:
cl_summary_text += f"- {cluster_summaries_by_size[str(num_clusters)][cl_id].replace(' gender terms', '').replace('; ethnicity terms:', ',')} \n"
return (
[c[0] for c in totals],
(
clusters_df.style.background_gradient(
axis=None, vmin=0, vmax=100, cmap="YlGnBu"
)
.format(precision=1)
.to_html()
),
gr.update(value=cl_summary_text),
)
def get_image(model, fname, score):
return (
professions_dset.select(
professions_df[
(professions_df["image_path"] == fname)
& (professions_df["model"] == model)
].index
)["image"][0],
" ".join(fname.split("/")[0].split("_")[4:])
+ f" | {score:.2f}"
+ f" | {models[model]}",
)
def show_examplars(num_clusters, prof_name, cl_id, confidence_threshold=0.6):
# only show images where the similarity to the centroid is > confidence_threshold
examplars_dict = clusters_dicts[num_clusters]["All"][prof_name][
"cluster_examplars"
][str(cl_id)]
l = [
tuple(img)
for img in examplars_dict["close"]
+ examplars_dict["mid"][:2]
+ examplars_dict["far"]
]
l = [
img
for i, img in enumerate(l)
if img[0] > confidence_threshold and img not in l[:i]
]
return (
[get_image(model, fname, score) for score, model, fname in l],
gr.update(
label=f"Generations for profession ''{prof_name}'' assigned to cluster {cl_id} of {num_clusters}"
),
)
with gr.Blocks(title=TITLE) as demo:
gr.Markdown(_INTRO)
gr.HTML(
"""<span style="color:red" font-size:smaller>⚠️ DISCLAIMER: the images displayed by this tool were generated by text-to-image systems and may depict offensive stereotypes or contain explicit content.</span>"""
)
with gr.Tab("Professions Overview"):
gr.Markdown(
"""
Select one or more professions and models from the dropdowns on the left to see which clusters are most representative for this combination.
Try choosing different numbers of clusters to see if the results change, and then go to the 'Profession Focus' tab to go more in-depth into these results.
The `Labor Women` column provided for comparison corresponds to the gender ratio reported by the
[U.S. Bureau of Labor Statistics](https://www.bls.gov/cps/cpsaat11.htm) for each profession.
"""
)
with gr.Row():
with gr.Column(scale=1):
gr.Markdown("Select the parameters here:")
num_clusters = gr.Radio(
[12, 24, 48],
value=12,
label="How many clusters do you want to use to represent identities?",
)
model_choices = gr.Dropdown(
[
"All Models",
"Stable Diffusion 1.4",
"Stable Diffusion 2",
"Dall-E 2",
],
value="All Models",
label="Which models do you want to compare?",
interactive=True,
)
profession_choices_overview = gr.Dropdown(
professions,
value=[
"all professions",
"CEO",
"director",
"social assistant",
"social worker",
],
label="Which professions do you want to compare?",
multiselect=True,
interactive=True,
)
with gr.Column(scale=3):
with gr.Row():
table = gr.HTML(
label="Profession assignment per cluster", wrap=True
)
with gr.Row():
# clusters = gr.Dataframe(type="array", visible=False, col_count=1)
clusters = gr.Textbox(label="clusters", visible=False)
gr.Markdown(
"""
##### What do the clusters mean?
Below is a summary of the identity cluster compositions.
For more details, see the [companion demo](https://huggingface.co/spaces/tti-bias/DiffusionFaceClustering):
"""
)
with gr.Row():
with gr.Accordion(label="Cluster summaries", open=True):
cluster_descriptions_table = gr.Text(
"TODO", label="Cluster summaries", show_label=False
)
with gr.Tab("Profession Focus"):
with gr.Row():
with gr.Column():
gr.Markdown(
"Select a profession to visualize and see which clusters and identity groups are most represented in the profession, as well as some examples of generated images below."
)
profession_choice_focus = gr.Dropdown(
choices=professions,
value="scientist",
label="Select profession:",
)
num_clusters_focus = gr.Radio(
[12, 24, 48],
value=12,
label="How many clusters do you want to use to represent identities?",
)
with gr.Column():
plot = gr.Plot(
label=f"Makeup of the cluster assignments for profession {profession_choice_focus}"
)
with gr.Row():
with gr.Column():
gr.Markdown(
"""
##### What do the clusters mean?
Below is a summary of the identity cluster compositions.
For more details, see the [companion demo](https://huggingface.co/spaces/tti-bias/DiffusionFaceClustering):
"""
)
with gr.Accordion(label="Cluster summaries", open=True):
cluster_descriptions = gr.Text(
"TODO", label="Cluster summaries", show_label=False
)
with gr.Column():
gr.Markdown(
"""
##### What's in the clusters?
You can show examples of profession images assigned to each identity cluster by selecting one here:
"""
)
with gr.Accordion(label="Cluster selection", open=True):
cluster_id_focus = gr.Dropdown(
choices=[i for i in range(num_clusters_focus.value)],
value=0,
label="Select cluster to visualize:",
)
with gr.Row():
examplars_plot = gr.Gallery(
label="Profession images assigned to the selected cluster."
).style(grid=4, height="auto", container=True)
demo.load(
make_profession_table,
[num_clusters, profession_choices_overview, model_choices],
[clusters, table, cluster_descriptions_table],
queue=False,
)
demo.load(
make_profession_plot,
[num_clusters_focus, profession_choice_focus],
[plot, cluster_id_focus, cluster_descriptions],
queue=False,
)
demo.load(
show_examplars,
[
num_clusters_focus,
profession_choice_focus,
cluster_id_focus,
],
[examplars_plot, examplars_plot],
queue=False,
)
for var in [num_clusters, model_choices, profession_choices_overview]:
var.change(
make_profession_table,
[num_clusters, profession_choices_overview, model_choices],
[clusters, table, cluster_descriptions_table],
queue=False,
)
for var in [num_clusters_focus, profession_choice_focus]:
var.change(
make_profession_plot,
[num_clusters_focus, profession_choice_focus],
[plot, cluster_id_focus, cluster_descriptions],
queue=False,
)
for var in [num_clusters_focus, profession_choice_focus, cluster_id_focus]:
var.change(
show_examplars,
[
num_clusters_focus,
profession_choice_focus,
cluster_id_focus,
],
[examplars_plot, examplars_plot],
queue=False,
)
if __name__ == "__main__":
demo.queue().launch(debug=True)