all_Comps

app.py

@@ -15,7 +15,6 @@ 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]
@@ -91,8 +90,15 @@ def make_profession_plot(num_clusters, prof_name):
     )
     df = pd.DataFrame.from_dict(pre_pandas)
     prof_plot = df.plot(kind="bar", barmode="group")
-    return prof_plot, gr.update(
-        choices=[k for k, _ in sorted_cl_scores], value=sorted_cl_scores[0][0]
+    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),
     )
 
 
@@ -143,43 +149,76 @@ def make_profession_table(num_clusters, prof_names, mod_name, max_cols=8):
         for prof_name, prof_clusters in professions_list_clusters
     ]
     clusters_df = pd.DataFrame.from_dict(prof_list_pre_pandas)
-    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()
+    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)
+                (professions_df["image_path"] == fname)
+                & (professions_df["model"] == model)
             ].index
         )["image"][0],
-        " ".join(fname.split("/")[0].split("_")[4:]) + f" | {score:.2f}" + f" | {models[model]}"
+        " ".join(fname.split("/")[0].split("_")[4:])
+        + f" | {score:.2f}"
+        + f" | {models[model]}",
     )
 
 
-def show_examplars(num_clusters, prof_name, cl_id, confidence_threshold=0.5):
-    # only show images where the similarity to the centroid is > 0.7
+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]]
+    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}")
+        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("# 🤗 Diffusion Cluster Explorer")
-    gr.Markdown("This tool helps you explore the different clusters that we discovered in the images generated by 3 text-to-image models: Dall-E 2, Stable Diffusion v.1.4 and v.2. This work was done in the scope of the [Stable Bias Project](https://huggingface.co/spaces/society-ethics/StableBias).")
+    gr.Markdown(
+        """
+        # Identity Biases in Diffusion Models: Professions
+        
+        This tool helps you explore the different clusters that we discovered in the images generated by 3 text-to-image models: Dall-E 2, Stable Diffusion v.1.4 and v.2.
+        This work was done in the scope of the [Stable Bias Project](https://huggingface.co/spaces/society-ethics/StableBias).
+        """
+    )
+    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.")
+        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."
+        )
         with gr.Row():
             with gr.Column(scale=1):
                 gr.Markdown("Select the parameters here:")
@@ -211,86 +250,118 @@ with gr.Blocks(title=TITLE) as demo:
                     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)
-                demo.load(
-                    make_profession_table,
-                    [num_clusters, profession_choices_overview, model_choices],
-                    [clusters, table],
-                    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],
-                        queue=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/society-ethics/DiffusionFaceClustering):
+                        """
                     )
-
+                with gr.Row():
+                    with gr.Accordion(label="Cluster summaries", open=False):
+                        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 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.")
-                num_clusters_focus = gr.Radio(
-                    [12, 24, 48],
-                    value=12,
-                    label="How many clusters do you want to use to represent identities?",
+                gr.Markdown(
+                    "Select 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="social worker",
+                    value="scientist",
                     label="Select profession:",
                 )
-                gr.Markdown(
-                    "You can show examples of profession images assigned to each cluster:"
-                )
-                cluster_id_focus = gr.Dropdown(
-                    choices=[i for i in range(num_clusters_focus.value)],
-                    value=0,
-                    label="Select cluster to visualize:",
+                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}"
                 )
-                demo.load(
-                    make_profession_plot,
-                    [num_clusters_focus, profession_choice_focus],
-                    [plot, cluster_id_focus],
-                    queue=False,
+        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/society-ethics/DiffusionFaceClustering):
+                    """
+                )
+                with gr.Accordion(label="Cluster summaries", open=False):
+                    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:
+                    """
                 )
-                for var in [num_clusters_focus, profession_choice_focus]:
-                    var.change(
-                        make_profession_plot,
-                        [num_clusters_focus, profession_choice_focus],
-                        [plot, cluster_id_focus],
-                        queue=False,
+                with gr.Accordion(label="Cluster selection", open=False):
+                    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(
-                show_examplars,
-                [
-                    num_clusters_focus,
-                    profession_choice_focus,
-                    cluster_id_focus,
-                ],
-                [examplars_plot, examplars_plot],
-                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,
-                )
+    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__":