Add conditioning, row filtering, fix ranking, markdown header

app.py

@@ -27,6 +27,7 @@ EXPECTED_KEY_TO_COLNAME = OrderedDict(
         ("rank", "Rank"),  # Just for columns order
         ("model", "Model"),  # Just for columns order
         ("model_size", "Model Size (Million)"),  # Just for columns order
+        ("conditioning", "Conditioning"),
         ("embedding_dim", "Embedding Dimension"),
     ]
     + [
@@ -95,11 +96,13 @@ def get_data_from_hub():
     return pd.DataFrame(df_list, columns=EXPECTED_KEY_TO_COLNAME.values())
 
 
-def filter_columns(df, k_filter, d_filter):
+def filter_dataframe(df, k_filter, d_filter, c_filter):
+    # ===== FILTER COLUMNS
     # Fixed column positions
     selected_columns = [
         EXPECTED_KEY_TO_COLNAME["rank"],
         EXPECTED_KEY_TO_COLNAME["model"],
+        EXPECTED_KEY_TO_COLNAME["conditioning"],
         EXPECTED_KEY_TO_COLNAME["model_size"],
         EXPECTED_KEY_TO_COLNAME["embedding_dim"],
     ]
@@ -124,12 +127,20 @@ def filter_columns(df, k_filter, d_filter):
     selected_columns.append(EXPECTED_KEY_TO_COLNAME["n_dists"])
     datatypes.append("number")
 
+    df = df[selected_columns]
+
+    # ===== FILTER ROWS
+    df = df[df[EXPECTED_KEY_TO_COLNAME["conditioning"]].isin(c_filter)]
+
     return df[selected_columns], datatypes
 
 
 def add_rank(df):
     main_metrics = df["R@1 +1M Dist."].str.split("±").str[0].astype(float)
-    df["Rank"] = main_metrics.argsort() + 1
+
+    # Argsort is smallest to largest so we reverse it
+    # We add 1 to start the rank at 1 instead of 0
+    df["Rank"] = main_metrics.argsort().values[::-1] + 1
     return df
 
 
@@ -141,14 +152,16 @@ def save_current_leaderboard(df):
     return filename
 
 
-def load_lrvsf_models(k_filter, d_filter, csv_file):
+def load_lrvsf_models(k_filter, d_filter, c_filter, csv_file):
     # Remove previous tmpfile
     if csv_file:
         os.remove(csv_file)
 
     df = get_data_from_hub()
     df = add_rank(df)
-    df, datatypes = filter_columns(df, k_filter, d_filter)
+    df, datatypes = filter_dataframe(df, k_filter, d_filter, c_filter)
+    df = df.sort_values(by="Rank")
+
     filename = save_current_leaderboard(df)
 
     outputs = [
@@ -163,7 +176,11 @@ if __name__ == "__main__":
     with gr.Blocks() as demo:
         gr.Markdown(
             """
-                # LAION - Referred Visual Search - Fashion : Leaderboard
+                # LAION - Referred Visual Search - Fashion 👗 Leaderboard
+
+                - To submit, refer to the [LAION-RVS-Fashion Benchmark repository](https://github.com/Simon-Lepage/LRVSF-Benchmark).
+                - For details on the task and the dataset, refer to the [LRVSF paper](https://arxiv.org/abs/2306.02928).
+                - To download the leaderboard as CSV, click on the file below the table.
             """
         )
         with gr.Row():
@@ -175,6 +192,11 @@ if __name__ == "__main__":
                 value=DIST_EVALUATIONS,
                 label="Number of Distractors",
             )
+            c_filter = gr.CheckboxGroup(
+                choices=["category", "text"],
+                value=["category", "text"],
+                label="Conditioning",
+            )
 
         df_table = gr.Dataframe(type="pandas", interactive=False)
         csv_file = gr.File(interactive=False)
@@ -183,12 +205,12 @@ if __name__ == "__main__":
         # Actions
         refresh.click(
             load_lrvsf_models,
-            inputs=[k_filter, d_filter, csv_file],
+            inputs=[k_filter, d_filter, c_filter, csv_file],
             outputs=[df_table, csv_file],
         )
         demo.load(
             load_lrvsf_models,
-            inputs=[k_filter, d_filter, csv_file],
+            inputs=[k_filter, d_filter, c_filter, csv_file],
             outputs=[df_table, csv_file],
         )