Update

README.md

@@ -1,6 +1,6 @@
 ---
 title: I Bet You Did Not Mean That
-emoji: 🔎
+emoji: 🤔
 colorFrom: blue
 colorTo: indigo
 sdk: streamlit

app_lib/main.py

@@ -56,7 +56,7 @@ def main(device=torch.device("cuda" if torch.cuda.is_available() else "cpu")):
             st.error(error_message)
 
         with st.container():
-            significance_level, tau_max, r, cardinality = get_advanced_settings(
+            significance_level, tau_max, r, cardinality, dataset_name = get_advanced_settings(
                 concepts, concepts_ready
             )
 
@@ -79,7 +79,7 @@ def main(device=torch.device("cuda" if torch.cuda.is_available() else "cpu")):
                 class_name,
                 concepts,
                 cardinality,
-                "imagenette",
+                dataset_name,
                 model_name,
                 device,
             )

app_lib/test.py

@@ -168,7 +168,7 @@ def test(
 
     progress_bar = st.progress(
         0,
-        text=f"Testing concepts (can take a few minutes) [0 / {len(concepts)} completed]",
+        text=f"Testing concepts (can take up to a minute) [0 / {len(concepts)} completed]",
     )
 
     embedding = _load_dataset(dataset_name, model_name)
@@ -179,7 +179,7 @@ def test(
 
     progress_bar.progress(
         1 / (len(concepts) + 1),
-        text=f"Testing concepts (can take a few minutes) [0 / {len(concepts)} completed]",
+        text=f"Testing concepts (can take up to a minute) [0 / {len(concepts)} completed]",
     )
 
     with ThreadPoolExecutor() as executor:
@@ -202,7 +202,7 @@ def test(
             results.append(future.result())
             progress_bar.progress(
                 (idx + 2) / (len(concepts) + 1),
-                text=f"Testing concepts (can take a few minutes) [{idx + 1} / {len(concepts)} completed]",
+                text=f"Testing concepts (can take up to a minute) [{idx + 1} / {len(concepts)} completed]",
             )
 
     rejected = np.empty((testing_config.r, len(concepts)))

app_lib/user_input.py

@@ -2,7 +2,7 @@ import streamlit as st
 from PIL import Image
 from streamlit_image_select import image_select
 
-from app_lib.utils import SUPPORTED_MODELS
+from app_lib.utils import SUPPORTED_MODELS, SUPPORTED_DATASETS
 
 
 def _validate_class_name(class_name):
@@ -43,7 +43,7 @@ def _get_tau_max():
     DEFAULT = 200
     return int(
         st.slider(
-            "Duration of test",
+            "Length of test",
             help=" ".join(
                 [
                     "The maximum number of steps for each test.",
@@ -80,6 +80,7 @@ def _get_number_of_tests():
 
 
 def _get_cardinality(concepts, concepts_ready):
+    DEFAULT = lambda concepts: int(len(concepts) / 2)
     return st.slider(
         "Size of conditioning set",
         help=" ".join(
@@ -90,12 +91,28 @@ def _get_cardinality(concepts, concepts_ready):
         ),
         min_value=1,
         max_value=max(2, len(concepts) - 1),
-        value=int(len(concepts) / 2),
+        value=DEFAULT(concepts),
         step=1,
         disabled=st.session_state.disabled or not concepts_ready,
     )
 
 
+def _get_dataset_name():
+    DEFAULT = SUPPORTED_DATASETS.index("imagenette")
+    return st.selectbox(
+        "Dataset",
+        options=SUPPORTED_DATASETS,
+        index=DEFAULT,
+        help=" ".join(
+            [
+                "Name of the dataset to use to train sampler.",
+                "Defaults to Imagenette.",
+            ]
+        ),
+        disabled=st.session_state.disabled,
+    )
+
+
 def get_model_name():
     return st.selectbox(
         "Model to test",
@@ -150,10 +167,12 @@ def get_concepts():
 
 
 def get_advanced_settings(concepts, concepts_ready):
-    with st.popover("Advanced settings", disabled=st.session_state.disabled):
+    with st.expander("Advanced settings"):
+        dataset_name = _get_dataset_name()
         significance_level = _get_significance_level()
         tau_max = _get_tau_max()
         r = _get_number_of_tests()
         cardinality = _get_cardinality(concepts, concepts_ready)
+        st.divider()
 
-    return significance_level, tau_max, r, cardinality
+    return significance_level, tau_max, r, cardinality, dataset_name

app_lib/utils.py

@@ -20,7 +20,7 @@ with open(supported_models_path, "r") as f:
 
 
 SUPPORTED_DATASETS = []
-with open(supported_models_path, "r") as f:
+with open(supported_datasets_path, "r") as f:
     for line in f:
         dataset_name = line.strip()
         SUPPORTED_DATASETS.append(dataset_name)

app_lib/viz.py

@@ -27,7 +27,7 @@ def viz_results():
     results = st.session_state.results
 
     if results is None:
-        st.info("Run tests to show results", icon="ℹ️")
+        st.info("Test concepts to show results", icon="ℹ️")
     else:
         rank_tab, wealth_tab = st.tabs(["Rank of importance", "Wealth process"])
 

header.md

@@ -1,3 +1,4 @@
 # 🤔 I Bet You Did Not Mean That
 
-Official 🤗 Space for the paper [*I Bet You Did Not Mean That: Testing Semantic Importance via Betting*](https://arxiv.org/pdf/2405.19146), by [Jacopo Teneggi](https://jacopoteneggi.github.io) and [Jeremias Sulam](https://sites.google.com/view/jsulam).
+Test the importance of semantic concepts for the predictions of a classifier. [[paper]](https://arxiv.org/pdf/2405.19146) [[code]](https://github.com/Sulam-Group/IBYDMT)
+

style.css

@@ -19,8 +19,23 @@ h1 {
     justify-content: center;
 }
 
-[data-testid="stVerticalBlock"]:has(> [data-testid="stPopover"]) {
+[data-testid="stVerticalBlock"]:has(> [data-testid="stExpander"]) {
     display: block;
+
+    details {
+        border: 0;
+    }
+
+    summary {
+        padding: 0;
+        display: inline-flex;
+        align-items: center;
+        width: fit-content;
+    }
+
+    hr {
+        margin-top: 0;
+    }
 }
 
 [data-testid="stPopover"] {
@@ -46,9 +61,7 @@ h1 {
     }
 }
 
-[data-testid="stSpinner"] {
-    >div {
-        display: flex;
-        justify-content: center;
-    }
+[data-testid="stSpinner"]>div {
+    display: flex;
+    justify-content: center;
 }