Update examples

app.py

@@ -1,5 +1,5 @@
-
 import base64
+import os.path
 from io import BytesIO
 from pathlib import Path
 
@@ -14,45 +14,53 @@ from matplotlib import rcParams
 from msclip.inference import run_inference_classification
 
 rcParams["font.size"] = 9
+rcParams["axes.titlesize"] = 9
 IMG_PX = 300
 
 EXAMPLES = {
     "EuroSAT": {
         "images": glob.glob("examples/eurosat/*.tif"),
         "classes": [
-            "AnnualCrop","Forest","HerbaceousVegetation","Highway","Industrial",
-            "Pasture","PermanentCrop","Residential","River","SeaLake"
+            "Annual crop", "Forest", "Herbaceous vegetation", "Highway", "Industrial",
+            "Pasture", "Permanent crop", "Residential", "River", "Sea lake"
         ]
     },
     "Meter-ML": {
         "images": glob.glob("examples/meterml/*.tif"),
         "classes": [
-            "Todo"
+            "Concentrated animal feeding operations",
+            "Landfills",
+            "Coal mines",
+            "Other features",
+            "Natural gas processing plants",
+            "Oil refineries and petroleum terminals",
+            "Wastewater treatment plants",
         ]
     },
     "TerraMesh": {
         "images": glob.glob("examples/terramesh/*.tif"),
         "classes": [
-            "Agriculture", "Beach", "River", "Ice", "Fields"
+            "Village", "Beach", "River", "Ice", "Fields", "Mountains", "Desert"
         ]
     },
 }
 
 
 def load_eurosat_example():
-    return EXAMPLES["EuroSAT"]["images"], ",".join(EXAMPLES["EuroSAT"]["classes"])
+    return EXAMPLES["EuroSAT"]["images"], ", ".join(EXAMPLES["EuroSAT"]["classes"])
 
 
 def load_meterml_example():
-    return EXAMPLES["Meter-ML"]["images"], ",".join(EXAMPLES["Meter-ML"]["classes"])
+    return EXAMPLES["Meter-ML"]["images"], ", ".join(EXAMPLES["Meter-ML"]["classes"])
 
 
 def load_terramesh_example():
-    return EXAMPLES["TerraMesh"]["images"], ",".join(EXAMPLES["TerraMesh"]["classes"])
+    return EXAMPLES["TerraMesh"]["images"], ", ".join(EXAMPLES["TerraMesh"]["classes"])
 
 
 pastel1_hex = [mpl.colors.to_hex(c) for c in mpl.colormaps["Pastel1"].colors]
 
+
 def build_colormap(class_names):
     return {c: pastel1_hex[i % len(pastel1_hex)] for i, c in enumerate(sorted(class_names))}
 
@@ -74,7 +82,7 @@ def _rgb_smooth_quantiles(array, tolerance=0.02, scaling=0.5, default=2000):
     array = np.where(array <= limit_high, array, limit_high + (array - limit_high) * scaling)
 
     # Update scaling params using a 10th of the tolerance for max value
-    limit_low, limit_high = np.quantile(array, q=[tolerance/10, 1. - tolerance/10])
+    limit_low, limit_high = np.quantile(array, q=[tolerance / 10, 1. - tolerance / 10])
     limit_high = limit_high.clip(default, 20000)  # Scale only pixels above default value
     limit_low = limit_low.clip(0, 500)  # Scale only pixels below 500
     limit_low = np.where(median > default / 2, limit_low, 0)  # Make image only darker if it is not dark already
@@ -121,7 +129,7 @@ def _img_to_b64(path: str | Path) -> str:
     return base64.b64encode(buf.getvalue()).decode()
 
 
-def _bar_chart(top_scores, cmap) -> str:
+def _bar_chart(top_scores, img_name, cmap) -> str:
     scores = top_scores.values.tolist()
     labels = top_scores.index.tolist()
     while len(scores) < 3:
@@ -138,10 +146,16 @@ def _bar_chart(top_scores, cmap) -> str:
     ax.set_xlim(0, 1)
     ax.invert_yaxis()
     ax.axis("off")
+    img_name = os.path.splitext(img_name)[0]
+    if len(img_name) > 25:
+        img_name = img_name[:23] + "..."
+    ax.set_title(img_name)
 
     for i, (cls, val) in enumerate(zip(labels, scores)):
+        if len(cls) > 25:
+            cls = cls[:23] + "..."
         if val > 0:  # skip padded rows
-            ax.text(0.02, i+0.03, f"{cls}  ({round(val * 100)}%)", ha="left", va="center")
+            ax.text(0.02, i + 0.03, f"{cls}  ({round(val * 100)}%)", ha="left", va="center")
 
     buf = BytesIO()
     fig.savefig(buf, format="png", dpi=300, bbox_inches="tight", transparent=True)
@@ -154,7 +168,7 @@ def classify(images, class_text):
     class_names = [c.strip() for c in class_text.split(",") if c.strip()]
     cards = []
 
-    df = run_inference_classification(image_path=images, class_names=class_names)  # one row per call
+    df = run_inference_classification(image_path=images, class_names=class_names, verbose=False)
     for img_path, (id, row) in zip(images, df.iterrows()):
         scores = row[2:].astype(float)  # drop filename column
         top = scores.sort_values(ascending=False)[:3]
@@ -166,21 +180,18 @@ def classify(images, class_text):
           <img src="data:image/png;base64,{_img_to_b64(img_path)}"
                style="width:{IMG_PX}px;height:{IMG_PX}px;object-fit:cover;
                       border-radius:8px;box-shadow:0 2px 6px rgba(0,0,0,.15);display:block;margin:auto;">
-          {_bar_chart(top, cmap)}
+          {_bar_chart(top, os.path.basename(img_path), cmap)}
         </div>""")
 
     return (
-        "<div style='display:flex;flex-wrap:wrap;justify-content:center;'>"
-        + "".join(cards)
-        + "</div>"
+            "<div style='display:flex;flex-wrap:wrap;justify-content:center;'>"
+            + "".join(cards)
+            + "</div>"
     )
 
 
 # UI
 
-DEFAULT_CLASSES = ["Forest", "River", "Buildings", "Agriculture", "Mountain", "Snow"]
-
-# with gr.Blocks(css=".gradio-container") as demo:
 with gr.Blocks(
         css="""
             .gradio-container
@@ -188,8 +199,9 @@ with gr.Blocks(
             #result_box.gr-skeleton {min-height:280px !important;}
             """) as demo:
     gr.Markdown("## Zero‑shot Classification with Llama3-MS‑CLIP")
-    gr.Markdown("Provide Sentinel-2 tif files with all 12 or 13 bands and define the class names. "
-                "You can also load one of the three provided example sets with class names that you can modify. The example images are comming from [EuroSAT](), [Meter-ML](), and [TerraMesh](). "
+    gr.Markdown("Provide Sentinel-2 L2A tif files with all 12 bands and define the class names for running zero-shot classification. "
+                "You can also use S-2 L1C files with 13 bands but the model might not work as well (e.g., misclassifing forests as sea because of the differrently scaled values). "
+                "We provide three sets of example images with class names that you can modify. The examples are from [EuroSAT](https://arxiv.org/abs/1709.00029), [Meter-ML](https://arxiv.org/abs/2207.11166), and [TerraMesh](https://arxiv.org/abs/2504.11172) (We downloaded S-2 L2A images for the same locations). "
                 "The images are classified based on the similarity between the images embeddings and text embeddings. "
                 "You find more information in the [model card](https://huggingface.co/ibm-esa-geospatial/Llama3-MS-CLIP-base) and the [paper](https://arxiv.org/abs/2503.15969). ")
     with gr.Row():
@@ -197,7 +209,8 @@ with gr.Blocks(
             label="Upload S-2 images", file_count="multiple", type="filepath"
         )
         cls_in = gr.Textbox(
-            value=", ".join(DEFAULT_CLASSES),
+            value=", ".join(["Forest", "River", "Buildings", "Agriculture", "Mountain", "Snow"]),
+            # some default classes
             label="Class names (comma‑separated)",
         )
 
@@ -243,6 +256,5 @@ with gr.Blocks(
         outputs=out_html,
     )
 
-
 if __name__ == "__main__":
     demo.launch()