feat: assign unknowns via vss

sdcat/cluster/cluster.py

@@ -86,7 +86,7 @@ def run_vss(image_t: tuple[str, np.array], vss_url: str, project: str, vss_thres
     :param vss_url: url for vss service
     :param project: project name in vss
     :param top_k: number of vss to use for prediction; 1, 3, 5 etc.
-    :return:
+    :return: best prediction and score
     """
     url_vss = f"{vss_url}/{top_k}/{project}"
     debug(f"URL: {url_vss} threshold: {vss_threshold}")
@@ -96,13 +96,13 @@ def run_vss(image_t: tuple[str, np.array], vss_url: str, project: str, vss_thres
 
     if response.status_code != 200:
         err(f"Error processing images: {response.text}")
-        return None, None
+        return "", 0.0
 
     predictions = response.json()["predictions"]
     debug(f"Predictions: {predictions}")
 
     if len(predictions) == 0:
-        return None, None
+        return "", 0.0
 
     scores = response.json()["scores"][0]
     # Scores are  1 - score, so we need to invert them
@@ -112,7 +112,7 @@ def run_vss(image_t: tuple[str, np.array], vss_url: str, project: str, vss_thres
 
     if best_pred is None:
         err(f"No majority prediction for {image_t[0]}")
-        return None, None
+        return "", 0.0
 
     return best_pred, best_score
 
@@ -514,14 +514,27 @@ def cluster_vits(
             # Run the VSS service to assign the cluster to a class
             image_t = read_image(exemplar['image_path'])
             best_prediction, best_score = run_vss(image_t, vss_url=vss_url, vss_threshold=.1, project='901103-biodiversity', top_k=1)
-            if best_prediction is None:
+            if len(best_prediction) == 0:
                 warn(f'No predictions found for {exemplar["image_path"]}')
                 continue
             # Assign the class to the cluster in df_dets
             info(f'Assigning {cluster_id} to class {best_prediction} with score {best_score}')
             df_dets.loc[df_dets['cluster'] == cluster_id, 'class'] = best_prediction
             df_dets.loc[df_dets['cluster'] == cluster_id, 'score'] = best_score
 
+        # Try to assign everything not in a cluster to a class
+        unknowns = df_dets[df_dets['cluster'] == -1]
+        for idx, row in unknowns.iterrows():
+            image_t = read_image(row['crop_path'])
+            best_prediction, best_score = run_vss(image_t, vss_url=vss_url, vss_threshold=.1, project='901103-biodiversity', top_k=1)
+            if len(best_prediction) == 0:
+                warn(f'No predictions found for {row["crop_path"]}')
+                continue
+            # Assign the class to the cluster in df_dets
+            info(f'Assigning {row["crop_path"]} to class {best_prediction} with score {best_score}')
+            df_dets.loc[df_dets['crop_path'] == row['crop_path'], 'class'] = best_prediction
+            df_dets.loc[df_dets['crop_path'] == row['crop_path'], 'score'] = best_score
+
     # Save the exemplar embeddings with the model type
     exemplar_df['model'] = model
     exemplar_df.to_csv(output_path / f'{prefix}_exemplars.csv', index=False)