Fix similarity score

Author: qbonnard

api/commands.py

@@ -48,6 +48,11 @@ def insert_images(
     }
 
 
+def similarity_score(distance):
+    # 2 is the maximum distance between normalised vectors
+    return 100 * (1 - distance / 2)
+
+
 def search(model_name, url, limit=10):
     embedding = embeddings[model_name].extract(load_image_from_url(url))
     search_results = collections[model_name].search(
@@ -68,7 +73,7 @@ def search(model_name, url, limit=10):
         {
             "url": hit.id,
             "metadata": json.loads(hit.entity.get("metadata")),
-            "similarity": 100 * (1 - hit.distance),
+            "similarity": similarity_score(hit.distance),
         }
         for hit in search_results[0]
     ]
@@ -84,7 +89,7 @@ def compare(model_name, url_left, url_right):
     # it's a bit overkill anyway if we don't compare with vectors from the db
     if metrics[model_name] == "L2":
         # _squared_ L2, to be consistent with the distances in milvus' search
-        return 100 * (1 - np.sum(np.square(np.array(left) - np.array(right))))
+        return similarity_score(np.sum(np.square(np.array(left) - np.array(right))))
 
     raise RuntimeError(
         "Distance calculation has not been implemented in the API. "

api/tests/test_commands.py

@@ -40,7 +40,7 @@ def test_crud(mock_model):
 
     assert [
         {
-            "similarity": pytest.approx(11.650939784262505),
+            "similarity": pytest.approx(55.82546989213125),
             "metadata": metadata,
             "url": TEST_URLS[0],
         }
@@ -118,7 +118,7 @@ def test_insert_without_replacing(mock_model):
 
 
 def test_compare():
-    assert pytest.approx(11.650939784262505) == commands["compare"](
+    assert pytest.approx(55.82546989213125) == commands["compare"](
         "vit_b32", TEST_URLS[0], TEST_URLS[1]
     )
 

api/tests/test_embeddings.py

@@ -3,6 +3,7 @@
 
 import random
 from PIL import Image
+import math
 
 from ..embeddings import load_image_from_url, embeddings
 
@@ -26,13 +27,18 @@ def test_load_image_from_url_and_reject_small_ones():
     )
 
 
+def squared_l2(v):
+    return sum([x * x for x in v])
+
+
 def test_extract_vit_b32():
     random.seed(2023)
     image_data = bytes([random.randint(0, 255) for _ in range(500 * 500 * 3)])
     image = Image.frombytes("RGB", (500, 500), image_data)
 
     b32 = embeddings["vit_b32"]
     embedding = b32.extract(image)
+    assert pytest.approx(1) == squared_l2(embedding)
     assert 512 == len(embedding)
     assert pytest.approx(-1.00633253128035) == sum(embedding)