init

handler.py

@@ -0,0 +1,29 @@
+from typing import Dict, List, Any
+import numpy as np
+from transformers import CLIPProcessor, CLIPModel
+from PIL import Image
+from io import BytesIO
+import base64
+
+class EndpointHandler():
+    def __init__(self, path=""):
+        # Preload all the elements you are going to need at inference.
+        self.model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
+        self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+
+    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
+        """
+       data args:
+            inputs (:obj: `str` | `PIL.Image` | `np.array`)
+            kwargs
+      Return:
+            A :obj:`list` | `dict`: will be serialized and returned
+        """
+
+        words = data.pop("words", data)
+        image = Image.open(BytesIO(base64.b64decode(data['image'])))
+        inputs = self.processor(text=words, images=image, return_tensors="pt", padding=True)
+        outputs = self.model(**inputs)
+        embeddings = outputs.image_embeds.detach().numpy().flatten().tolist()
+        return {"embeddings": embeddings}

requirements.txt

@@ -0,0 +1,2 @@
+pillow
+numpy