Create pipeline.py

pipeline.py

@@ -0,0 +1,43 @@
+from typing import Dict, List, Any
+import PIL
+import torch
+import base64
+import os
+import io
+from transformers import ViTImageProcessor, VisionEncoderDecoderModel
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+
+class PreTrainedPipeline():
+    def __init__(self, path=""):
+        self.model = VisionEncoderDecoderModel.from_pretrained(
+            pretrained_model_name_or_path=path,
+            config=os.path.join(path, 'config.json')
+        )
+        self.model.eval()
+        self.model = self.model.to(device)
+
+        self.processor = ViTImageProcessor.from_pretrained(
+            pretrained_model_name_or_path=os.path.join(
+                path, 'preprocessor_config.json')
+        )
+
+    def __call__(self, data: Any) -> Dict[str, List[float]]:
+        """
+        Args:
+            data (:str:):
+                Inputs should be an image encoded in base 64.
+        Return:
+            A :obj:`dict`:. The object returned should be a dict like
+                {"feature_vector": [0.6331314444541931,...,-0.7866355180740356,]} containing :
+                - "feature_vector": A list of floats corresponding to the image embedding.
+        """
+        # decode base64 image to PIL
+        image = PIL.Image.open(io.BytesIO(base64.b64decode(data)))
+
+        inputs = self.processor(images=image, return_tensors="pt")
+        outputs = self.model.encoder(**inputs)
+        feature_vector = outputs.last_hidden_state[0, 0, :].tolist()
+        # postprocess the prediction
+        return {"feature_vector": feature_vector}