modified nomic encode images

nomic_fewshot.py

@@ -228,34 +228,26 @@ class NomicVisionEncoderONNX:
 
     def encode_images(self, images: list[Image.Image]) -> np.ndarray:
         rgb = [img.convert("RGB") for img in images]
-        processed = self.processor(images=rgb, return_tensors="np")
-
+        processed = self.processor(images=rgb, return_tensors="pt")
+    
         if "pixel_values" not in processed:
             raise RuntimeError(f"Processor did not return pixel_values. Keys={list(processed.keys())}")
-
-        pixel_values = processed["pixel_values"]
-        pixel_values = (
-            pixel_values.numpy().astype(np.float32)
-            if hasattr(pixel_values, "numpy")
-            else np.asarray(pixel_values, dtype=np.float32)
-        )
-
-        feeds = {}
+    
+        pixel_values = processed["pixel_values"].detach().cpu().numpy().astype(np.float32)
+    
         if self._pixel_name is None:
             raise RuntimeError(f"Could not find pixel input in ONNX inputs: {self.input_names}")
-        feeds[self._pixel_name] = pixel_values
-
-        outputs = self.session.run(self.output_names, feeds)
+    
+        outputs = self.session.run(self.output_names, {self._pixel_name: pixel_values})
         main_out = _pick_output(outputs, self.output_names, kind="vision")
-
-        # Current PyTorch behavior: CLS token from last_hidden_state
+    
         if main_out.ndim == 3:
             embs = main_out[:, 0, :]
         elif main_out.ndim == 2:
             embs = main_out
         else:
             raise RuntimeError(f"Unexpected vision output rank: {main_out.ndim}")
-
+    
         return _l2_normalize(embs, axis=1)