add resizing

infer.py

@@ -33,15 +33,18 @@ def infer_pipe(pipe, test_image, task_name, seed, device, video_depth=False):
 
         if video_depth == False:
             test_image = Image.open(test_image).convert('RGB')
-        test_image = np.array(test_image).astype(np.float16)
+        test_image = np.array(test_image).astype(np.float32)
         if max(test_image.shape[:2]) > 1024:
             # resize for a maximum size of 1024
             scale = 1024 / max(test_image.shape[:2])
-            test_image = cv2.resize(test_image, (0, 0), fx=scale, fy=scale)
         elif min(test_image.shape[:2]) < 256:
             # resize for a minimum size of 256
             scale = 256 / min(test_image.shape[:2])
-            test_image = cv2.resize(test_image, (0, 0), fx=scale, fy=scale)
+        else:
+            scale = 1.0
+        new_shape = (int(test_image.shape[1] * scale), int(test_image.shape[0] * scale))
+        test_image = cv2.resize(test_image, new_shape)
+        test_image = test_image.astype(np.float16)
         test_image = torch.tensor(test_image).permute(2,0,1).unsqueeze(0)
         test_image = test_image / 127.5 - 1.0 
         test_image = test_image.to(device)