Fix RKNN data layout

onnx/rknnrun.py

@@ -46,7 +46,7 @@ for k, v in inputs.items():
 
 # 4. run vision encoder using RKNN
 start_time = time.time()
-image_features = rknn_vision_encoder.inference(inputs=[inputs["pixel_values"]])[0]
+image_features = rknn_vision_encoder.inference(inputs=[inputs["pixel_values"]], data_format='nchw')[0]
 end_time = time.time()
 vision_encoder_time = (end_time - start_time) * 1000
 total_time += vision_encoder_time