Because during conversion you quantized the model from float32 to int8. Check my [part=3](https://medium.com/analytics-vidhya/facenet-on-modile-part-3-cc6f6d5752d6) to understand it more deeply.