bartvelp / edge_tpu_mnist Goto Github PK

Hi, thanks for sharing this example. I was wondering if I am missing something regarding the normalisation. When you train the model you convert the MNIST from uint8 to float by dividing for 255. However, in the inference step you use the input to the model as uint8 (which I understand being due EdgeTPU on Coral constrains). However, it seems the model that was trained using floats is being feed now with uint8. How does the model become aware of this? It is during the post-quantization step that in some way a model that was trained using floats between 0 and 1 will be ready to receive uint8?

Thanks