Remove base64 encoding about az-deep-realtime-score HOT 8 CLOSED

This looks more like the official way to do it https://towardsdatascience.com/how-to-deploy-machine-learning-models-with-tensorflow-part-1-make-your-model-ready-for-serving-776a14ec3198

from az-deep-realtime-score.

Funny Google advocates base64 encoding for their service https://cloud.google.com/vision/automl/docs/base64 :D

from az-deep-realtime-score.

I have seen this as well in a tutorial, definitely needs to be fixed.

from az-deep-realtime-score.

@fboylu Do you have a link? The tutorials in AML repo still use the old way https://github.com/Azure/MachineLearningNotebooks/blob/master/how-to-use-azureml/training-with-deep-learning/train-hyperparameter-tune-deploy-with-pytorch/pytorch_score.py

from az-deep-realtime-score.

Look here at the bottom where they test, there is no base64, did you not mean that?
https://github.com/Azure/MachineLearningNotebooks/blob/master/how-to-use-azureml/training-with-deep-learning/train-hyperparameter-tune-deploy-with-pytorch/train-hyperparameter-tune-deploy-with-pytorch.ipynb

from az-deep-realtime-score.

Thanks @fboylu. Looking at it it does this {'data': input_data.tolist()} which is worse than base64. It is the numpy array values like so [54, 14, 45 ....] which will be huge. The method we are talking about will preserve compression.

from az-deep-realtime-score.

I agree, I didn't like .tolist() either but interestingly tensorflow serving does the same... Look here https://www.tensorflow.org/serving/tutorials/Serving_REST_simple, the request part.

from az-deep-realtime-score.

Ha ok, with fashion mnist the overhead is negligable. I wonder what they do for object detection?

from az-deep-realtime-score.