in_tensor_dtype use np or tf types

[FrancescoSaverioZuppichini]

It would be nice to use directly np or tf types to define in_tensor_dtype. eg:

{'in_tensor_name': 'inputs', 'in_tensor_dtype': np.float32, 'data': np.random.rand(10,2)}

Also, what about make float the default type? So if somebody is lazy it can just not pass in_tensor_dtype?

Btw, awesome work! You save my day :)

Cheers,

Francesco Saverio

[stianlp]

Using np or tf types
So I am not sure how (or even if it's possible) to map all types listed in types.proto (https://github.com/epigramai/tfserving-python-predict-client/blob/master/protos/types.proto) to np types.

For instance, what np type would you pass if your model input is TF_DOUBLE? And I don't see any type string in numpy, so here the Python type str would probably be best to use?

There are also types like DT_HALF, DT_RESOURCE that are not in numpy, so I am not sure how these should be mapped.

Numpy and tf data types are listed here:
https://docs.scipy.org/doc/numpy-1.13.0/user/basics.types.html
https://www.tensorflow.org/versions/r1.1/programmers_guide/dims_types

I don't want to use tf types (like tf.float32) because I don't want to pull TensorFlow into the client. What I like most about this client is that it can be used by anyone writing Python without having to install TensorFlow.

Use float as default dtype

Yes, that would be convenient. Perhaps the default type also can be set in the client's constructor? So if type is not specified in request_data's objects, check the client's default data type wich defaults to 'DT_FLOAT', but can be overridden by the developer with PredictClient(..., default_data_type='DT_FLOAT').

Feel free to do the change yourself and make a PR. I might find time during this or next week.

[FrancescoSaverioZuppichini]

Dear @stiansel ,

Thank you for your very quick response. I agree on your point on types, what I think it can be nice to scan the input and see if there is a numpy array and directly map it to the correct time since it may be not obvious for the user to understand where the types come from.

I don't want to use tf types (like tf.float32) because I don't want to pull TensorFlow into the client. What I like most about this client is that it can be used by anyone writing Python without having to install TensorFlow.

Totally agree.

Feel free to do the change yourself and make a PR. I might find time during this or next week.

I am a little busy in these days, I will do it in the weekend

Thank you again for the package, you saved my day!

All the best,

Francesco Saverio