Using customized tensor name for prediction after deploying the model

[vbvg2008]

Please fill out the form below.

System Information

• Framework (e.g. TensorFlow) / Algorithm (e.g. KMeans): TensorFlow
• Framework Version: 1.10
• Python Version: 2.7
• CPU or GPU: GPU
• Are you using a custom image: No

Describe the problem

All of the Sagemaker tensorflow tutorial uses "inputs" as the input name for prediction. for example, in "tensorflow_resnet_cifar10_with_tensorboard" example, the prediction is simply done by

random_image = np.random.rand(32, 32, 3)
predictor.predict(random_image_data)

but what if I trained a tensorflow model with a customized input name? say my model inputs only accepts a dictionary like this:
inputs = {"input_image": tf.placeholder(tf.float32, [None, 32, 32, 3])}

Although I successfully trained and deployed the model, but when I try to do the prediction such as:

random_image = np.random.rand(32, 32, 3)
predictor.predict({'input_image':random_image})

it gives me error saying my data is not JSON serializable. How do I use a customized name as input name for prediction?

[mvsusp]

Hello @vbvg2008 ,

I believe that this issue is related to aws/sagemaker-tensorflow-training-toolkit#71.
We are in the process to release a newer container with these codefixes. I will update you in this ticket when it happens.

Thanks for using SageMaker!

[winstonaws]

@mvsusp The problem is fixed on the container side, but now we need to make a fix to the Python SDK to serialize numpy arrays in dicts correctly.

@vbvg2008 in the meantime, if you want to unblock yourself, try the workaround described here:
aws/sagemaker-tensorflow-training-toolkit#71 (comment)

Basically, convert your numpy arrays into lists before putting them in the dict.

[tlelson]

hi @winstonaws,
I am running into an issue that seems like it may be related. I have logged it under issue #413.
If you have any idea on how to coerce my input data into an acceptable format I would very much appreciate it.

[icywang86rui]

hi @eL0ck,

We have merged in a change to accept dicts with numpy arrays on the python sdk side - #404 .
Your original way to do the prediction should work now:

random_image = np.random.rand(32, 32, 3)
predictor.predict({'input_image':random_image})

Please let us know if the problems persists.