Industrialise code by having it run on-demand on Google Cloud

[pvhee]

• Package up classifier into module that can run on Google Cloud ML
• Store network weights (save TF model somehow?) into the cloud
• Invoke classifier on-demand using Cloud Functions. Input should be image, output will be digit. Build front-end if needed.

Good reads

• https://medium.com/onfido-tech/higher-level-apis-in-tensorflow-67bfb602e6c0
• https://medium.com/tooso/serving-tensorflow-predictions-with-python-and-aws-lambda-facb4ab87ddd
• https://cloud.google.com/solutions/building-a-serverless-ml-model Export model to Google Cloud and invoke from a node.js Google Function
• https://github.com/GoogleCloudPlatform/training-data-analyst/blob/master/blogs/babyweight/babyweight.ipynb shows how to export a model, infer on Google Cloud ML, then build a python function to predict input using the model on Cloud ML (this could be wrapped up as a function)
• https://github.com/GoogleCloudPlatform/cloudml-samples/blob/master/census/keras/trainer/model.py shows how to run a Keras model on Google Cloud

[pvhee]

Also consider AWS, see https://aws.amazon.com/about-aws/whats-new/2017/10/aws-deep-learning-ami-now-supports-pytorch-keras-2-and-latest-deep-learning-frameworks/

--> we can probably use Sagemaker https://aws.amazon.com/sagemaker/ so no need to use gcloud

[pvhee]

For Google, we can turn Keras model into a TensorFlow estimator, then serve predictions using TensorFlow Serving as per https://cloud.google.com/blog/big-data/2017/12/new-in-tensorflow-14-converting-a-keras-model-to-a-tensorflow-estimator

[pvhee]

Having issues with exporting our model (into a directory) as our input receiver function is incorrect. To help debug this, it is useful to inspect the SavedModel using

saved_model_cli show --dir export/1521414113/ --tag_set serve --signature_def serving_default

which returns

The given SavedModel SignatureDef contains the following input(s):
inputs['examples'] tensor_info:
    dtype: DT_STRING
    shape: (-1)
    name: input_example_tensor:0
The given SavedModel SignatureDef contains the following output(s):
outputs['labels'] tensor_info:
    dtype: DT_FLOAT
    shape: (-1, 6, 11)
    name: labels/truediv:0
Method name is: tensorflow/serving/predict

[pvhee]

Options around sending image data as input: https://stackoverflow.com/questions/46216095/using-gcloud-ml-serving-for-large-images/46222990#46222990

[pvhee]

Convert an image (daisy.jpg) into b64 encoded JSON string

python -c 'import base64, sys, json; img = base64.b64encode(open(sys.argv[1], "rb").read()); print json.dumps({"key":"0", "image_bytes": {"b64": img}})' daisy.jpg

[pvhee]

Latest update: we can generate predictions using Cloud ML Prediction API via https://github.com/pvhee/deeplearning_experiments/blob/master/svhn/predict_gcloud.py. This returns for example

python -m svhn.predict_gcloud
[ 2  3  4 10 10 10]
[ 3  2  1  5 10 10]
[ 4  1  0  3 10 10]
[ 3  1  1  0 10 10]
[ 2  1  0 10 10 10]

✌️

Next up are:

• [ ] update https://github.com/pvhee/digit-recognizer-serverless to receive base64-encoded images, resize, then upload as JSON to CloudML
• [ ] update https://github.com/pvhee/digit-recognizer-app to base64-encode images, pass on the function above and display the result in the front-end

Almost there... 👍