Skip to content
TensorFlow serving can handle a variable batch size when doing predictions. I never understood how to configure this and also the shape of the results returned. Finally figuring this out, here’s the changes to our previous serving setup to accept a variable number of images to classify for our model.
Serving input function First thing is to update our serving input receiver function placeholder. In the past we had set the placeholder to have a shape of , for variable batch size, this is as easy as setting it to [None].
Azure provides low priority VMs that we can use to save some money when utilising GPUs for machine learning. All the major cloud providers offer spare compute for a discount: Google cloud has Preemptible VMs, AWS has spot instances and Azure has low priority VMs. On Azure these low priority VMs can only be provisioned in Azure Batch or virtual machine scale sets and we can use the latter for hosting a jupyter environment.
Here we’ll look at exporting our previously trained dog and cat classifier and call that with local or remote files to test it out. To do this, I’ll use TensorFlow Serving in a docker container and use a python client to call to the remote host.
_Update 12th June, 2018: I used the gRPC interface here, but TensorFlow serving now has a REST API that could be beneficial or of more interest_
This notebook is available as a codelab
TensorFlow Hub was announced at TensorFlow Dev Summit 2018 and promises to reduce the effort required to use existing machine learning models and weights in your own custom model. From the overview page
TensorFlow Hub is a library to foster the publication, discovery, and consumption of reusable parts of machine learning models. A module is a self-contained piece of a TensorFlow graph, along with its weights and assets, that can be reused across different tasks in a process known as transfer learning.
Azure Batch AI provides us the PaaS opportunity to use GPU resources in the cloud. The basis is to use virtual machines in a managed cluster (i.e. you don’t have to maintain them) and run jobs as you see fit. For my use case, the opportunity of low-priority VMs to reduce the cost of using GPU machines is also particularly promising. What I’ll run through is running our first job on Azure Batch AI.
The TensorFlow canned estimators got promoted to core in version 1.3 to make training and evaluation of machine learning models very easy. This API allows you to describe your input data (categorical, numeric, embedding etc) through the use of feature columns. The estimator API also allows you to write a custom model for your unique job, and the feature columns capabilities can be utilised here as well to simplify or enhance things.
Understanding the shape of your model is sometimes non-trivial when it comes to machine learning. Look at convolutional neural nets with the number of filters, padding, kernel sizes etc and it’s quickly evident why understanding what shapes your inputs and outputs are will keep you sane and reduce the time spent digging into strange errors. TensorFlow’s RNN API exposed me to similar frustrations and misunderstandings about what I was expected to give it and what I was getting in return.
TFrecord files are TensorFlow’s suggested data format, although they are very difficult to inspect given their binary nature. Inspecting the contents of existing record files and ensuring the data in your input pipeline is as you expect is a good technique to have.
Inspecting TFRecord values The first trick is reading in the tfrecord files and inspecting their values in python. As you’d expect, the TensorFlow API allows this (although a little hidden down).
I’m a big advocate of the cloud and it’s ability to provide just enough resources ad hoc. You can use whatever you want, and pay for it just when using it. In machine learning there are services such as Google’s ML Engine or Azure’s upcoming Batch AI but during development, data preprocessing etc sometimes you want immediate iterative processes. In these cases, you can’t go past a Jupyter notebook and in this case, running that on a VM.
In 2017, there seems no doubt that if you aren’t running your ML training on a GPU you just aren’t doing things right ?. At home, my only computer is my MacBook Pro which is great to develop on, but would take an extremely long time to train something such as an image classification task. Saying this, I’d love to have a GPU machine at home, but I also love the opportunity to use this hardware in the cloud without having to power it, upgrade it and generally take care of something you actually own.
I’ve had a few attempts at getting TensorFlow estimators into a serving host and a client I can use to query them. Finally getting it working, I thought I’d write up the steps for reproduction.
Assumptions and Prerequisites The first assumption is that you have already trained your estimator (say the tf.estimator.DNNRegressor) and this is now in the variable estimator. You also have a list of feature columns as is standard in a variable feature_columns.
Embeddings can be used in machine learning to represent data and take advantage of reducing the dimensionality of the dataset and learning some latent factors between data points. Commonly this is used with words to say, reduce a 400,000 word vector to a 50 dimensional vector, but could equally be used to map post codes or other token encoded data. Another use case might be in recommender systems GloVe (Global Vectors for Word Representation) was developed at Stanford and more information can be found here.
An MNIST classifier is the go-to introduction for machine learning. Tensorflow is no different, and evolves to the Deep MNIST for Experts to include convolution, max pooling, dense layers and dropout: a good overview of ML layers for image problems. The downside of this is it doesn’t make use of Tensorflow’s new tf.estimator high level APIs. These provide all sorts of benefits for free than the usual sess.run TensorFlow tutorials you see online.
A common format for storing images and labels is a tree directory structure with the data directory containing a set of directories named by their label and each containing samples for said label. Often transfer learning that is used for image classification may provide data in this structure.
Update May 2018: If you would like an approach that doesn’t prepare into TFRecords, utilising tf.data and reading directly from disk, I have done this in when making the input function for my Dogs vs Cats transfer learning classifier.
TFRecords are TensorFlow’s native binary data format and is the recommended way to store your data for streaming data. Using the TFRecordReader is also a very convenient way to subsequently get these records into your model.
The data We will use the well known MNIST dataset for handwritten digit recognition as a sample. This is easily retrieved from tensorflow via:
from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets( "/tmp/tensorflow/mnist/input_data", reshape=False ) We then have mnist.
I recently completed Stanford Machine Learning on Coursera taught by Andrew Ng. It is a fantastic course for anyone interested, plus it’s free! The course uses MATLAB or Octave for the programming assignments. With the release of Azure Machine Learning, this seemed like a good exercise to reimplement these assignments on Azure ML to learn the tool.
Creating an Azure ML Workspace To get started, you will need an Azure subscription.