Variational Autoencoders with Tensorflow Probability Layers

Variational Autoencoders with Tensorflow Probability Layers

In this post, we demonstrated how to combine deep learning with probabilistic programming: we built a variational autoencoder that used TFP Layers to pass the output of a Keras Sequential model to a probability distribution in TFP.

At the 2019 TensorFlow Developer Summit, we announced TensorFlow Probability (TFP) Layers. In that presentation, we showed how to build a powerful regression model in very few lines of code. Here, we will show how easy it is to make a Variational Autoencoder (VAE) using TFP Layers.

TensorFlow Probability Layers

TFP Layers provides a high-level API for composing distributions with deep networks using Keras. This API makes it easy to build models that combine deep learning and probabilistic programming. For example, we can parameterize a probability distribution with the output of a deep network. We will use this approach here.

Variational Autoencoders and the ELBO

Variational Autoencoders (VAEs) are popular generative models being used in many different domains, including collaborative filtering, image compression, reinforcement learning, and generation of music and sketches.

In the traditional derivation of a VAE, we imagine some process that generates the data, such as a latent variable generative model. Consider the process of drawing digits, as in MNIST. Suppose that before you draw the digit, you first decide which digit you will draw, imagining some fuzzy picture in your head. Then, you put pen to paper and try to create the picture in the real world. We can formalize this two step process:

  1. You sample some latent representation z from some prior distribution z ~ p(z). This is the fuzzy picture in your head — let’s say of a “3”.
  2. Based on your sample, you draw the actual picture representation x, modeled itself as a stochastic process x ~ p(x|z). This captures the idea that each time you write a “3”, it looks at least a little different.

Thus, when a handwritten digit is created, we imagine some of the variation is due to some kind of signal inherent to the process, such as the class identity of the MNIST digit, and some of that variation is due to noise, such as differences in the precise angles of the lines among different samples of the same digit drawn by the same person. Broadly, a VAE is an attempt to try to separate the signal from the noise with an explicit model of both processes.

To train this objective, we maximize the ELBO (Evidence Lower BOund) objective:

Where the three probability density functions are:

  • p(z), the prior on the latent representation z,
  • q(z|x), the variational encoder, and
  • p(x|z), the decoder — how likely is the image x given the latent representation z.

The ELBO is a lower bound on log p(x), the log probability of an observed data point. The first integral in the ELBO equation is the reconstruction term. It asks how likely we are to start at an image x, encode it to z, decode it, and get back the original x. The second term is the KL divergence term. It measures how close together our encoder and prior are; you can think of this term as just trying to keep our encoder honest. If our encoder generates z samples that are too unlikely given our prior, the objective is worse than if it generates z samples more typical of the prior. Thus, the encoder should differ from the prior only if the cost of doing so is outweighed by the benefit in the reconstruction term.

Code

From the description above, we can see that it is natural to model three different components individually: the prior p(z), the variational encoder q(z|x), and the decoder p(x|z). You can follow along by running this colab, which trains a VAE on MNIST in a few minutes on a cloud GPU.

Prior

The simplest prior typically used in a VAE is an isotropic Gaussian:

tfd = tfp.distributions
encoded_size = 16
prior = tfd.Independent(tfd.Normal(loc=tf.zeros(encoded_size), scale=1),
reinterpreted_batch_ndims=1)

Here, we have just created a TFP independent Gaussian distribution with no learned parameters, and we have specified that our latent variable, z, will have 16 dimensions.

Encoder

For our encoder distribution, we will use a full-covariance Gaussian distribution, with its mean and covariance matrices parameterized by the output of a neural network. This may sound complicated, but it is very easy to express with TFP Layers:

tfpl = tfp.layers
encoder = tfk.Sequential([
    tfkl.InputLayer(input_shape=input_shape),
    tfkl.Lambda(lambda x: tf.cast(x, tf.float32) - 0.5),
    tfkl.Conv2D(base_depth, 5, strides=1,
                padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2D(base_depth, 5, strides=2,
                padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2D(2 * base_depth, 5, strides=1,
                padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2D(2 * base_depth, 5, strides=2,
                padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2D(4 * encoded_size, 7, strides=1,
                padding='valid', activation=tf.nn.leaky_relu),
    tfkl.Flatten(),
    tfkl.Dense(tfpl.MultivariateNormalTriL.params_size(encoded_size),
               activation=None),
    tfpl.MultivariateNormalTriL(
        encoded_size,
        activity_regularizer=tfpl.KLDivergenceRegularizer(prior, weight=1.0)),
])

The encoder is just a normal Keras Sequential model, consisting of convolutions and dense layers, but the output is passed to a TFP Layer, MultivariateNormalTril(), which transparently splits the activations from the final Dense() layer into the parts needed to specify both the mean and the (lower triangular) covariance matrix, the parameters of a Multivariate Normal. We used a helper, tfpl.MultivariateNormalTriL.params_size(encoded_size), to make the Dense() layer output the correct number of activations (i.e., the distribution’s parameters). Finally, we said that the distribution should contribute a “regularization” term to the final loss. Specifically, we are adding the KL divergence between the encoder and the prior to the loss, which is the KL term in the ELBO that we described above. (Fun fact: we can turn this VAE into a β-VAE simply by changing the weight argument to something other than 1!)

Decoder

For our decoder, we will use a simple “mean-field decoder”, which in this case will be a pixel-independent Bernoulli distribution:

decoder = tfk.Sequential([
    tfkl.InputLayer(input_shape=[encoded_size]),
    tfkl.Reshape([1, 1, encoded_size]),
    tfkl.Conv2DTranspose(2 * base_depth, 7, strides=1,
                         padding='valid', activation=tf.nn.leaky_relu),
    tfkl.Conv2DTranspose(2 * base_depth, 5, strides=1,
                         padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2DTranspose(2 * base_depth, 5, strides=2,
                         padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2DTranspose(base_depth, 5, strides=1,
                         padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2DTranspose(base_depth, 5, strides=2,
                         padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2DTranspose(base_depth, 5, strides=1,
                         padding='same', activation=tf.nn.leaky_relu),
    tfkl.Conv2D(filters=1, kernel_size=5, strides=1,
                padding='same', activation=None),
    tfkl.Flatten(),
    tfpl.IndependentBernoulli(input_shape, tfd.Bernoulli.logits),
])

The form here is essentially the same as the encoder, but now we are using transposed convolutions to take our latent representation, which is a 16 dimensional vector, and turn it into a 28 x 28 x 1 tensor. That final tensor parameterizes the pixel-independent Bernoulli distribution.

Loss

We are now ready to build the full model and specify the rest of the loss function.

vae = tfk.Model(inputs=encoder.inputs,
outputs=decoder(encoder.outputs[0]))

Our model is just a Keras Model where the outputs are defined as the composition of the encoder and the decoder. Since the encoder already added the KL term to the loss, we need to specify only the reconstruction loss (the first term of the ELBO above).

negative_log_likelihood = lambda x, rv_x: -rv_x.log_prob(x)

vae.compile(optimizer=tf.optimizers.Adam(learning_rate=1e-3),
loss=negative_log_likelihood)

The loss function takes two arguments — the original input, x, and the output of the model. We call that _rv_x_ because it is a random variable. This example demonstrates some of the core magic of TFP Layers — even though Keras and Tensorflow view the TFP Layers as outputting tensors, TFP Layers are actually Distribution objects. Thus, we can make our loss function be the negative log likelihood of the data given the model: -rv_x.log_prob(x).

Wait, what?

It is worth taking a moment to understand what TFP Layers is actually doing to integrate transparently with Keras. As we said, the output of a TFP Layer is actually a Distribution object. We can check that with the following code:

x = eval_dataset.make_one_shot_iterator().get_next()[0][:10]
xhat = vae(x)
assert isinstance(xhat, tfd.Distribution)

But if a TFP Layer returns a Distribution, what happens when we compose the decoder with the output of the encoder: decoder_model(encoder_model.outputs[0]))? Well, in order for Keras to view the encoder distribution as a Tensor, TFP Layers actually “reifies” the distribution as a sample from that distribution, which is just a fancy way of saying that Keras sees the Distribution object as the Tensor we would have gotten, had we called encoder_model.sample(). But, when we need to access the Distribution object directly, we can — just like we do in the loss function when we call rv_x.log_prob(x). TFP Layers provides the distribution-like and Tensor-like behaviors automatically, so you don’t need to worry about Keras getting confused.

Training

Training the model is as easy as training any Keras model: we just call vae_model.fit():

vae.fit(train_dataset,
        epochs=15,
validation_data=eval_dataset)

With this model, we are able to get an ELBO of around 115 nats (the nat is the natural logarithm equivalent of the bit — 115 nats is around 165 bits). Of course, this performance isn’t state-of-the-art, but it is easy to make any of the three components more powerful starting from this basic setup. Also, it already generates nice looking digits!

Decoder modes generated by encoding images from the MNIST test set.

Decoder modes generated by sampling from the prior.

Summary

In this post, we demonstrated how to combine deep learning with probabilistic programming: we built a variational autoencoder that used TFP Layers to pass the output of a Keras Sequential model to a probability distribution in TFP.

We utilized the tensor-like and distribution-like semantics of TFP layers to make our code relatively straightforward.

Machine Learning Tutorial with Python, Jupyter, KSQL and TensorFlow

Machine Learning Tutorial with Python, Jupyter, KSQL and TensorFlow

Machine Learning With Python, Jupyter, KSQL, and TensorFlow. This post focuses on how the Kafka ecosystem can help solve the impedance mismatch between data scientists, data engineers and production engineers.

Machine Learning With Python, Jupyter, KSQL, and TensorFlow. This post focuses on how the Kafka ecosystem can help solve the impedance mismatch between data scientists, data engineers and production engineers.

Building a scalable, reliable, and performant machine learning (ML) infrastructure is not easy. It takes much more effort than just building an analytic model with Python and your favorite machine learning framework.

Uber, which already runs their scalable and framework-independent machine learning platform Michelangelo for many use cases in production, wrote a good summary:

When Michelangelo started, the most urgent and highest impact use cases were some very high scale problems, which led us to build around Apache Spark (for large-scale data processing and model training) and Java (for low latency, high throughput online serving). This structure worked well for production training and deployment of many models but left a lot to be desired in terms of overhead, flexibility, and ease of use, especially during early prototyping and experimentation [where Notebooks and Python shine].
Uber expanded Michelangelo “to serve any kind of Python model from any source to support other Machine Learning and Deep Learning frameworks like PyTorch and TensorFlow [instead of just using Spark for everything].”

So why did Uber (and many other tech companies) build its own platform and framework-independent machine learning infrastructure?

The posts How to Build and Deploy Scalable Machine Learning in Production with Apache Kafka and Using Apache Kafka to Drive Cutting-Edge Machine Learning describe the benefits of leveraging the Apache Kafka ® ecosystem as a central, scalable, and mission-critical nervous system. It allows real-time data ingestion, processing, model deployment, and monitoring in a reliable and scalable way.

This post focuses on how the Kafka ecosystem can help solve the impedance mismatch between data scientists, data engineers, and production engineers. By leveraging it to build your own scalable machine learning infrastructure and also make your data scientists happy, you can solve the same problems for which Uber built its own ML platform, Michelangelo.


You may also like:A Complete Machine Learning Project Walk-Through in Python


Impedance Mismatch Between Data Scientists, Data Engineers and Production Engineers

Based on what I’ve seen in the field, an impedance mismatch between data scientists, data engineers, and production engineers is the main reason why companies struggle to bring analytic models into production to add business value.

The following diagram illustrates the different required steps and corresponding roles as part of the impedance mismatch in a machine learning lifecycle:

Impedance mismatch between model development and model deployment

Data scientists love Python, period. Therefore, the majority of machine learning/deep learning frameworks focus on Python APIs. Both the stablest and most cutting edge APIs, as well as the majority of examples and tutorials, use Python APIs. In addition to Python support, there is typically support for other programming languages, including JavaScript for web integration and Java for platform integration-though oftentimes with fewer features and less maturity. No matter what other platforms are supported, chances are very high that your data scientists will build and train their analytic models with Python.

There is an impedance mismatch between model development using Python, its tool stack and a scalable, reliable data platform with low latency, high throughput, zero data loss and 24/7 availability requirements needed for data ingestion, preprocessing, model deployment and monitoring at scale. Python, in practice, is not the most well-known technology for these requirements. However, it is a great client for a data platform like Apache Kafka.

The problem is that writing the machine learning source code to train an analytic model with Python and the machine learning framework of your choice is just a very small part of a real-world machine learning infrastructure. You need to think about the whole model lifecycle. The following image represents this hidden technical debt in machine learning systems (showing how small the “ML code” part is):

Thus, you need to train and deploy the model built to a scalable production environment in order to reliably make use of it. This can either be built natively around the Kafka ecosystem, or you could use Kafka just for ingestion into another storage and processing cluster such as HDFS or AWS S3 with Spark. There are many tradeoffs between Kafka, Spark, and several other scalable infrastructures, but that discussion is out of scope for this post. For now, we’ll focus on Kafka.

Different solutions in the industry solve certain parts of the impedance mismatch between data scientists, data engineers, and production engineers. Let’s take a look at some of these options:

  • Official standards like Open Neural Network Exchange (ONNX), Portable Format for Analytics (PFA) or Predictive Model Markup Language (PMML): A data scientist builds a model with Python. The Java developer imports it in Java for production deployment. This approach supports different frameworks, products, and cloud services. You do not have to rely on the same framework or product for training and model deployment. Consider ONNX, a relatively new standard for deep learning — it already supports TensorFlow, PyTorch, and MXNet. These standards have pros and cons. Some people like and use them; many don’t.
  • Developer-focused frameworks like Deeplearning4j: These frameworks are built for software engineers to build the whole machine learning lifecycle on the Java platform, not just model deployment and monitoring, but also preprocessing and training. You can still import other models if you want (e.g., Deeplearning4j lets you import Keras models). This option is great if you: a) have data scientists who can write Java or b) have software engineers who understand machine learning concepts enough to build analytic models.
  • AutoML for building analytic models with limited machine learning experience: This way, domain experts can build and deploy analytic models with a button click. The AutoML engine provides an interface for others to use the model for predictions.
  • Embedding model binaries into applications: The output of model training is an analytic model. For instance, you can write Python code to train and generate a TensorFlow model. Depending on the framework, the output can be text files, Java source code, or binary files. For example, TensorFlow generates a model artifact with Protobuf, JSON, and other files. No matter what format the output of your machine learning framework is, it can be embedded into applications to use for predictions via the framework’s API (e.g., you can load a TensorFlow model from a Java application through TensorFlow’s Java API).
  • Managed model server in the public cloud like Google Cloud Machine Learning Engine: The cloud provider takes over the burden of availability and reliability. The data scientist “just” deploys its trained model, and production engineers can access it. The key tradeoff is that this requires RPC communication to perform model inference.

While all these solutions help data scientists, data engineers, and production engineers to work better together, there are underlying challenges within the hidden debts:

  • Data collection (i.e., integration) and preprocessing need to run at scale

  • Configuration needs to be shared and automated for continuous builds and integration tests

  • The serving and monitoring infrastructure need to fit into your overall enterprise architecture and tool stack

So how can the Kafka ecosystem help here?

Apache Kafka as a Key Component for Solving the Impedance Mismatch

In many cases, it is best to provide experts with the tools they like and know well. The challenge is to combine the different toolsets and still build an integrated system, as well as a continuous, scalable, machine learning workflow. Therefore, Kafka is not competitive but complementary to the discussed alternatives when it comes to solving the impedance mismatch between the data scientist and developer.

The data engineer builds a scalable integration pipeline using Kafka as infrastructure and Python for integration and preprocessing statements. The data scientist can build their model with Python or any other preferred tool. The production engineer gets the analytic models (either manually or through any automated, continuous integration setup) from the data scientist and embeds them into their Kafka application to deploy it in production. Or, the team works together and builds everything with Java and a framework like Deeplearning4j.

Any option can pair well with Apache Kafka. Pick the pieces you need, whether it’s Kafka core for data transportation, Kafka Connect for data integration, or Kafka Streams/KSQL for data preprocessing. Many components can be used for both model training and model inference. Write once and use in both scenarios as shown in the following diagram:

Leveraging the Apache Kafka ecosystem for a machine learning infrastructure

Monitoring the complete environment in real time and at scale is also a common task for Kafka. A huge benefit is that you only build a highly reliable and scalable pipeline once but use it for both parts of a machine learning infrastructure. And you can use it in any environment: in the cloud, in on-prem datacenters, or at the edges where IoT devices are.

Say you wanted to build one integration pipeline from MQTT to Kafka with KSQL for data preprocessing and use Kafka Connect for data ingestion into HDFS, AWS S3, or Google Cloud Storage, where you do the model training. The same integration pipeline, or at least parts of it, can be reused for model inference. New MQTT input data can directly be used in real time to make predictions.

We just explained various alternatives to solving the impedance mismatch between data scientists and software engineers in Kafka environments. Now, let’s discuss one specific option in the next section, which is probably the most convenient for data scientists: leveraging Kafka from a Jupyter Notebook with KSQL statements and combining it with TensorFlow and Keras to train a neural network.

Data Scientists Combining Python and Jupyter With Scalable Streaming Architectures

Data scientists use tools like Jupyter Notebooks to analyze, transform, enrich, filter, and process data. The preprocessed data is then used to train analytic models with machine learning/deep learning frameworks like TensorFlow.

However, some data scientists do not even know “bread-and-butter” concepts of software engineers, such as version control systems like GitHub or continuous integration tools like Jenkins.

This raises the question of how to combine the Python experience of data scientists with the benefits of Apache Kafka as a battle-tested, highly scalable data processing and streaming platform.

Apache Kafka and KSQL for Data Scientists and Data Engineers

Kafka offers integration options that can be used with Python, like Confluent’s Python Client for Apache Kafka or Confluent REST Proxy for HTTP integration. But this is not really a convenient way for data scientists who are used to quickly and interactively analyzing and preprocessing data before model training and evaluation. Rapid prototyping is typically used here.

KSQL enables data scientists to take a look at Kafka event streams and implement continuous stream processing from their well-known and loved Python environments like Jupyter by writing simple SQL-like statements for interactive analysis and data preprocessing.

The following Python example executes an interactive query from a Kafka stream leveraging the open source framework ksql-python, which adds a Python layer on top of KSQL’s REST interface. Here are a few lines of the Python code using KSQL from a Jupyter Notebook:

The result of such a KSQL query is a Python generator object, which you can easily process with other Python libraries. This feels much more Python native and is analogous to NumPy, pandas, scikit-learn and other widespread Python libraries.

Similarly to rapid prototyping with these libraries, you can do interactive queries and data preprocessing with ksql-python. Check out the KSQL quick start and KSQL recipes to understand how to write a KSQL query to easily filter, transform, enrich, or aggregate data. While KSQL is running continuous queries, you can also use it for interactive analysis and use the LIMIT keyword like in ANSI SQL if you just want to get a specific number of rows.

So what’s the big deal? You understand that KSQL can feel Python-native with the ksql-python library, but why use KSQL instead of or in addition to your well-known and favorite Python libraries for analyzing and processing data?

The key difference is that these KSQL queries can also be deployed in production afterwards. KSQL offers you all the features from Kafka under the hood like high scalability, reliability, and failover handling. The same KSQL statement that you use in your Jupyter Notebook for interactive analysis and preprocessing can scale to millions of messages per second. Fault tolerant. With zero data loss and exactly once semantics. This is very important and valuable for bringing together the Python-loving data scientist with the highly scalable and reliable production infrastructure.

Just to be clear: KSQL + Python is not the all-rounder for every data engineering task, and it does not replace the existing Python toolset. But it is a great option in the toolbox of data scientists and data engineers, and it adds new possibilities like getting real-time updates of incoming information as the source data changes or updating a deployed model with a new and improved version.

Jupyter Notebook for Fraud Detection With Python KSQL and TensorFlow/Keras

Let’s now take a look at a detailed example using the combination of KSQL and Python. It involves advanced code examples using ksql-python and other widespread components from Python’s machine learning ecosystem, like NumPy, pandas, TensorFlow, and Keras.

The use case is fraud detection for credit card payments. We use a test dataset from Kaggle as a foundation to train an unsupervised autoencoder to detect anomalies and potential fraud in payments. The focus of this example is not just model training, but the whole machine learning infrastructure, including data ingestion, data preprocessing, model training, model deployment, and monitoring. All of this needs to be scalable, reliable, and performant.

For the full running example and more details, see the documentation.

Let’s take a look at a few snippets of the Jupyter Notebook.

Connection to KSQL server and creation of a KSQL stream using Python:

from ksql import KSQLAPI
client = KSQLAPI('http://localhost:8088')

client.create_stream(table_name='creditcardfraud_source',
                     columns_type=['Id bigint', 'Timestamp varchar', 'User varchar', 'Time int', 'V1 double', 'V2 double', 'V3 double', 'V4 double', 'V5 double', 'V6 double', 'V7 double', 'V8 double', 'V9 double', 'V10 double', 'V11 double', 'V12 double', 'V13 double', 'V14 double', 'V15 double', 'V16 double', 'V17 double', 'V18 double', 'V19 double', 'V20 double', 'V21 double', 'V22 double', 'V23 double', 'V24 double', 'V25 double', 'V26 double', 'V27 double', 'V28 double', 'Amount double', 'Class string'],
                     topic='creditcardfraud_source',
                     value_format='DELIMITED')

Preprocessing incoming payment information using Python:

  • Filter columns that are not needed

  • Filter messages where column "class" is empty

  • Change data format to Avro for convenient and further processing

client.create_stream_as(table_name='creditcardfraud_preprocessed_avro',
                     select_columns=['Time', 'V1', 'V2', 'V3', 'V4', 'V5', 'V6', 'V7', 'V8', 'V9', 'V10', 'V11', 'V12', 'V13', 'V14', 'V15', 'V16', 'V17', 'V18', 'V19', 'V20', 'V21', 'V22', 'V23', 'V24', 'V25', 'V26', 'V27', 'V28', 'Amount', 'Class'],
                     src_table='creditcardfraud_source',
                     conditions='Class IS NOT NULL',
                     kafka_topic='creditcardfraud_preprocessed_avro',
                     value_format='AVRO')

Some more examples for possible data wrangling and preprocessing with KSQL:

  • Drop columns, filter messages where value “class” is empty and change data format to Avro:
CREATE STREAM creditcardfraud_preprocessed_avro WITH (VALUE_FORMAT='AVRO', KAFKA_TOPIC='creditcardfraud_preprocessed_avro') AS SELECT Time,  V1 , V2 , V3 , V4 , V5 , V6 , V7 , V8 , V9 , V10 , V11 , V12 , V13 , V14 , V15 , V16 , V17 , V18 , V19 , V20 , V21 , V22 , V23 , V24 , V25 , V26 , V27 , V28 , Amount , Class FROM creditcardfraud_source WHERE Class IS NOT NULL;

  • Anonymization (mask the two leftmost characters, e.g., “Hans” becomes “**ns”):
SELECT Id, MASK_LEFT(User, 2) FROM creditcardfraud_source;

  • Augmentation (add -1 if “class” is null):
SELECT Id, IFNULL(Class, -1) FROM creditcardfraud_source;

  • Merge/join data frames:
CREATE STREAM creditcardfraud_per_user WITH (VALUE_FORMAT='AVRO', KAFKA_TOPIC='creditcardfraud_preprocessed_avro') AS SELECT Time,  V1 , V2 , V3 , V4 , V5 , V6 , V7 , V8 , V9 , V10 , V11 , V12 , V13 , V14 , V15 , V16 , V17 , V18 , V19 , V20 , V21 , V22 , V23 , V24 , V25 , V26 , V27 , V28 , Amount , Class FROM creditcardfraud_enahnced c INNER JOIN USERS u on c.userid = u.userid WHERE V1 > 5 AND V2 IS NOT NULL AND u.CITY LIKE 'Premium%';

The Jupyter Notebook contains the full example. We use Python + KSQL for integration, data preprocessing, and interactive analysis and combine them with various other libraries from a common Python machine learning tool stack for prototyping and model training:

  • Arrays/matrices processing with NumPy and pandas

  • ML-specific processing (split train/test, etc.) with scikit-learn

  • Interactive analysis through data visualisations with Matplotlib

  • ML training + evaluation with TensorFlow and Keras

Model inference and visualisation are done in the Jupyter notebook, too. After you have built an accurate model, you can deploy it anywhere to make predictions and leverage the same integration pipeline for model training. Some examples of model deployment in Kafka environments are:

  • Analytic models (TensorFlow, Keras, H2O and Deeplearning4j) embedded in Kafka Streams microservices

  • Anomaly detection of IoT sensor data with a model embedded into a KSQL UDF

  • RPC communication between Kafka Streams application and model server (TensorFlow Serving)

Python, KSQL, and Jupyter for Prototyping, Demos, and Production Deployments

As you can see, both in theory (Google’s paper Hidden Technical Debt in Machine Learning Systems) and in practice (Uber’s machine learning platform Michelangelo), it is not a simple task to build a scalable, reliable, and performant machine learning infrastructure.

The impedance mismatch between data scientists, data engineers, and production engineers must be resolved in order for machine learning projects to deliver real business value. This requires using the right tool for the job and understanding how to combine them. You can use Python and Jupyter for prototyping and demos (often Kafka and KSQL might be overhead here and not needed if you just want to do fast, simple prototyping on a historical dataset) or combine Python and Jupyter with your whole development lifecycle up to production deployments at scale.

Integration of Kafka event streams and KSQL statements into Jupyter Notebooks allows you to:

  • Use the preferred existing environment of the data scientist (including Python and Jupyter) and combine it with Kafka and KSQL to integrate and continuously process real-time streaming data by using a simple Python wrapper API to execute KSQL queries

  • Easily connect to real-time streaming data instead of just historical batches of data (maybe from the last day, week or month, e.g., coming in via CSV files)

  • Merge different concepts like streaming event-based sensor data coming from Kafka with Python programming concepts like generators or dictionaries objects, which you can use for your Python data processing tools or ML frameworks like NumPy, pandas, or scikit-learn

  • Reuse the same logic for integration, preprocessing, and monitoring and move it from your Jupyter Notebook and prototyping or demos to large-scale test and production systems

Python for prototyping and Apache Kafka for a scalable streaming platform are not rival technology stacks. They work together very well, especially if you use “helper tools” like Jupyter Notebooks and KSQL.

Please try it out and let us know your thoughts. How do you leverage the Apache Kafka ecosystem in your machine learning projects?

TensorFlow Vs PyTorch: Comparison of the Machine Learning Libraries

TensorFlow Vs PyTorch: Comparison of the Machine Learning Libraries

Libraries play an important role when developers decide to work in Machine Learning or Deep Learning researches. In this article, we list down 10 comparisons between TensorFlow and PyTorch these two Machine Learning Libraries.

According to this article, a survey based on a sample of 1,616 ML developers and data scientists, for every one developer using PyTorch, there are 3.4 developers using TensorFlow. In this article, we list down 10 comparisons between these two Machine Learning Libraries

1 - Origin

PyTorch has been developed by Facebook which is based on Torch while TensorFlow, an open sourced Machine Learning Library, developed by Google Brain is based on the idea of data flow graphs for building models.

2 - Features

TensorFlow has some attracting features such as TensorBoard which serves as a great option while visualising a Machine Learning model, it also has TensorFlow Serving which is a specific grpc server that is used during the deployment of models in production. On the other hand, PyTorch has several distinguished features too such as dynamic computation graphs, naive support for Python, support for CUDA which ensures less time for running the code and increase in performance.

3 - Community

TensorFlow is adopted by many researchers of various fields like academics, business organisations, etc. It has a much bigger community than PyTorch which implies that it is easier to find for resources or solutions in TensorFlow. There is a vast amount of tutorials, codes, as well as support in TensorFlow and PyTorch, being the newcomer into play as compared to TensorFlow, it lacks these benefits.

4 - Visualisation

Visualisation plays as a protagonist while presenting any project in an organisation. TensorFlow has TensorBoard for visualising Machine Learning models which helps during training the model and spot the errors quickly. It is a real-time representation of the graphs of a model which not only depicts the graphic representation but also shows the accuracy graphs in real-time. This eye-catching feature is lacked by PyTorch.

5 - Defining Computational Graphs

In TensorFlow, defining computational graph is a lengthy process as you have to build and run the computations within sessions. Also, you will have to use other parameters such as placeholders, variable scoping, etc. On the other hand, Python wins this point as it has the dynamic computation graphs which help id building the graphs dynamically. Here, the graph is built at every point of execution and you can manipulate the graph at run-time.

6 - Debugging

PyTorch being the dynamic computational process, the debugging process is a painless method. You can easily use Python debugging tools like pdb or ipdb, etc. for instance, you can put “pdb.set_trace()” at any line of code and then proceed for executions of further computations, pinpoint the cause of the errors, etc. While, for TensorFlow you have to use the TensorFlow debugger tool, tfdbg which lets you view the internal structure and states of running TensorFlow graphs during training and inference.

7 - Deployment

For now, deployment in TensorFlow is much more supportive as compared to PyTorch. It has the advantage of TensorFlow Serving which is a flexible, high-performance serving system for deploying Machine Learning models, designed for production environments. However, in PyTorch, you can use the Microframework for Python, Flask for deployment of models.

8 - Documentation

The documentation of both frameworks is broadly available as there are examples and tutorials in abundance for both the libraries. You can say, it is a tie between both the frameworks.

Click here for TensorFlow documentation and click here for PyTorch documentation.

9 - Serialisation

The serialisation in TensorFlow can be said as one of the advantages for this framework users. Here, you can save your entire graph as a protocol buffer and then later it can be loaded in other supported languages, however, PyTorch lacks this feature. 

10 - Device Management

By default, Tensorflow maps nearly all of the GPU memory of all GPUs visible to the process which is a comedown but here it automatically presumes that you want to run your code on the GPU because of the well-set defaults and thus result in fair management of the device. On the other hand, PyTorch keeps track of the currently selected GPU and all the CUDA tensors which will be allocated.


TensorFlow Extended (TFX): Machine Learning Pipelines

TensorFlow Extended (TFX): Machine Learning Pipelines

TensorFlow Extended (TFX): Machine Learning Pipelines

TensorFlow Extended (TFX): Machine Learning Pipelines

Speaker: Martin Andrews

Event: Google I/O Recap 2019 Singapore AI - From Model to Device by BigDataX

Thanks for watching

If you liked this post, share it with all of your programming buddies!

Follow us on Facebook | Twitter

Further reading about TensorFlow and Machine Learning

Machine Learning In Node.js With TensorFlow.js

Machine Learning A-Z™: Hands-On Python & R In Data Science

TensorFlow is dead, long live TensorFlow!

A Complete Machine Learning Project Walk-Through in Python

Top 18 Machine Learning Platforms For Developers