Keras-APIs, SavedModels, TensorBoard, Keras-Tuner and more. On June 26 of 2019, I will be giving a TensorFlow (TF) 2.0 workshop at the . Aside from the happiness of being representing as the workshop host, I am very happy to talk about TF 2.0. PAPIs.io LATAM conference in São Paulo Daitan The idea of the workshop is to highlight what has changed from the previous 1.x version of TF. In this text, you can follow along with the main topics we are going to discuss. And of course, have a look at the for practical code. Colab notebook Introduction to TensorFlow 2.0 TensorFlow is a general purpose high-performance computing library open sourced by Google in 2015. Since the beginning, its main focus was to provide high-performance APIs for building Neural Networks (NNs). However, with the advance of time and interest by the Machine Learning (ML) community, the lib has grown to a full ML ecosystem. Currently, the library is experiencing its largest set of changes since its birth. TensorFlow 2.0 is currently in beta and brings many changes compared to TF 1.x. Let’s dive into the main ones. Eager Execution By Default To start, eager execution is the default way of running TF code. As you might recall, to build a Neural Net in TF 1.x, we needed to define this abstract data structure called a Graph. Also, (as you probably have tried), if we attempted to print one of the graph nodes, we would not see the values we were expecting. Instead, we would see a reference to the graph node. To actually, run the graph, we needed to use an encapsulation called a Session. And using the Session.run() method, we could pass Python data to the graph and actually train our models. TF 1.x code example. With eager execution, this changes. Now, TensorFlow code can be run like normal Python code. Eagerly. Meaning that operations are created and evaluated at once. Tensorflow 2.0 code example. TensorFlow 2.0 code looks a lot like NumPy code. In fact, TensorFlow and NumPy objects can easily be switched from one to the other. Hence, you do not need to worry about , , , etc. placeholders Sessions feed_dictionaties API Cleanup Many APIs like tf.gans, tf.app, tf.contrib, tf.flags are either gone or moved to separate repositories. However, one of the most important cleanups relates to how we build models. You may remember that in TF 1.x we have many more than 1 or 2 different ways of building/training ML models. are all possible APIs one can use to build NNs is TF 1.x. That not to include the Sequence to Sequence APIs in TF 1.x. And most of the time, it was not clear which one to choose for each situation. Tf.slim, tf.layers, tf.contrib.layers, tf.keras Although many of these APIs have great features, they did not seem to converge to a common way of development. Moreover, if we trained a model in one of these APIs, it was not straight forward to reuse that code using the other ones. In TF 2.0, tf.keras is the recommended high-level API. As we will see, Keras API tries to address all possible use cases. The Beginners API From TF 1.x to 2.0, the beginner API did not change much. But now, Keras is the default and recommended high-level API. In summary, Keras is a set of layers that describes how to build neural networks using a clear standard. Basically, when we install TensorFlow using pip, we get the full Keras API plus some additional functionalities. <a href="https://medium.com/media/746016574c2b973f2c367bae1f52500a/href">https://medium.com/media/746016574c2b973f2c367bae1f52500a/href</a> The beginner’s API is called Sequential. It basically defines a neural network as a stack of layers. Besides its simplicity, it has some advantages. Note that we define our model in terms of a data structure (a stack of layers). As a result, it minimizes the probability of making errors due to model definition. Keras-Tuner is a dedicated library for hyper-parameter tuning of Keras models. As of this writing, the lib is in pre-alpha status but works fine on Colab with tf.keras and Tensorflow 2.0 beta. Keras-tuner It is a very simple concept. First, need to define a model building function that returns a compiled keras model. The function takes as input a parameter called hp. Using hp, we can define a range of candidate values that we can sample hyper-parameters values. Below we build a simple model and optimize over 3 hyper-parameters. For the hidden units, we sample integer values between a pre-defined range. For dropout and learning rate, we choose at random, between some specified values. <a href="https://medium.com/media/87b409d311a23be814935c70c892fbc3/href">https://medium.com/media/87b409d311a23be814935c70c892fbc3/href</a> Then, we create a tuner object. In this case, it implements a Random Search Policy. Lastly, we can start optimization using the method. It has the same signature as . search() fit() In the end, we can check the tuner summary results and choose the best model(s). Note that training logs and model checkpoints are all saved in the directory folder (my_logs). Also, the choice of minimizing or maximizing the objective (validation accuracy) is automatically infered. Have a look at their to learn more. Github page The Advanced API The moment you see this type of implementation it goes back to Object Oriented programming. Here, your model is a Python class that extends . Model subclassing is an idea inspired by Chainer and relates very much to how PyTorch defines models. tf.keras.Model With model Subclassing, we define the model layers in the class constructor. And the method handles the definition and execution of the forward pass. call() <a href="https://medium.com/media/b723e4f079049629c6ecf160b44e8b30/href">https://medium.com/media/b723e4f079049629c6ecf160b44e8b30/href</a> Subclassing has many advantages. It is easier to perform a model inspection. We can, (using breakpoint debugging), stop at a given line and inspect the model’s activations or logits. However, with great flexibility comes more bugs. Model Subclassing requires more attention and knowledge from the programmer. In general, your code is more prominent to errors (like model wiring). Defining the Training Loop The easiest way to train a model in TF 2.0 is by using the method. supports both types of models, Sequential and Subclassing. The only adjustment you need to do, if using model Subclassing, is to override the () class method, otherwise, you can through it away. Other than that, you should be able to use with either or standard NumPy nd-arrays as input. fit() fit() compute_output_shape fit() tf.data.Dataset However, if you want a clear understanding of what is going on with the gradients or the loss, you can use the Gradient Tape. That is especially useful if you are doing research. Using Gradient Tape, one can manually define each step of a training procedure. Each of the basic steps in training a neural net such as: Forward pass Loss function evaluation Backward pass Gradient descent step is separately specified. This is much more intuitive if one wants to get a feel of how a Neural Net is trained. If you want to check the loss values w.r.t the model weights or the gradient vectors itself, you can just print them out. Gradient Tape gives much more flexibility. But just like Subclassing vs Sequential, more flexibility comes with an extra cost. Compared to the method, here we need to define a training loop manually. As a natural consequence, it makes the code more prominent to bugs and harder to debug. I believe that is a great trade off that works ideally for code engineers (looking for standardized code), compared to researchers who usually are interested in developing something new. fit() Also, using we can easily setup TensorBoard as we see next. fit() Setting up TensorBoard You can easily setup an instance of TensorBoard using the method. It also works on Jupyter/Colab notebooks. fit() In this case, you add TensorBoard as a callback to the fit method. As long as you are using the method, it works on both: Sequential and the Subclassing APIs. fit() <a href="https://medium.com/media/8e571f25855d39b5156c0e801b7d5ab3/href">https://medium.com/media/8e571f25855d39b5156c0e801b7d5ab3/href</a> If you choose to use Model Subclassing and write the training loop yourself (using Grading Tape), you also need to define TensorBoard manually. It involves creating the summary files, using and specifying which variables you want to visualize. tf.summary.create_file_writer(), As a worth noting point, there are many callbacks you can use. Some of the more useful ones are: EarlyStopping: As the name implies, it sets up a rule to stop training when a monitored quantity has stopped improving. ReduceLROnPlateau: Reduce the learning rate when a metric has stopped improving. TerminateOnNaN: Callback that terminates training when a NaN loss is encountered. LambdaCallback: Callback for creating simple, custom callbacks on-the-fly. You can check the complete list at . TensorFlow 2.0 callbacks Extracting Performance of your EagerCode If you choose to train your model using Gradient Tape, you will notice a substantial decrease in performance. Executing TF code eagerly is good for understanding, but it fails on performance. To avoid this problem, TF 2.0 introduces . tf.function Basically, if you decorate a python function with tf.function, you are asking TensorFlow to take your function and convert it to a TF high-performance abstraction. <a href="https://medium.com/media/3bae3b7a9313cf6eb4cd045c228fa46d/href">https://medium.com/media/3bae3b7a9313cf6eb4cd045c228fa46d/href</a> It means that the function will be marked for JIT compilation so that TensorFlow runs it as a graph. As a result, you get the performance benefits of TF 1.x (graphs) such as node pruning, kernel fusion, etc. In short, the idea of TF 2.0 is that you can devise your code into smaller functions. Then, you can annotate the ones you wish using , to get this extra performance. It is best to decorate functions that represent the largest computing bottlenecks. These are usually the training loops or the model’s forward pass. tf.function Note that when you decorate a function with , you loose some of the benefits of eager execution. In other words, you will not be able to setup breakpoints or use inside that section of code. tf.function print() Save and Restore Models Another great lack of standardization in TF 1.x is how we save/load trained models for production. TF 2.0 also tries to address this problem by defining a single API. Instead of having many ways of saving models, TF 2.0 standardize to an abstraction called the SavedModel. There is no much to say here. If you create a Sequential model or extend your class using , your class inherits from . As a result, you can serialize your model to a SavedModel object. tf.keras.Model tf.train.Checkpoints <a href="https://medium.com/media/bbdb1f5348751be4faa93268df37f229/href">https://medium.com/media/bbdb1f5348751be4faa93268df37f229/href</a> SavedModels are integrated with the TensorFlow ecosystem. In other words, you will be able to deploy it to many different devices. These include mobile phones, edge devices, and servers. Converting to TF-Lite If you want to deploy a SavedModel to embedded devices like Raspberry Pi, Edge TPUs or your phone, use the TF Lite converter. Note that in 2.0, the TFLiteConverter does not support frozen GraphDefs (usually generated in TF 1.x). If you want to convert a frozen GraphDefs to run in TF 2.0, you can use the tf.compat.v1.TFLiteConverter. It is very common to perform post-training quantization before deploying to embedded devices. To do it with the TFLiteConverter, set the optimizations flag to “OPTIMIZE_FOR_SIZE”. This will quantize the model’s weights from floating point to 8-bits of precision. It will reduce the model size and improve latency with little degradation in model accuracy. <a href="https://medium.com/media/93eff0a267f3eaf143c8df4ebc7f9f15/href">https://medium.com/media/93eff0a267f3eaf143c8df4ebc7f9f15/href</a> Note that this is an experimental flag, and it is subject to changes. Converting to TensorFlow.js To close up, we can also take the same SavedModel object and convert it to TensorFlow.js format. Then, we can load it using Javascript and run your model on the Browser. <a href="https://medium.com/media/5af60ecf15b9b16801ee2961fb0ec65d/href">https://medium.com/media/5af60ecf15b9b16801ee2961fb0ec65d/href</a> First, you need to install TensorFlow.js via pip. Then, use the script to take your trained-model and convert to Javascript compatible code. Finally, you can load it and perform inference in Javascript. tensorflowjs_converter You can also train models using Tesnorflow.js on the Browser. Conclusions To close off, I would like to mention some other capabilities of 2.0. First, we have seen that adding more layers to a Sequential or Subclassing model is very straightforward. And, although TF covers most of the popular layers like Conv2D, TransposeConv2D etc; you can always find yourself in a situation where you need something that is not available. That is especially true if you are reproducing some paper or doing research. The good news is that we can develop our own Custom layers. Following the same Keras API, we can create a class and extend it to . In fact, we can create custom activation functions, regularization layers, or metrics following a very similar pattern. Here is a about it. tf.keras.Layer good resource Also, we can convert existing TensorFlow 1.x code to TF 2.0. For this end, the TF team created the utility. tf_upgrade_v2 <a href="https://medium.com/media/a8c1c383cc8fd7c5f70e041804249814/href">https://medium.com/media/a8c1c383cc8fd7c5f70e041804249814/href</a> This script does not convert TF 1.x code to 2.0 idiomatics. It basically uses module for functions that got their namespaces changed. Also, if your legacy code uses , the script will not be able to convert it. You will probably need to use additional libraries or use the new TF 2.0 version of the missing functions. tf.compat.v1 tf.contrib Thanks for reading.
