Copyright 2019 The TensorFlow Authors.

Custom Federated Algorithms, Part 2: Implementing Federated Averaging

This tutorial is the second part of a two-part series that demonstrates how to implement custom types of federated algorithms in TFF using the Federated Core (FC), which serves as a foundation for the Federated Learning (FL) layer (tff.learning).

We encourage you to first read the first part of this series, which introduce some of the key concepts and programming abstractions used here.

This second part of the series uses the mechanisms introduced in the first part to implement a simple version of federated training and evaluation algorithms.

We encourage you to review the image classification and text generation tutorials for a higher-level and more gentle introduction to TFF's Federated Learning APIs, as they will help you put the concepts we describe here in context.

Before we start

Before we start, try to run the following "Hello World" example to make sure your environment is correctly setup. If it doesn't work, please refer to the Installation guide for instructions.

Implementing Federated Averaging

As in Federated Learning for Image Classification, we are going to use the MNIST example, but since this is intended as a low-level tutorial, we are going to bypass the Keras API and tff.simulation, write raw model code, and construct a federated data set from scratch.

Preparing federated data sets

For the sake of a demonstration, we're going to simulate a scenario in which we have data from 10 users, and each of the users contributes knowledge how to recognize a different digit. This is about as non-i.i.d. as it gets.

First, let's load the standard MNIST data:

The data comes as Numpy arrays, one with images and another with digit labels, both with the first dimension going over the individual examples. Let's write a helper function that formats it in a way compatible with how we feed federated sequences into TFF computations, i.e., as a list of lists - the outer list ranging over the users (digits), the inner ones ranging over batches of data in each client's sequence. As is customary, we will structure each batch as a pair of tensors named x and y, each with the leading batch dimension. While at it, we'll also flatten each image into a 784-element vector and rescale the pixels in it into the 0..1 range, so that we don't have to clutter the model logic with data conversions.

As a quick sanity check, let's look at the Y tensor in the last batch of data contributed by the fifth client (the one corresponding to the digit 5).

Just to be sure, let's also look at the image corresponding to the last element of that batch.

On combining TensorFlow and TFF

In this tutorial, for compactness we immediately decorate functions that introduce TensorFlow logic with tff.tf_computation. However, for more complex logic, this is not the pattern we recommend. Debugging TensorFlow can already be a challenge, and debugging TensorFlow after it has been fully serialized and then re-imported necessarily loses some metadata and limits interactivity, making debugging even more of a challenge.

Therefore, we strongly recommend writing complex TF logic as stand-alone Python functions (that is, without tff.tf_computation decoration). This way the TensorFlow logic can be developed and tested using TF best practices and tools (like eager mode), before serializing the computation for TFF (e.g., by invoking tff.tf_computation with a Python function as the argument).

Defining a loss function

Now that we have the data, let's define a loss function that we can use for training. First, let's define the type of input as a TFF named tuple. Since the size of data batches may vary, we set the batch dimension to None to indicate that the size of this dimension is unknown.

You may be wondering why we can't just define an ordinary Python type. Recall the discussion in part 1, where we explained that while we can express the logic of TFF computations using Python, under the hood TFF computations are not Python. The symbol BATCH_TYPE defined above represents an abstract TFF type specification. It is important to distinguish this abstract TFF type from concrete Python representation types, e.g., containers such as dict or collections.namedtuple that may be used to represent the TFF type in the body of a Python function. Unlike Python, TFF has a single abstract type constructor tff.NamedTupleType for tuple-like containers, with elements that can be individually named or left unnamed. This type is also used to model formal parameters of computations, as TFF computations can formally only declare one parameter and one result - you will see examples of this shortly.

Let's now define the TFF type of model parameters, again as a TFF named tuple of weights and bias.

With those definitions in place, now we can define the loss for a given model, over a single batch. Note the usage of @tf.function decorator inside the @tff.tf_computation decorator. This allows us to write TF using Python like semantics even though were inside a tf.Graph context created by the tff.tf_computation decorator.

As expected, computation batch_loss returns float32 loss given the model and a single data batch. Note how the MODEL_TYPE and BATCH_TYPE have been lumped together into a 2-tuple of formal parameters; you can recognize the type of batch_loss as (<MODEL_TYPE,BATCH_TYPE> -> float32).

As a sanity check, let's construct an initial model filled with zeros and compute the loss over the batch of data we visualized above.

Note that we feed the TFF computation with the initial model defined as a dict, even though the body of the Python function that defines it consumes model parameters as model.weight and model.bias. The arguments of the call to batch_loss aren't simply passed to the body of that function.

What happens when we invoke batch_loss? The Python body of batch_loss has already been traced and serialized in the above cell where it was defined. TFF acts as the caller to batch_loss at the computation definition time, and as the target of invocation at the time batch_loss is invoked. In both roles, TFF serves as the bridge between TFF's abstract type system and Python representation types. At the invocation time, TFF will accept most standard Python container types (dict, list, tuple, collections.namedtuple, etc.) as concrete representations of abstract TFF tuples. Also, although as noted above, TFF computations formally only accept a single parameter, you can use the familiar Python call syntax with positional and/or keyword arguments in case where the type of the parameter is a tuple - it works as expected.

Gradient descent on a single batch

Now, let's define a computation that uses this loss function to perform a single step of gradient descent. Note how in defining this function, we use batch_loss as a subcomponent. You can invoke a computation constructed with tff.tf_computation inside the body of another computation, though typically this is not necessary - as noted above, because serialization looses some debugging information, it is often preferable for more complex computations to write and test all the TensorFlow without the tff.tf_computation decorator.

When you invoke a Python function decorated with tff.tf_computation within the body of another such function, the logic of the inner TFF computation is embedded (essentially, inlined) in the logic of the outer one. As noted above, if you are writing both computations, it is likely preferable to make the inner function (batch_loss in this case) a regular Python or tf.function rather than a tff.tf_computation. However, here we illustrate that calling one tff.tf_computation inside another basically works as expected. This may be necessary if, for example, you do not have the Python code defining batch_loss, but only its serialized TFF representation.

Now, let's apply this function a few times to the initial model to see whether the loss decreases.

Gradient descent on a sequence of local data

Now, since batch_train appears to work, let's write a similar training function local_train that consumes the entire sequence of all batches from one user instead of just a single batch. The new computation will need to now consume tff.SequenceType(BATCH_TYPE) instead of BATCH_TYPE.

There are quite a few details buried in this short section of code, let's go over them one by one.

First, while we could have implemented this logic entirely in TensorFlow, relying on tf.data.Dataset.reduce to process the sequence similarly to how we've done it earlier, we've opted this time to express the logic in the glue language, as a tff.federated_computation. We've used the federated operator tff.sequence_reduce to perform the reduction.

The operator tff.sequence_reduce is used similarly to tf.data.Dataset.reduce. You can think of it as essentially the same as tf.data.Dataset.reduce, but for use inside federated computations, which as you may remember, cannot contain TensorFlow code. It is a template operator with a formal parameter 3-tuple that consists of a sequence of T-typed elements, the initial state of the reduction (we'll refer to it abstractly as zero) of some type U, and the reduction operator of type (<U,T> -> U) that alters the state of the reduction by processing a single element. The result is the final state of the reduction, after processing all elements in a sequential order. In our example, the state of the reduction is the model trained on a prefix of the data, and the elements are data batches.

Second, note that we have again used one computation (batch_train) as a component within another (local_train), but not directly. We can't use it as a reduction operator because it takes an additional parameter - the learning rate. To resolve this, we define an embedded federated computation batch_fn that binds to the local_train's parameter learning_rate in its body. It is allowed for a child computation defined this way to capture a formal parameter of its parent as long as the child computation is not invoked outside the body of its parent. You can think of this pattern as an equivalent of functools.partial in Python.

The practical implication of capturing learning_rate this way is, of course, that the same learning rate value is used across all batches.

Now, let's try the newly defined local training function on the entire sequence of data from the same user who contributed the sample batch (digit 5).

Did it work? To answer this question, we need to implement evaluation.

Local evaluation

Here's one way to implement local evaluation by adding up the losses across all data batches (we could have just as well computed the average; we'll leave it as an exercise for the reader).

Again, there are a few new elements illustrated by this code, let's go over them one by one.

First, we have used two new federated operators for processing sequences: tff.sequence_map that takes a mapping function T->U and a sequence of T, and emits a sequence of U obtained by applying the mapping function pointwise, and tff.sequence_sum that just adds all the elements. Here, we map each data batch to a loss value, and then add the resulting loss values to compute the total loss.

Note that we could have again used tff.sequence_reduce, but this wouldn't be the best choice - the reduction process is, by definition, sequential, whereas the mapping and sum can be computed in parallel. When given a choice, it's best to stick with operators that don't constrain implementation choices, so that when our TFF computation is compiled in the future to be deployed to a specific environment, one can take full advantage of all potential opportunities for a faster, more scalable, more resource-efficient execution.

Second, note that just as in local_train, the component function we need (batch_loss) takes more parameters than what the federated operator (tff.sequence_map) expects, so we again define a partial, this time inline by directly wrapping a lambda as a tff.federated_computation. Using wrappers inline with a function as an argument is the recommended way to use tff.tf_computation to embed TensorFlow logic in TFF.

Now, let's see whether our training worked.

Indeed, the loss decreased. But what happens if we evaluated it on another user's data?

As expected, things got worse. The model was trained to recognize 5, and has never seen a 0. This brings the question - how did the local training impact the quality of the model from the global perspective?

Federated evaluation

This is the point in our journey where we finally circle back to federated types and federated computations - the topic that we started with. Here's a pair of TFF types definitions for the model that originates at the server, and the data that remains on the clients.

With all the definitions introduced so far, expressing federated evaluation in TFF is a one-liner - we distribute the model to clients, let each client invoke local evaluation on its local portion of data, and then average out the loss. Here's one way to write this.

We've already seen examples of tff.federated_mean and tff.federated_map in simpler scenarios, and at the intuitive level, they work as expected, but there's more in this section of code than meets the eye, so let's go over it carefully.

First, let's break down the let each client invoke local evaluation on its local portion of data part. As you may recall from the preceding sections, local_eval has a type signature of the form (<MODEL_TYPE, LOCAL_DATA_TYPE> -> float32).

The federated operator tff.federated_map is a template that accepts as a parameter a 2-tuple that consists of the mapping function of some type T->U and a federated value of type {T}@CLIENTS (i.e., with member constituents of the same type as the parameter of the mapping function), and returns a result of type {U}@CLIENTS.

Since we're feeding local_eval as a mapping function to apply on a per-client basis, the second argument should be of a federated type {<MODEL_TYPE, LOCAL_DATA_TYPE>}@CLIENTS, i.e., in the nomenclature of the preceding sections, it should be a federated tuple. Each client should hold a full set of arguments for local_eval as a member consituent. Instead, we're feeding it a 2-element Python list. What's happening here?

Indeed, this is an example of an implicit type cast in TFF, similar to implicit type casts you may have encountered elsewhere, e.g., when you feed an int to a function that accepts a float. Implicit casting is used scarcily at this point, but we plan to make it more pervasive in TFF as a way to minimize boilerplate.

The implicit cast that's applied in this case is the equivalence between federated tuples of the form {<X,Y>}@Z, and tuples of federated values <{X}@Z,{Y}@Z>. While formally, these two are different type signatures, looking at it from the programmers's perspective, each device in Z holds two units of data X and Y. What happens here is not unlike zip in Python, and indeed, we offer an operator tff.federated_zip that allows you to perform such conversions explicity. When the tff.federated_map encounters a tuple as a second argument, it simply invokes tff.federated_zip for you.

Given the above, you should now be able to recognize the expression tff.federated_broadcast(model) as representing a value of TFF type {MODEL_TYPE}@CLIENTS, and data as a value of TFF type {LOCAL_DATA_TYPE}@CLIENTS (or simply CLIENT_DATA_TYPE), the two getting filtered together through an implicit tff.federated_zip to form the second argument to tff.federated_map.

The operator tff.federated_broadcast, as you'd expect, simply transfers data from the server to the clients.

Now, let's see how our local training affected the average loss in the system.

Indeed, as expected, the loss has increased. In order to improve the model for all users, we'll need to train in on everyone's data.

Federated training

The simplest way to implement federated training is to locally train, and then average the models. This uses the same building blocks and patters we've already discussed, as you can see below.

Note that in the full-featured implementation of Federated Averaging provided by tff.learning, rather than averaging the models, we prefer to average model deltas, for a number of reasons, e.g., the ability to clip the update norms, for compression, etc.

Let's see whether the training works by running a few rounds of training and comparing the average loss before and after.

For completeness, let's now also run on the test data to confirm that our model generalizes well.

This concludes our tutorial.

Of course, our simplified example doesn't reflect a number of things you'd need to do in a more realistic scenario - for example, we haven't computed metrics other than loss. We encourage you to study the implementation of federated averaging in tff.learning as a more complete example, and as a way to demonstrate some of the coding practices we'd like to encourage.