Github url


by deepmind

deepmind /sonnet

TensorFlow-based neural network library

8.4K Stars 1.2K Forks Last release: 4 months ago (v2.0.0) Apache License 2.0 816 Commits 36 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:



Documentation | Examples

Sonnet is a library built on top of TensorFlow 2designed to provide simple, composable abstractions for machine learning research.


Sonnet has been designed and built by researchers at DeepMind. It can be used to construct neural networks for many different purposes (un/supervised learning, reinforcement learning, ...). We find it is a successful abstraction for our organization, you might too!

More specifically, Sonnet provides a simple but powerful programming model centered around a single concept:


. Modules can hold references to parameters, other modules and methods that apply some function on the user input. Sonnet ships with many predefined modules (e.g.






) and some predefined networks of modules (e.g.


) but users are also encouraged to build their own modules.

Unlike many frameworks Sonnet is extremely unopinionated about how you will use your modules. Modules are designed to be self contained and entirely decoupled from one another. Sonnet does not ship with a training framework and users are encouraged to build their own or adopt those built by others.

Sonnet is also designed to be simple to understand, our code is (hopefully!) clear and focussed. Where we have picked defaults (e.g. defaults for initial parameter values) we try to point out why.

Getting Started


The easiest way to try Sonnet is to use Google Colab which offers a free Python notebook attached to a GPU or TPU.


To get started install TensorFlow 2.0 and Sonnet 2:

$ pip install tensorflow-gpu tensorflow-probability $ pip install dm-sonnet

You can run the following to verify things installed correctly:

import tensorflow as tf import sonnet as snt print("TensorFlow version {}".format(tf.\_\_version\_\_)) print("Sonnet version {}".format(snt.\_\_version\_\_))

Using existing modules

Sonnet ships with a number of built in modules that you can trivially use. For example to define an MLP we can use the


module to call a sequence of modules, passing the output of a given module as the input for the next module. We can use




to actually define our computation:

mlp = snt.Sequential([snt.Linear(1024), tf.nn.relu, snt.Linear(10),])

To use our module we need to "call" it. The


module (and most modules) define a


method that means you can call them by name:

logits = mlp(tf.random.normal([batch\_size, input\_size]))

It is also very common to request all the parameters for your module. Most modules in Sonnet create their parameters the first time they are called with some input (since in most cases the shape of the parameters is a function of the input). Sonnet modules provide two properties for accessing parameters.



property returns all


s that are referenced by the given module:

all\_variables = mlp.variables

It is worth noting that


s are not just used for parameters of your model. For example they are used to hold state in metrics used in


. In most cases users retrieve the module variables to pass them to an optimizer to be updated. In this case non-trainable variables should typically not be in that list as they are updated via a different mechanism. TensorFlow has a built in mechanism to mark variables as "trainable" (parameters of your model) vs. non-trainable (other variables). Sonnet provides a mechanism to gather all trainable variables from your module which is probably what you want to pass to an optimizer:

model\_parameters = mlp.trainable\_variables

Building your own module

Sonnet strongly encourages users to subclass


to define their own modules. Let's start by creating a simple


layer called



class MyLinear(snt.Module): def \_\_init\_\_(self, output\_size, name=None): super(MyLinear, self).\_\_init\_\_(name=name) self.output\_size = output\_size @snt.once def \_initialize(self, x): initial\_w = tf.random.normal([x.shape[1], self.output\_size]) self.w = tf.Variable(initial\_w, name="w") self.b = tf.Variable(tf.zeros([self.output\_size]), name="b") def \_\_call\_\_(self, x): self.\_initialize(x) return tf.matmul(x, self.w) + self.b

Using this module is trivial:

mod = MyLinear(32) mod(tf.ones([batch\_size, input\_size]))

By subclassing


you get many nice properties for free. For example a default implementation of


which shows constructor arguments (very useful for debugging and introspection):

\>\>\> print(repr(mod)) MyLinear(output\_size=10)

You also get the





\>\>\> mod.variables (<tf.variable shape="(10,)" ...>,
 <tf.variable shape="(1," ...>)

You may notice the


prefix on the variables above. This is because Sonnet modules also enter the modules name scope whenever methods are called. By entering the module name scope we provide a much more useful graph for tools like TensorBoard to consume (e.g. all operations that occur inside my_linear will be in a group called my_linear).

Additionally your module will now support TensorFlow checkpointing and saved model which are advanced features covered later.


Sonnet supports multiple serialization formats. The simplest format we support is Python's


, and all built in modules are tested to make sure they can be saved/loaded via pickle in the same Python process. In general we discourage the use of pickle, it is not well supported by many parts of TensorFlow and in our experience can be quite brittle.

TensorFlow Checkpointing


TensorFlow checkpointing can be used to save the value of parameters periodically during training. This can be useful to save the progress of training in case your program crashes or is stopped. Sonnet is designed to work cleanly with TensorFlow checkpointing:

checkpoint\_root = "/tmp/checkpoints" checkpoint\_name = "example" save\_prefix = os.path.join(checkpoint\_root, checkpoint\_name) my\_module = create\_my\_sonnet\_module() # Can be anything extending snt.Module. # A `Checkpoint` object manages checkpointing of the TensorFlow state associated # with the objects passed to it's constructor. Note that Checkpoint supports # restore on create, meaning that the variables of `my_module` do \*\*not\*\* need # to be created before you restore from a checkpoint (their value will be # restored when they are created). checkpoint = tf.train.Checkpoint(module=my\_module) # Most training scripts will want to restore from a checkpoint if one exists. This #&nbsp;would be the case if you interrupted your training (e.g. to use your GPU for # something else, or in a cloud environment if your instance is preempted). latest = tf.train.latest\_checkpoint(checkpoint\_root) if latest is not None: checkpoint.restore(latest) for step\_num in range(num\_steps): train(my\_module) # During training we will occasionally save the values of weights. Note that # this is a blocking call and can be slow (typically we are writing to the #&nbsp;slowest storage on the machine). If you have a more reliable setup it might be # appropriate to save less frequently. if step\_num and not step\_num % 1000:\_prefix) # Make sure to save your final values!!\_prefix)

TensorFlow Saved Model


TensorFlow saved models can be used to save a copy of your network that is decoupled from the Python source for it. This is enabled by saving a TensorFlow graph describing the computation and a checkpoint containing the value of weights.

The first thing to do in order to create a saved model is to create a


that you want to save:

my\_module = snt.nets.MLP([1024, 1024, 10]) my\_module(tf.ones([1, input\_size]))

Next, we need to create another module describing the specific parts of our model that we want to export. We advise doing this (rather than modifying the original model in-place) so you have fine grained control over what is actually exported. This is typically important to avoid creating very large saved models, and such that you only share the parts of your model you want to (e.g. you only want to share the generator for a GAN but keep the discriminator private).

@tf.function(input\_signature=[tf.TensorSpec([None, input\_size])]) def inference(x): return my\_module(x) to\_save = snt.Module() to\_save.inference = inference to\_save.all\_variables = list(my\_module.variables) tf.saved\\_save, "/tmp/example\_saved\_model")

We now have a saved model in the



$ ls -lh /tmp/example\_saved\_model total 24K drwxrwsr-t 2 tomhennigan 154432098 4.0K Apr 28 00:14 assets -rw-rw-r-- 1 tomhennigan 154432098 14K Apr 28 00:15 saved\_model.pb drwxrwsr-t 2 tomhennigan 154432098 4.0K Apr 28 00:15 variables

Loading this model is simple and can be done on a different machine without any of the Python code that built the saved model:

loaded = tf.saved\_model.load("/tmp/example\_saved\_model") # Use the inference method. Note this doesn't run the Python code from `to_save` # but instead uses the TensorFlow Graph that is part of the saved model. loaded.inference(tf.ones([1, input\_size])) # The all\_variables property can be used to retrieve the restored variables. assert len(loaded.all\_variables) \> 0

Note that the loaded object is not a Sonnet module, it is a container object that has the specific methods (e.g.


) and properties (e.g.


) that we added in the previous block.

Distributed training


Sonnet has support for distributed training usingcustom TensorFlow distribution strategies.

A key difference between Sonnet and distributed training using


is that Sonnet modules and optimizers do not behave differently when run under distribution strategies (e.g. we do not average your gradients or sync your batch norm stats). We believe that users should be in full control of these aspects of their training and they should not be baked into the library. The trade off here is that you need to implement these features in your training script (typically this is just 2 lines of code to all reduce your gradients before applying your optimizer) or swap in modules that are explicitly distribution aware (e.g.



Our distributed Cifar-10example walks through doing multi-GPU training with Sonnet.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.