View Source Nx.Defn (Nx v0.7.3)

Numerical functions.

A numerical function is a subset of Elixir tailored for numerical computations. For example, the following function:

defmodule MyModule do
  import Nx.Defn

  defn softmax(t) do
    Nx.exp(t) / Nx.sum(Nx.exp(t))
  end
end

will work with scalars, vector, matrices, and n-dimensional tensors. Depending on your compiler of choice, the code can even be JIT-compiled and run either on the CPU or GPU.

To support these features, defn is a subset of Elixir. It replaces Elixir's Kernel by Nx.Defn.Kernel. Nx.Defn.Kernel provides tensor-aware operators, such as +, -, etc, while also preserving many high-level constructs known to Elixir developers, such as pipe operator, aliases, conditionals, pattern-matching, the access syntax, and more:

For example, the code above can also be written as:

defmodule MyModule do
  import Nx.Defn

  defn softmax(t) do
    t
    |> Nx.exp()
    |> then(& &1 / Nx.sum(&1))
  end
end

Please consult Nx.Defn.Kernel for a complete reference.

Some of the functions in this module may also be used within defn.

Operators

defn attempts to keep as close to the Elixir semantics as possible but that's not achievable. For example, mathematical and bitwise operators (+, -, &&&, <<<, etc.) in Elixir work on numbers, which means mapping them to tensors is straight-forward and they largely preserve the same semantics, except they are now multi-dimensional.

On the other hand, the logical operators and, or, and not work with booleans in Elixir (true and false), which map to 0 and 1 in defn.

Therefore, when working with logical operators inside defn, 0 is considered false and all other numbers are considered true, which is represented as the number 1. For example, in defn, 0 and 1 as well as 0 and 2 return 0, while 1 and 1 or 1 and -1 will return 1.

The same semantics apply to conditional expressions inside defn, such as if, while, etc.

JIT compilers

The power of Nx.Defn is given by its compilers. The default compiler is Nx.Defn.Evaluator, which evalutes the code. You can use jit/3 to compile a function on the fly using a different compiler, such as EXLA:

fun = Nx.Defn.jit(&MyModule.softmax/1, compiler: EXLA)
fun.(my_tensor)

The above will return an anonymous function that optimizes, compiles, and run softmax on the fly on the CPU (or the GPU) if available. EXLA, in particular, also exports a EXLA.jit/2 function for convenience.

defn functions are compiled when they are invoked, based on the type and shapes of the tensors given as arguments. Therefore compilation may be quite time consuming on the first invocation. The compilation is then cached based on the tensors shapes and types. Calling the same function with a tensor of different values but same shape and type means no recompilation is performed.

For those interested in writing custom compilers, see Nx.Defn.Compiler.

Invoking custom Elixir code

Inside defn you can only call other defn functions and the functions in the Nx module. However, it is possible to use transforms, defined with either deftransform or deftransformp to invoke any Elixir code.

You can call code which was defined with deftransform from another module:

defmodule MyRemoteModule do
  import Nx.Defn

  deftransform remote_elixir_code(value) do
    IO.inspect(value)
  end
end

defn add_and_mult(a, b, c) do
  res = a * b + c
  MyRemoteModule.remote_elixir_code(res)
end

You can also define and call a private transform defined through deftransformp:

defn add_and_mult(a, b, c) do
  res = a * b + c
  custom_elixir_code(res)
end

deftransformp custom_elixir_code(value), do: IO.inspect(value)

The only difference between using deftransform and deftransformp is wether you want to expose and share the code with other modules, just like def and defp.

Transforms are useful to manipulate tensor expressions or Elixir data structures without the constraints of defn.

Inputs and outputs types

Nx and defn expect the arguments to be numbers, tensors, or one composite data type that implements Nx.LazyContainer. Tuples and maps implement Nx.LazyContainer by default. As previously described, defn are cached based on the shape, type, and names of the input tensors, but not their values.

defn also accepts two special arguments: functions (or tuples of functions) and lists (most commonly as keyword lists). Those values are passed as is to numerical definitions and cached as a whole. For this reason, you must never capture tensors in functions or pass tensors in keyword lists.

When numbers are given as arguments, they are always immediately converted to tensors on invocation. If you want to keep numbers as is or if you want to pass any other value to numerical definitions, they must be given as keyword lists.

Default arguments

defn functions support default arguments. They are typically used as options. For example, imagine you want to create a function named zeros, which returns a tensor of zeroes with a given type and shape. It could be implemented like this:

defn zeros(opts \\ []) do
  opts = keyword!(opts, type: {:f, 32}, shape: {})
  Nx.broadcast(Nx.tensor(0, type: opts[:type]), opts[:shape])
end

The function above accepts opts which are then validated and given default values via the keyword!/2 function. Note that while it is possible to access options via the Access syntax, such as opts[:shape], it is not possible to directly call functions in the Keyword module inside defn. To freely manipulate any Elixir value inside defn, you have to use transforms, as described in the "Invoking custom Elixir code" section.

Important! When it comes to JIT compilation, each different set of options (as well as anonymous functions) will lead to a different compilation of the numerical function.

Furthermore, if tensors are given through keyword lists, they won't be cached effectively. Tensors in defn are cached based on their shape and type, not their value, but this is not true if the tensor is given via a default argument or captured by an anonymous function. For this reason, it is extremely discouraged to pass tensors through anonymous functions and default arguments.

Working with maps and structs

While Nx supports maps in defn, you must be careful if your numerical definitions are receiving maps and returning maps. For example, imagine this code:

defn update_a(map) do
  %{map | a: Nx.add(map.a, 1)}
end

The following code increments the value under the key :a by 1. However, because the function receives the whole map and returns the whole map, it means if the map has 120 keys, the whole map will be copied to the CPU/GPU, and then brought back.

However, if you do this instead:

defn update_a(map) do
  Nx.add(map.a, 1)
end

And then update the map on Elixir, outside of defn:

%{map | a: update_a(map)}

Nx will only send the parts of the map that matters.

Recursion and loops

Given numerical definition first build a representation of your code, it is not possible to write recursive (nor tail recursive) code inside defn. Instead, one must use Nx.Defn.Kernel.while/4.

Summary

Functions

Compiles the given anonymous function with the given tensor shapes.

Wraps an anonymous function to return its underlying defn expression.

Invokes the anonymous function to return its underlying defn expression.

Gets the default options for the current process.

Sets the default options for defn in the current process.

Defines a public numerical function.

Defines a private numerical function.

Can be used to define bodiless clauses for multi-clause transforms.

Defines a transform that executes the given fun with arg when building defn expressions.

Private function version for deftransform/1

Private function version for deftransform/2

Sets the default options globally.

Receives an anonymous function and returns a new anonymous function that returns the gradient of the input function when invoked.

Computes the gradient of the given var on fun.

Wraps an anonymous function with just-in-time compilation.

Invokes the anonymous function with just-in-time compilation.

Starts streaming the given anonymous function with just-in-time compilation.

Returns a backend corresponding to the compiler options.

Receives an anonymous function and returns a new anonymous function that returns the value and gradient of the input function when invoked.

Computes the value and gradient of the given var on fun with an optional data transformation.

Functions

Link to this function

compile(fun, template_args, opts \\ [])

View Source

Compiles the given anonymous function with the given tensor shapes.

While jit/2 compiles a function just-in time based on the input shapes, this function precompiles the given anonymous function based on the input shapes. This can be beneficial for large numerical definitions, where the cache mechanism in jit/2 may take miliseconds.

For example, take the following definition:

defn softmax(t), do: Nx.exp(t) / Nx.sum(Nx.exp(t))

You can jit and then apply it as:

fun = Nx.Defn.compile(&softmax/1, [Nx.template({3}, {:s, 64})], compiler: EXLA)
fun.(Nx.tensor([1, 2, 3]))

You can also pass a mixture of templates and options when compiling a function. In such cases, you must only pass the inputs when invoking the compiled function, as the options will already be embedded in its compiled value:

fun = Nx.Defn.compile(&Nx.sum/2, [Nx.template({2, 2}, {:s, 64}), [axes: [1]]])
fun.(Nx.iota({2, 2}))

If the input tensors do not match the shape of the tensors given on compilation, it will raise.

Options

  • :compiler - the compiler for the JIT compilation

  • :hooks - a map of hooks to execute. See Nx.Defn.Kernel.hook/3

Link to this function

debug_expr(fun, opts \\ [])

View Source

Wraps an anonymous function to return its underlying defn expression.

Warning

This function must be invoked for debugging purposes only.

Options

Link to this function

debug_expr_apply(fun, args, opts \\ [])

View Source

Invokes the anonymous function to return its underlying defn expression.

Warning

This function must be invoked for debugging purposes only.

It accepts the same options as debug_expr/2.

Gets the default options for the current process.

Link to this function

default_options(options)

View Source

Sets the default options for defn in the current process.

The options defined here apply to all future invocations of defn done by the current process. It also applies to calls to the jit/3 and stream/3 functions in this module.

The default options are stored only in the process dictionary and override any global options. This means if you start a separate process, such as Task, the default options must be set on the new process too.

The function returns the values that were previously set as default options.

This function must be used only for scripting and testing.

Examples

iex> Nx.Defn.default_options(compiler: EXLA, client: :cuda)
iex> Nx.Defn.default_options()
[compiler: EXLA, client: :cuda]
Link to this macro

defn(call, list)

View Source (macro)

Defines a public numerical function.

Link to this macro

defnp(call, list)

View Source (macro)

Defines a private numerical function.

Private numerical functions are always inlined by their callers at compilation time. This happens to all local function calls within defn.

Link to this macro

deftransform(call)

View Source (macro)

Can be used to define bodiless clauses for multi-clause transforms.

See also: deftransform/2

Examples

deftransform foo(bar, baz \ 1)
deftransform foo(bar, 1), do: bar
deftransform foo(bar, baz), do: bar + baz
Link to this macro

deftransform(call, list)

View Source (macro)

Defines a transform that executes the given fun with arg when building defn expressions.

Example

Take the following defn expression:

defn tanh_power(a, b) do
  Nx.tanh(a) + Nx.pow(b, 2)
end

Let's see a trivial example, which is to use IO.inspect/1 to print a tensor expression at definition time:

defn tanh_power(a, b) do
  Nx.tanh(a) + Nx.pow(b, 2) |> my_inspect()
end

deftransformp my_inspect(expr), do: IO.inspect(expr)

Or:

defn tanh_power(a, b) do
  res = Nx.tanh(a) + Nx.pow(b, 2)
  my_inspect(res)
  res
end

When invoked in both cases, it will print the expression being built by defn:

#Nx.Defn.Expr<
  parameter a
  parameter c
  b = tanh [ a ] ()
  d = pow [ c, 2 ] ()
  e = add [ b, d ] ()
>

Although, for convenience, you might use print_expr/2 instead.

Link to this macro

deftransformp(call)

View Source (macro)

Private function version for deftransform/1

Link to this macro

deftransformp(call, list)

View Source (macro)

Private function version for deftransform/2

Link to this function

global_default_options(options)

View Source

Sets the default options globally.

The options defined here apply to all future invocations of defn. It also applies to calls to the jit/3 and stream/3 functions in this module.

You must avoid calling this function at runtime and mostly for testing purposes. You may also set in your test environment using configuration:

config :nx, :default_defn_options, [compiler: EXLA, client: :cuda]

The function returns the values that were previously set as global default options.

Receives an anonymous function and returns a new anonymous function that returns the gradient of the input function when invoked.

Examples

iex> fun = Nx.Defn.grad(fn x -> Nx.sin(x) end)
iex> fun.(Nx.tensor(0))
#Nx.Tensor<
  f32
  1.0
>

Computes the gradient of the given var on fun.

The result of the grad function must be a scalar tensor. If a non-scalar tensor is given, it is assumed the additional dimensions are batch dimensions.

Examples

defn tanh_grad(t) do
  grad(t, &Nx.tanh/1)
end

To differentiate on multiple vars, pass a tuple as first argument:

defn tanh_power_grad(a, b) do
  grad({a, b}, fn {a, b} -> Nx.tanh(a) + Nx.pow(b, 2) end)
end

var_or_vars can be any Nx.Container with one or multiple tensors.

Wraps an anonymous function with just-in-time compilation.

Once invoked, the wrapped anonymous function will perform just in time compilation with the configured compiler. For example, take the following definition:

defn softmax(t), do: Nx.exp(t) / Nx.sum(Nx.exp(t))

You can jit and then apply it as:

fun = Nx.Defn.jit(&softmax/1, compiler: EXLA)
fun.(Nx.tensor([1, 2, 3]))

Options

  • :compiler - the compiler for the JIT compilation

  • :hooks - a map of hooks to execute. See Nx.Defn.Kernel.hook/3

  • :on_conflict - what to do if a JIT compilation is already in place. It may be :raise (the default), :force (forces a new JIT compilation), or :reuse (reuses the exiting JIT compilation). It is not recommended to set the :compiler option when reusing.

Link to this function

jit_apply(fun, args, opts \\ [])

View Source

Invokes the anonymous function with just-in-time compilation.

This function is equivalent to calling jit/2 and then applying the given arguments to the anonymous function.

For example, take the following definition:

defn softmax(t), do: Nx.exp(t) / Nx.sum(Nx.exp(t))

You can jit_apply/3 it as:

Nx.Defn.jit_apply(&softmax/1, [Nx.tensor([1, 2, 3])], compiler: EXLA)

It accepts the same options as jit/2.

Link to this function

stream(fun, args, opts \\ [])

View Source

Starts streaming the given anonymous function with just-in-time compilation.

At least two arguments are expected:

  1. The first argument is a tensor template of the data to be streamed in

  2. The second argument is a tensor with the stream initial state

The streaming function must return a two element tuple, the first element is the data to be sent and the second is the accumulator.

For each streamed chunk, you must call Nx.Stream.send/2 and Nx.Stream.recv/1. You don't need to call recv immediately after send, but doing so can be a useful mechanism to provide backpressure. Once all chunks are sent, you must use Nx.Stream.done/1 to receive the accumulated result. Let's see an example:

defmodule Streamed do
  import Nx.Defn

  defn sum(tensor, acc) do
    {acc, tensor + acc}
  end
end

Now let's invoke it:

stream = Nx.Defn.stream(&Streamed.sum/2, [Nx.template({}, {:s, 64}), 0])

for i <- 1..5 do
  Nx.Stream.send(stream, i)
  IO.inspect {:chunk, Nx.Stream.recv(stream)}
end

IO.inspect {:result, Nx.Stream.done(stream)}

It will print:

{:chunk, 0}
{:chunk, 1}
{:chunk, 2}
{:chunk, 3}
{:chunk, 4}
{:result, 5}

Options

Beware: deadlocks

Some backends (such as XLA) place locks around devices. For example, if you start streaming on the GPU, you cannot perform any other operation on the GPU until streaming is over.

This means if we modify the loop above to the following:

for i <- 1..5 do
  Nx.Stream.send(stream, Nx.tensor(i) |> Nx.multiply(2))
  IO.inspect {:chunk, Nx.Stream.recv(stream)}
end

The loop may deadlock at the time it performs the multiplication. In practice, this means you should perform the streaming on the GPU and the remaining operations on the CPU. If you only have a single device (i.e. only a CPU), then it may not be possible to perform the above and you will have to restructure your code to manipulate the input before streaming starts.

Returns a backend corresponding to the compiler options.

The backend matches the backend used for outputs from computations defined by the given compiler.

Receives an anonymous function and returns a new anonymous function that returns the value and gradient of the input function when invoked.

Examples

iex> fun = Nx.Defn.value_and_grad(fn x -> Nx.sin(x) end)
iex> {value, grad} = fun.(Nx.tensor(0))
iex> value
#Nx.Tensor<
  f32
  0.0
>
iex> grad
#Nx.Tensor<
  f32
  1.0
>
Link to this function

value_and_grad(var_or_vars, fun, transform \\ & &1)

View Source

Computes the value and gradient of the given var on fun with an optional data transformation.

It returns a tuple with the value and the gradient.

Examples

defn tanh_grad(t) do
  value_and_grad(t, &Nx.tanh/1)
end

To differentiate on multiple vars, pass a tuple as first argument:

defn tanh_power_grad(a, b) do
  value_and_grad({a, b}, fn {a, b} -> Nx.tanh(a) + Nx.pow(b, 2) end)
end

var_or_vars can be any Nx.Container with one or multiple tensors.

transform allows you to transform the expression before the gradient is calculated. This enables optimizations that reuse parts of expressions. As an example, consider the following objective function:

defn objective(predict_fn, loss_fn, params, inputs, targets) do
  preds = predict_fn.(params, inputs)
  loss = loss_fn.(preds, targets)
  {preds, loss}
end

You can compute the gradient with respect to just the loss function by applying a transform:

{{preds, loss}, gradient} = value_and_grad(params, &objective(predict_fn, loss_fn, &1, inputs, targets), &elem(&1, 1))

preds can be re-used to compute other metrics such as accuracy, absolute error, etc. without having to do another forward pass.