Training

To actually train a model we need four things:

• A objective function, that evaluates how well a model is doing given some input data.
• The trainable parameters of the model.
• A collection of data points that will be provided to the objective function.
• An optimiser that will update the model parameters appropriately.

Training a model is typically an iterative process, where we go over the data set, calculate the objective function over the datapoints, and optimise that. This can be visualised in the form of a simple loop.

for d in datapoints

  # `d` should produce a collection of arguments
  # to the loss function

  # Calculate the gradients of the parameters
  # with respect to the loss function
  grads = Flux.gradient(parameters) do
    loss(d...)
  end

  # Update the parameters based on the chosen
  # optimiser (opt)
  Flux.Optimise.update!(opt, parameters, grads)
end

To make it easy, Flux defines train!:

train!(loss, params, data, opt; cb)

There are plenty of examples in the model zoo, and more information can be found on Custom Training Loops.

Loss Functions

The objective function must return a number representing how far the model is from its target – the loss of the model. The loss function that we defined in basics will work as an objective. In addition to custom losses, model can be trained in conjuction with the commonly used losses that are grouped under the Flux.Losses module. We can also define an objective in terms of some model:

m = Chain(
  Dense(784, 32, σ),
  Dense(32, 10), softmax)

loss(x, y) = Flux.Losses.mse(m(x), y)
ps = Flux.params(m)

# later
Flux.train!(loss, ps, data, opt)

The objective will almost always be defined in terms of some cost function that measures the distance of the prediction m(x) from the target y. Flux has several of these built in, like mse for mean squared error or crossentropy for cross entropy loss, but you can calculate it however you want. For a list of all built-in loss functions, check out the losses reference.

At first glance it may seem strange that the model that we want to train is not part of the input arguments of Flux.train! too. However the target of the optimizer is not the model itself, but the objective function that represents the departure between modelled and observed data. In other words, the model is implicitly defined in the objective function, and there is no need to give it explicitly. Passing the objective function instead of the model and a cost function separately provides more flexibility, and the possibility of optimizing the calculations.

Model parameters

The model to be trained must have a set of tracked parameters that are used to calculate the gradients of the objective function. In the basics section it is explained how to create models with such parameters. The second argument of the function Flux.train! must be an object containing those parameters, which can be obtained from a model m as params(m).

Such an object contains a reference to the model's parameters, not a copy, such that after their training, the model behaves according to their updated values.

Handling all the parameters on a layer by layer basis is explained in the Layer Helpers section. Also, for freezing model parameters, see the Advanced Usage Guide.

params(model)
params(layers...)

julia> params(Chain(Dense(ones(2,3)), softmax))  # unpacks Flux models
Params([[1.0 1.0 1.0; 1.0 1.0 1.0], [0.0, 0.0]])

julia> bn = BatchNorm(2, relu)
BatchNorm(2, relu)  # 4 parameters, plus 4 non-trainable

julia> params(bn)  # only the trainable parameters
Params([Float32[0.0, 0.0], Float32[1.0, 1.0]])

julia> params([1, 2, 3], [4])  # one or more arrays of numbers
Params([[1, 2, 3], [4]])

julia> params([[1, 2, 3], [4]])  # unpacks array of arrays
Params([[1, 2, 3], [4]])

julia> params(1, [2 2], (alpha=[3,3,3], beta=Ref(4), gamma=sin))  # ignores scalars, unpacks NamedTuples
Params([[2 2], [3, 3, 3]])

Datasets

The data argument of train! provides a collection of data to train with (usually a set of inputs x and target outputs y). For example, here's a dummy dataset with only one data point:

x = rand(784)
y = rand(10)
data = [(x, y)]

Flux.train! will call loss(x, y), calculate gradients, update the weights and then move on to the next data point if there is one. We can train the model on the same data three times:

data = [(x, y), (x, y), (x, y)]
# Or equivalently
using IterTools: ncycle
data = ncycle([(x, y)], 3)

It's common to load the xs and ys separately. In this case you can use zip:

xs = [rand(784), rand(784), rand(784)]
ys = [rand( 10), rand( 10), rand( 10)]
data = zip(xs, ys)

Training data can be conveniently partitioned for mini-batch training using the Flux.Data.DataLoader type:

X = rand(28, 28, 60000)
Y = rand(0:9, 60000)
data = DataLoader((X, Y), batchsize=128) 

Note that, by default, train! only loops over the data once (a single "epoch"). A convenient way to run multiple epochs from the REPL is provided by @epochs.

julia> using Flux: @epochs

julia> @epochs 2 println("hello")
INFO: Epoch 1
hello
INFO: Epoch 2
hello

julia> @epochs 2 Flux.train!(...)
# Train for two epochs

@epochs N body

julia> Flux.@epochs 2 println("hello")
[ Info: Epoch 1
hello
[ Info: Epoch 2
hello

Callbacks

train! takes an additional argument, cb, that's used for callbacks so that you can observe the training process. For example:

train!(objective, ps, data, opt, cb = () -> println("training"))

Callbacks are called for every batch of training data. You can slow this down using Flux.throttle(f, timeout) which prevents f from being called more than once every timeout seconds.

A more typical callback might look like this:

test_x, test_y = # ... create single batch of test data ...
evalcb() = @show(loss(test_x, test_y))
throttled_cb = throttle(evalcb, 5)
Flux.@epochs 20 Flux.train!(objective, ps, data, opt, cb = throttled_cb)

Calling Flux.stop() in a callback will exit the training loop early.

cb = function ()
  accuracy() > 0.9 && Flux.stop()
end

Custom Training loops

The Flux.train! function can be very convenient, especially for simple problems. Its also very flexible with the use of callbacks. But for some problems its much cleaner to write your own custom training loop. An example follows that works similar to the default Flux.train but with no callbacks. You don't need callbacks if you just code the calls to your functions directly into the loop. E.g. in the places marked with comments.

function my_custom_train!(loss, ps, data, opt)
  # training_loss is declared local so it will be available for logging outside the gradient calculation.
  local training_loss
  ps = Params(ps)
  for d in data
    gs = gradient(ps) do
      training_loss = loss(d...)
      # Code inserted here will be differentiated, unless you need that gradient information
      # it is better to do the work outside this block.
      return training_loss
    end
    # Insert whatever code you want here that needs training_loss, e.g. logging.
    # logging_callback(training_loss)
    # Insert what ever code you want here that needs gradient.
    # E.g. logging with TensorBoardLogger.jl as histogram so you can see if it is becoming huge.
    update!(opt, ps, gs)
    # Here you might like to check validation set accuracy, and break out to do early stopping.
  end
end

You could simplify this further, for example by hard-coding in the loss function.

Another possibility is to use Zygote.pullback to access the training loss and the gradient simultaneously.

function my_custom_train!(loss, ps, data, opt)
  ps = Params(ps)
  for d in data
    # back is a method that computes the product of the gradient so far with its argument.
    train_loss, back = Zygote.pullback(() -> loss(d...), ps)
    # Insert whatever code you want here that needs training_loss, e.g. logging.
    # logging_callback(training_loss)
    # Apply back() to the correct type of 1.0 to get the gradient of loss.
    gs = back(one(train_loss))
    # Insert what ever code you want here that needs gradient.
    # E.g. logging with TensorBoardLogger.jl as histogram so you can see if it is becoming huge.
    update!(opt, ps, gs)
    # Here you might like to check validation set accuracy, and break out to do early stopping.
  end
end