The autograd package enables automatic differentiation of NDArray operations. In machine learning applications, autograd is often used to calculate the gradients of loss functions with respect to parameters.

autograd records computation history on the fly to calculate gradients later. This is only enabled inside a with autograd.record(): block. A with auto_grad.pause() block can be used inside a record() block to temporarily disable recording.

To compute gradient with respect to an NDArray x, first call x.attach_grad() to allocate space for the gradient. Then, start a with autograd.record() block, and do some computation. Finally, call backward() on the result:

>>> x = gluon.nd.array([1,2,3,4])
>>> x.attach_grad()
>>> with gluon.autograd.record():
...     y = x * x + 1
>>> y.backward()
>>> print(x.grad)
[ 2.  4.  6.  8.]
<NDArray 4 @cpu(0)>

Some operators (Dropout, BatchNorm, etc) behave differently in when training and when making predictions. This can be controlled with train_mode and predict_mode scope.

By default, gluon is in predict_mode. A with autograd.record() block by default turns on train_mode (equivalent to with autograd.record(train_mode=True)). To compute a gradient in prediction mode (as when generating adversarial examples), call record with train_mode=False and then call backward(train_mode=False)

Although training usually coincides with recording, this isn't always the case. To control training vs predict_mode without changing recording vs not recording, Use a with autograd.train_mode(): or with autograd.predict_mode(): block.

Detailed tutorials are available in Part 1 of the gluon book.

Autograd for NDArray.

gluon.autograd.set_recording(is_recording)

Set status to recording/not recording. When recording, graph will be constructed for gradient computation.

Parameters:	is_recording (bool)
Returns:
Return type:	previous state before this set.

gluon.autograd.set_training(train_mode)

Set status to training/predicting. This affects ctx.is_train in operator running context. For example, Dropout will drop inputs randomly when train_mode=True while simply passing through if train_mode=False.

Parameters:	train_mode (bool)
Returns:
Return type:	previous state before this set.

gluon.autograd.is_recording()

Get status on recording/not recording.

Returns:
Return type:	Current state of recording.

gluon.autograd.is_training()

Get status on training/predicting.

Returns:
Return type:	Current state of training/predicting.

gluon.autograd.record(train_mode=True)

Returns an autograd recording scope context to be used in 'with' statement and captures code that needs gradients to be calculated.

Note: When forwarding with train_mode=False, the corresponding backward

should also use train_mode=False, otherwise gradient is undefined.

Example:

with autograd.record():
    output = net(data)
    L = loss(output, label)
    grads = L.backward(L.parameters)
    updater.step(grads)

Parameters:	train_mode (bool, default True) -- Whether the forward pass is in training or predicting mode. This controls the behavior of some layers such as Dropout, BatchNorm.

gluon.autograd.pause(train_mode=False)

Returns a scope context to be used in 'with' statement for codes that do not need gradients to be calculated.

Example:

with autograd.record():
    y = model(x)
    backward([y])
    with autograd.pause():
        # testing, IO, gradient updates...

Parameters:	train_mode (bool, default False) -- Whether to do forward for training or predicting.

gluon.autograd.train_mode()

Returns a scope context to be used in 'with' statement in which forward pass behavior is set to training mode, without changing the recording states.

Example:

y = model(x)
with autograd.train_mode():
    y = dropout(y)

gluon.autograd.predict_mode()

Returns a scope context to be used in 'with' statement in which forward pass behavior is set to inference mode, without changing the recording states.

Example:

with autograd.record():
    y = model(x)
    with autograd.predict_mode():
        y = sampling(y)
    backward([y])

gluon.autograd.mark_variables(variables, gradients, grad_reqs='write')

Mark NDArrays as variables to compute gradient for autograd.

Parameters:	variables (NDArray or list of NDArray) gradients (NDArray or list of NDArray) grad_reqs (str or list of str)

gluon.autograd.backward(heads, head_grads=None, retain_graph=False, train_mode=True)

Compute the gradients of heads w.r.t previously marked variables.

Parameters:	heads (NDArray or list of NDArray) -- Output NDArray(s) head_grads (NDArray or list of NDArray or None) -- Gradients with respect to heads. train_mode (bool, optional) -- Whether to do backward for training or predicting.

gluon.autograd.grad(heads, variables, head_grads=None, retain_graph=None, create_graph=False, train_mode=True)

Compute the gradients of heads w.r.t variables. Gradients will be returned as new NDArrays instead of stored into variable.grad. Supports recording gradient graph for computing higher order gradients.

gradients.

Parameters:	heads NDArray or list of NDArray -- Output NDArray(s) variables NDArray or list of NDArray -- Input variables to compute gradients for. head_grads NDArray or list of NDArray or None -- Gradients with respect to heads. retain_graph (bool) -- Whether to keep computation graph to differentiate again, instead of clearing history and release memory. Defaults to the same value as create_graph. create_graph (bool) -- Whether to record gradient graph for computing higher order train_mode (bool, optional) -- Whether to do backward for training or prediction.
Returns:	Gradients with respect to variables.
Return type:	NDArray or list of NDArray

Example:

>>> x = gluon.nd.ones((1,))
>>> x.attach_grad()
>>> with gluon.autograd.record():
...     z = gluon.nd.elemwise_add(gluon.nd.exp(x), x)
>>> dx = gluon.autograd.grad(z, [x], create_graph=True)
>>> dx.backward()
>>> print(dx.grad)
[ 3.71828175]
<NDArray 1 @cpu(0)>]

gluon.autograd.get_symbol(x)

Retrieve recorded computation history as Symbol.

Parameters:	x NDArray -- Array representing the head of computation graph.
Returns:	The retrieved Symbol.
Return type:	Symbol

class gluon.autograd.Function

User-defined differentiable function.

Function allows defining both forward and backward computation for custom operators. During gradient computation, the used-defined backward function will be used instead of the default chain-rule. You can also cast to numpy array and back for some operations in forward and backward.

For example, a stable sigmoid function can be defined as:

class sigmoid(Function):
    def forward(self, x):
        y = 1 / (1 + gluon.nd.exp(-x))
        self.save_for_backward(y)
        return y

    def backward(self, dy):
        # backward takes as many inputs as forward's return value,
        # and returns as many NDArrays as forward's arguments.
        y, = self.saved_tensors
        return y * (1-y)

forward(*inputs)

Forward computation.

backward(*output_grads)

Backward computation.

Takes as many inputs as forward's outputs, and returns as many NDArrays as forward's inputs.