Warning

This package is currently experimental and may change in the near future.

## Overview¶

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.

### Record vs Pause¶

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 = mx.nd.array([1,2,3,4])
...     y = x * x + 1
>>> y.backward()
[ 2.  4.  6.  8.]



## Train mode and Predict Mode¶

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, MXNet 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 MXNet gluon book.

 record Returns an autograd recording scope context to be used in ‘with’ statement and captures code that needs gradients to be calculated. pause Returns a scope context to be used in ‘with’ statement for codes that do not need gradients to be calculated. 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. 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. backward Compute the gradients of heads w.r.t previously marked variables. set_training Set status to training/predicting. is_training Get status on training/predicting. set_recording Set status to recording/not recording. is_recording Get status on recording/not recording. mark_variables Mark NDArrays as variables to compute gradient for autograd. Function User-defined differentiable function.

## API Reference¶

mxnet.autograd.set_recording(is_recording)

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

Parameters: is_recording (bool) – previous state before this set.
mxnet.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) – previous state before this set.
mxnet.autograd.is_recording()

Get status on recording/not recording.

Returns: Current state of recording.
mxnet.autograd.is_training()

Get status on training/predicting.

Returns: Current state of training/predicting.
mxnet.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():
y = model(x)
backward([y])
metric.update(...)
optim.step(...)

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.
mxnet.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])

Parameters: train_mode (bool, default False) – Whether to do forward for training or predicting.
mxnet.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)
y = dropout(y)

mxnet.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)
y = sampling(y)
backward([y])

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

Parameters: variables (NDArray or list of NDArray) – gradients (NDArray or list of NDArray) – grad_reqs (str or list of str) –
mxnet.autograd.backward(heads, head_grads=None, retain_graph=False, train_mode=True)

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.
mxnet.autograd.get_symbol(x)

Retrieve recorded computation history as Symbol.

Parameters: x (NDArray) – Array representing the head of computation graph. The retrieved Symbol. Symbol
class mxnet.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 + mx.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.