Automatic differentiation package - torch.autograd¶
torch.autograd provides classes and functions implementing automatic
differentiation of arbitrary scalar valued functions. It requires minimal
changes to the existing code - you only need to declare
for which gradients should be computed with the
As of now, we only support autograd for floating point
Tensor types (
half, float, double and bfloat16) and complex
Tensor types (cfloat, cdouble).
Computes the sum of gradients of given tensors with respect to graph leaves.
Computes and returns the sum of gradients of outputs with respect to the inputs.
Functional higher level API¶
This API is in beta. Even though the function signatures are very unlikely to change, major improvements to performances are planned before we consider this stable.
This section contains the higher level API for the autograd that builds on the basic API above and allows you to compute jacobians, hessians, etc.
This API works with user-provided functions that take only Tensors as input and return
If your function takes other arguments that are not Tensors or Tensors that don’t have requires_grad set,
you can use a lambda to capture them.
For example, for a function
f that takes three inputs, a Tensor for which we want the jacobian, another
tensor that should be considered constant and a boolean flag as
f(input, constant, flag=flag)
you can use it as
functional.jacobian(lambda x: f(x, constant, flag=flag), input).
Function that computes the Jacobian of a given function.
Function that computes the Hessian of a given scalar function.
Function that computes the dot product between a vector
Function that computes the dot product between the Jacobian of the given function at the point given by the inputs and a vector
Function that computes the dot product between a vector
Function that computes the dot product between the Hessian of a given scalar function and a vector
Locally disabling gradient computation¶
See Locally disabling gradient computation for more information on the differences between no-grad and inference mode as well as other related mechanisms that may be confused with the two.
Context-manager that disabled gradient calculation.
Context-manager that enables gradient calculation.
Context-manager that sets gradient calculation to on or off.
Context-manager that enables or disables inference mode
Default gradient layouts¶
param.grad is initially
param’s memory is non-overlapping and dense,
.gradis created with strides matching
.gradis created with rowmajor-contiguous strides.
param already has a non-sparse
.gradin-place, which preserves its strides.
.gradwith a new tensor
.grad + new grad, which attempts (but does not guarantee) matching the preexisting
The default behavior (letting
None before the first
backward(), such that their layout is created according to 1 or 2,
and retained over time according to 3 or 4) is recommended for best performance.
optimizer.zero_grad() will not affect
In fact, resetting all
None before each
accumulation phase, e.g.:
for iterations... ... for param in model.parameters(): param.grad = None loss.backward()
such that they’re recreated according to 1 or 2 every time,
is a valid alternative to
that may improve performance for some networks.
Manual gradient layouts¶
If you need manual control over
param.grad = a zeroed tensor with desired strides
before the first
backward(), and never reset it to
3 guarantees your layout is preserved as long as
4 indicates your layout is likely preserved even if
In-place operations on Tensors¶
Supporting in-place operations in autograd is a hard matter, and we discourage their use in most cases. Autograd’s aggressive buffer freeing and reuse makes it very efficient and there are very few occasions when in-place operations actually lower memory usage by any significant amount. Unless you’re operating under heavy memory pressure, you might never need to use them.
In-place correctness checks¶
Tensor s keep track of in-place operations applied to them, and
if the implementation detects that a tensor was saved for backward in one of
the functions, but it was modified in-place afterwards, an error will be raised
once backward pass is started. This ensures that if you’re using in-place
functions and not seeing any errors, you can be sure that the computed
gradients are correct.
The Variable API has been deprecated: Variables are no longer necessary to
use autograd with tensors. Autograd automatically supports Tensors with
requires_grad set to
True. Below please find a quick guide on what
Variable(tensor, requires_grad)still work as expected, but they return Tensors instead of Variables.
var.datais the same thing as
Methods such as
var.backward(), var.detach(), var.register_hook()now work on tensors with the same method names.
autograd_tensor = torch.randn((2, 3, 4), requires_grad=True)
Tensor autograd functions¶
This attribute is
All Tensors that have
Computes the gradient of current tensor w.r.t.
Returns a new Tensor, detached from the current graph.
Detaches the Tensor from the graph that created it, making it a leaf.
Registers a backward hook.
Base class to create custom autograd.Function
To create a custom autograd.Function, subclass this class and implement the
forward()and :meth`backward` static methods. Then, to use your custom op in the forward pass, call the class method
apply. Do not call
To ensure correctness and best performance, make sure you are calling the correct methods on
ctxand validating your backward function using
See Extending torch.autograd for more details on how to use this class.
>>> class Exp(Function): >>> @staticmethod >>> def forward(ctx, i): >>> result = i.exp() >>> ctx.save_for_backward(result) >>> return result >>> >>> @staticmethod >>> def backward(ctx, grad_output): >>> result, = ctx.saved_tensors >>> return grad_output * result >>> >>> # Use it by calling the apply method: >>> output = Exp.apply(input)
Defines a formula for differentiating the operation with backward mode automatic differentiation.
Performs the operation.
Context method mixins¶
When creating a new
Function, the following methods are available to ctx.
Marks given tensors as modified in an in-place operation.
Marks outputs as non-differentiable.
Saves given tensors for a future call to
Sets whether to materialize output grad tensors.
Numerical gradient checking¶
Check gradients computed via small finite differences against analytical gradients w.r.t.
Check gradients of gradients computed via small finite differences against analytical gradients w.r.t.
Autograd includes a profiler that lets you inspect the cost of different
operators inside your model - both on the CPU and GPU. There are two modes
implemented at the moment - CPU-only using
and nvprof based (registers both CPU and GPU activity) using
profile(enabled=True, *, use_cuda=False, record_shapes=False, with_flops=False, profile_memory=False, with_stack=False, with_modules=False, use_kineto=False, use_cpu=True)¶
Context manager that manages autograd profiler state and holds a summary of results. Under the hood it just records events of functions being executed in C++ and exposes those events to Python. You can wrap any code into it and it will only report runtime of PyTorch functions. Note: profiler is thread local and is automatically propagated into the async tasks
enabled (bool, optional) – Setting this to False makes this context manager a no-op.
use_cuda (bool, optional) – Enables timing of CUDA events as well using the cudaEvent API. Adds approximately 4us of overhead to each tensor operation.
record_shapes (bool, optional) – If shapes recording is set, information about input dimensions will be collected. This allows one to see which dimensions have been used under the hood and further group by them using prof.key_averages(group_by_input_shape=True). Please note that shape recording might skew your profiling data. It is recommended to use separate runs with and without shape recording to validate the timing. Most likely the skew will be negligible for bottom most events (in a case of nested function calls). But for higher level functions the total self cpu time might be artificially increased because of the shape collection.
with_flops (bool, optional) – If with_flops is set, the profiler will estimate the FLOPs (floating point operations) value using the operator’s input shape. This allows one to estimate the hardware performance. Currently, this option only works for the matrix multiplication and 2D convolution operators.
profile_memory (bool, optional) – track tensor memory allocation/deallocation.
with_stack (bool, optional) – record source information (file and line number) for the ops.
with_modules (bool) – record module hierarchy (including function names) corresponding to the callstack of the op. e.g. If module A’s forward call’s module B’s forward which contains an aten::add op, then aten::add’s module hierarchy is A.B Note that this support exist, at the moment, only for TorchScript models and not eager mode models.
use_kineto (bool, optional) – experimental, enable profiling with Kineto profiler.
use_cpu (bool, optional) – profile CPU events; setting to
use_kineto=Trueand can be used to lower the overhead for GPU-only profiling.
>>> x = torch.randn((1, 1), requires_grad=True) >>> with torch.autograd.profiler.profile() as prof: >>> for _ in range(100): # any normal python code, really! >>> y = x ** 2 >> y.backward() >>> # NOTE: some columns were removed for brevity >>> print(prof.key_averages().table(sort_by="self_cpu_time_total")) ----------------------------------- --------------- --------------- --------------- Name Self CPU total CPU time avg Number of Calls ----------------------------------- --------------- --------------- --------------- mul 32.048ms 32.048ms 200 pow 27.041ms 27.041ms 200 PowBackward0 9.727ms 55.483ms 100 torch::autograd::AccumulateGrad 9.148ms 9.148ms 100 torch::autograd::GraphRoot 691.816us 691.816us 100 ----------------------------------- --------------- --------------- ---------------
Exports an EventList as a Chrome tracing tools file.
Averages all function events over their keys.
Returns total time spent on CPU obtained as a sum of all self times across all the events.
Averages all events.
Context manager that makes every autograd operation emit an NVTX range.
It is useful when running the program under nvprof:
nvprof --profile-from-start off -o trace_name.prof -- <regular command here>
Unfortunately, there’s no way to force nvprof to flush the data it collected to disk, so for CUDA profiling one has to use this context manager to annotate nvprof traces and wait for the process to exit before inspecting them. Then, either NVIDIA Visual Profiler (nvvp) can be used to visualize the timeline, or
torch.autograd.profiler.load_nvprof()can load the results for inspection e.g. in Python REPL.
enabled (bool, optional, default=True) – Setting
enabled=Falsemakes this context manager a no-op. Default:
record_shapes (bool, optional, default=False) – If
record_shapes=True, the nvtx range wrapping each autograd op will append information about the sizes of Tensor arguments received by that op, in the following format:
[[arg0.size(0), arg0.size(1), ...], [arg1.size(0), arg1.size(1), ...], ...]Non-tensor arguments will be represented by
. Arguments will be listed in the order they are received by the backend op. Please note that this order may not match the order in which those arguments were passed on the Python side. Also note that shape recording may increase the overhead of nvtx range creation.
>>> with torch.cuda.profiler.profile(): ... model(x) # Warmup CUDA memory allocator and profiler ... with torch.autograd.profiler.emit_nvtx(): ... model(x)
When viewing a profile created using
emit_nvtxin the Nvidia Visual Profiler, correlating each backward-pass op with the corresponding forward-pass op can be difficult. To ease this task,
emit_nvtxappends sequence number information to the ranges it generates.
During the forward pass, each function range is decorated with
seqis a running counter, incremented each time a new backward Function object is created and stashed for backward. Thus, the
seq=<N>annotation associated with each forward function range tells you that if a backward Function object is created by this forward function, the backward object will receive sequence number N. During the backward pass, the top-level range wrapping each C++ backward Function’s
apply()call is decorated with
Mis the sequence number that the backward object was created with. By comparing
stashed seqnumbers in backward with
seqnumbers in forward, you can track down which forward op created each backward Function.
Any functions executed during the backward pass are also decorated with
seq=<N>. During default backward (with
create_graph=False) this information is irrelevant, and in fact,
Nmay simply be 0 for all such functions. Only the top-level ranges associated with backward Function objects’
apply()methods are useful, as a way to correlate these Function objects with the earlier forward pass.
If, on the other hand, a backward pass with
create_graph=Trueis underway (in other words, if you are setting up for a double-backward), each function’s execution during backward is given a nonzero, useful
seq=<N>. Those functions may themselves create Function objects to be executed later during double-backward, just as the original functions in the forward pass did. The relationship between backward and double-backward is conceptually the same as the relationship between forward and backward: The functions still emit current-sequence-number-tagged ranges, the Function objects they create still stash those sequence numbers, and during the eventual double-backward, the Function objects’
apply()ranges are still tagged with
stashed seqnumbers, which can be compared to seq numbers from the backward pass.
Opens an nvprof trace file and parses autograd annotations.
Context-manager that enable anomaly detection for the autograd engine.
This does two things:
Running the forward pass with detection enabled will allow the backward pass to print the traceback of the forward operation that created the failing backward function.
Any backward computation that generate “nan” value will raise an error.
This mode should be enabled only for debugging as the different tests will slow down your program execution.
>>> import torch >>> from torch import autograd >>> class MyFunc(autograd.Function): ... @staticmethod ... def forward(ctx, inp): ... return inp.clone() ... @staticmethod ... def backward(ctx, gO): ... # Error during the backward pass ... raise RuntimeError("Some error in backward") ... return gO.clone() >>> def run_fn(a): ... out = MyFunc.apply(a) ... return out.sum() >>> inp = torch.rand(10, 10, requires_grad=True) >>> out = run_fn(inp) >>> out.backward() Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/your/pytorch/install/torch/_tensor.py", line 93, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "/your/pytorch/install/torch/autograd/__init__.py", line 90, in backward allow_unreachable=True) # allow_unreachable flag File "/your/pytorch/install/torch/autograd/function.py", line 76, in apply return self._forward_cls.backward(self, *args) File "<stdin>", line 8, in backward RuntimeError: Some error in backward >>> with autograd.detect_anomaly(): ... inp = torch.rand(10, 10, requires_grad=True) ... out = run_fn(inp) ... out.backward() Traceback of forward call that caused the error: File "tmp.py", line 53, in <module> out = run_fn(inp) File "tmp.py", line 44, in run_fn out = MyFunc.apply(a) Traceback (most recent call last): File "<stdin>", line 4, in <module> File "/your/pytorch/install/torch/_tensor.py", line 93, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "/your/pytorch/install/torch/autograd/__init__.py", line 90, in backward allow_unreachable=True) # allow_unreachable flag File "/your/pytorch/install/torch/autograd/function.py", line 76, in apply return self._forward_cls.backward(self, *args) File "<stdin>", line 8, in backward RuntimeError: Some error in backward
Context-manager that sets the anomaly detection for the autograd engine on or off.
set_detect_anomalywill enable or disable the autograd anomaly detection based on its argument
mode. It can be used as a context-manager or as a function.
detect_anomalyabove for details of the anomaly detection behaviour.
mode (bool) – Flag whether to enable anomaly detection (
True), or disable (
Saved tensors default hooks¶
Some operations need intermediary results to be saved during the forward pass in order to execute the backward pass. You can define how these saved tensors should be packed / unpacked using hooks. A common application is to trade compute for memory by saving those intermediary results to disk or to CPU instead of leaving them on the GPU. This is especially useful if you notice your model fits on GPU during evaluation, but not training. Also see Hooks for saved tensors.
Context-manager that sets a pair of pack / unpack hooks for saved tensors.
Use this context-manager to define how intermediary results of an operation should be packed before saving, and unpacked on retrieval.
In that context, the
pack_hookfunction will be called everytime an operation saves a tensor for backward (this includes intermediary results saved using
save_for_backward()but also those recorded by a PyTorch-defined operation). The output of
pack_hookis then stored in the computation graph instead of the original tensor.
unpack_hookis called when the saved tensor needs to be accessed, namely when executing
torch.autograd.grad(). It takes as argument the packed object returned by
pack_hookand should return a tensor which has the same content as the original tensor (passed as input to the corresponding
The hooks should have the following signatures:
pack_hook(tensor: Tensor) -> Any
unpack_hook(Any) -> Tensor
where the return value of
pack_hookis a valid input to
In general, you want
unpack_hook(pack_hook(t))to be equal to
tin terms of value, size, dtype and device.
>>> def pack_hook(x): ... print("Packing", x) ... return x >>> >>> def unpack_hook(x): ... print("Unpacking", x) ... return x >>> >>> a = torch.ones(5, requires_grad=True) >>> b = torch.ones(5, requires_grad=True) * 2 >>> with torch.autograd.graph.saved_tensors_hooks(pack_hook, unpack_hook): ... y = a * b Packing tensor([1., 1., 1., 1., 1.]) Packing tensor([2., 2., 2., 2., 2.]) >>> y.sum().backward() Unpacking tensor([1., 1., 1., 1., 1.]) Unpacking tensor([2., 2., 2., 2., 2.])
Performing an inplace operation on the input to either hooks may lead to undefined behavior.
Only one pair of hooks is allowed at a time. Recursively nesting this context-manager is not yet supported.
Context-manager under which tensors saved by the forward pass will be stored on cpu, then retrieved for backward.
When performing operations within this context manager, intermediary results saved in the graph during the forward pass will be moved to CPU, then copied back to the original device when needed for the backward pass. If the graph was already on CPU, no tensor copy is performed.
Use this context-manager to trade compute for GPU memory usage (e.g. when your model doesn’t fit in GPU memory during training).
>>> a = torch.randn(5, requires_grad=True, device="cuda") >>> b = torch.randn(5, requires_grad=True, device="cuda") >>> c = torch.randn(5, requires_grad=True, device="cuda") >>> >>> def f(a, b, c): ... prod_1 = a * b # a and b are saved on GPU ... with torch.autograd.graph.save_on_cpu(): ... prod_2 = prod_1 * c # prod_1 and c are saved on CPU ... y = prod_2 * a # prod_2 and a are saved on GPU ... return y >>> >>> y = f(a, b, c) >>> del a, b, c # for illustration only >>> # the content of a, b, and prod_2 are still alive on GPU >>> # the content of prod_1 and c only live on CPU >>> y.sum().backward() # all CPU tensors are moved back to GPU, for backward >>> # all intermediary tensors are released (deleted) after the call to backward