This note will present an overview of how autograd works and records the operations. It’s not strictly necessary to understand all this, but we recommend getting familiar with it, as it will help you write more efficient, cleaner programs, and can aid you in debugging.
Excluding subgraphs from backward¶
Every Tensor has a flag:
requires_grad that allows for fine grained
exclusion of subgraphs from gradient computation and can increase efficiency.
If there’s a single input to an operation that requires gradient, its output will also require gradient. Conversely, only if all inputs don’t require gradient, the output also won’t require it. Backward computation is never performed in the subgraphs, where all Tensors didn’t require gradients.
>>> x = torch.randn(5, 5) # requires_grad=False by default >>> y = torch.randn(5, 5) # requires_grad=False by default >>> z = torch.randn((5, 5), requires_grad=True) >>> a = x + y >>> a.requires_grad False >>> b = a + z >>> b.requires_grad True
This is especially useful when you want to freeze part of your model, or you
know in advance that you’re not going to use gradients w.r.t. some parameters.
For example if you want to finetune a pretrained CNN, it’s enough to switch the
requires_grad flags in the frozen base, and no intermediate buffers will
be saved, until the computation gets to the last layer, where the affine
transform will use weights that require gradient, and the output of the network
will also require them.
model = torchvision.models.resnet18(pretrained=True) for param in model.parameters(): param.requires_grad = False # Replace the last fully-connected layer # Parameters of newly constructed modules have requires_grad=True by default model.fc = nn.Linear(512, 100) # Optimize only the classifier optimizer = optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9)
How autograd encodes the history¶
Autograd is reverse automatic differentiation system. Conceptually, autograd records a graph recording all of the operations that created the data as you execute operations, giving you a directed acyclic graph whose leaves are the input tensors and roots are the output tensors. By tracing this graph from roots to leaves, you can automatically compute the gradients using the chain rule.
Internally, autograd represents this graph as a graph of
Function objects (really expressions), which can be
apply() ed to compute the result of
evaluating the graph. When computing the forwards pass, autograd
simultaneously performs the requested computations and builds up a graph
representing the function that computes the gradient (the
attribute of each
torch.Tensor is an entry point into this graph).
When the forwards pass is completed, we evaluate this graph in the
backwards pass to compute the gradients.
An important thing to note is that the graph is recreated from scratch at every iteration, and this is exactly what allows for using arbitrary Python control flow statements, that can change the overall shape and size of the graph at every iteration. You don’t have to encode all possible paths before you launch the training - what you run is what you differentiate.
In-place operations with autograd¶
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.
There are two main reasons that limit the applicability of in-place operations:
In-place operations can potentially overwrite values required to compute gradients.
Every in-place operation actually requires the implementation to rewrite the computational graph. Out-of-place versions simply allocate new objects and keep references to the old graph, while in-place operations, require changing the creator of all inputs to the
Functionrepresenting this operation. This can be tricky, especially if there are many Tensors that reference the same storage (e.g. created by indexing or transposing), and in-place functions will actually raise an error if the storage of modified inputs is referenced by any other
In-place correctness checks¶
Every tensor keeps a version counter, that is incremented every time it is
marked dirty in any operation. When a Function saves any tensors for backward,
a version counter of their containing Tensor is saved as well. Once you access
self.saved_tensors it is checked, and if it is greater than the saved value
an error is raised. 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 autograd engine is responsible for running all the backward operations necessary to compute the backward pass. This section will describe all the details that can help you make the best use of it in a multithreaded environment.(this is relevant only for PyTorch 1.6+ as the behavior in previous version was different).
User could train their model with multithreading code (e.g. Hogwild training), and does not block on the concurrent backward computations, example code could be:
# Define a train function to be used in different threads def train_fn(): x = torch.ones(5, 5, requires_grad=True) # forward y = (x + 3) * (x + 4) * 0.5 # backward y.sum().backward() # potential optimizer update # User write their own threading code to drive the train_fn threads =  for _ in range(10): p = threading.Thread(target=train_fn, args=()) p.start() threads.append(p) for p in threads: p.join()
Note that some behaviors that user should be aware of:
Concurrency on CPU¶
When you run
grad() via python or C++ API in multiple
threads on CPU, you are expecting to see extra concurrency instead of
serializing all the backward calls in a specific order during execution
(behavior before PyTorch 1.6).
If you are calling
backward() on multiple thread concurrently but with
shared inputs (i.e. Hogwild CPU training). Since parameters are automatically
shared across threads, gradient accumulation might become non-deterministic on
backward calls across threads, because two backward calls might access and try
to accumulate the same
.grad attribute. This is technically not safe, and
it might result in racing condition and the result might be invalid to use.
But this is expected pattern if you are using the multithreading approach to
drive the whole training process but using shared parameters, user who use
multithreading should have the threading model in mind and should expect this
to happen. User could use the functional API
calculate the gradients instead of
backward() to avoid non-determinism.
If part of the autograd graph is shared between threads, i.e. run first
part of forward single thread, then run second part in multiple threads,
then the first part of graph is shared. In this case different threads
backward() on the same graph might have issue of
destroying the graph on the fly of one thread, and the other thread will
crash in this case. Autograd will error out to the user similar to what call
backward() twice with out
retain_graph=True, and let the user know
they should use
Thread Safety on Autograd Node¶
Since Autograd allows the caller thread to drive its backward execution for potential parallelism, it’s important that we ensure thread safety on CPU with parallel backwards that share part/whole of the GraphTask.
autograd.function is automatically thread safe because of GIL.
for built-in C++ Autograd Nodes(e.g. AccumulateGrad, CopySlices) and custom
autograd::Function, the Autograd Engine uses thread mutex locking to protect
thread safety on autograd Nodes that might have state write/read.
No thread safety on C++ hooks¶
Autograd relies on the user to write thread safe C++ hooks. If you want the hook to be correctly applied in multithreading environment, you will need to write proper thread locking code to ensure the hooks are thread safe.
Autograd for Complex Numbers¶
What notion of complex derivative does PyTorch use?¶
PyTorch follows JAX’s convention for autograd for Complex Numbers.
Suppose we have a function which we can decompose into functions u and v which compute the real and imaginary parts of the function:
def F(z): x, y = real(z), imag(z) return u(x, y) + v(x, y) * 1j
where is a unit imaginary number.
We define the for function at applied to a tangent vector as:
This is similar to the definition of the JVP for a function defined from , and the multiplication with is used to identify the result as a complex number.
We define the of at for a cotangent vector as:
In PyTorch, the VJP is mostly what we care about, as it is the computation performed when we do backward mode automatic differentiation. Notice that d and are negated in the formula above. Please look at the JAX docs to get explanation for the negative signs in the formula.
What happens if I call backward() on a complex scalar?¶
The gradient for a complex function is computed assuming the input function is a holomorphic function. This is because for general functions, the Jacobian has 4 real-valued degrees of freedom (as in the 2x2 Jacobian matrix above), so we can’t hope to represent all of them with in a complex number. However, for holomorphic functions, the gradient can be fully represented with complex numbers due to the Cauchy-Riemann equations that ensure that 2x2 Jacobians have the special form of a scale-and-rotate matrix in the complex plane, i.e. the action of a single complex number under multiplication. And so, we can obtain that gradient using backward which is just a call to vjp with covector 1.0.
The net effect of this assumption is that the partial derivatives of the imaginary part of the function
above) are discarded for
torch.autograd.backward() on a complex scalar
(e.g., this is equivalent to dropping the imaginary part of the loss before performing a backwards).
For any other desired behavior, you can specify the covector grad_output in
torch.autograd.backward() call accordingly.
How are the JVP and VJP defined for cross-domain functions?¶
Based on formulas above and the behavior we expect to see (going from should be an identity), we use the formula given below for cross-domain functions.
The and for a are defined as:
The and for a are defined as: