Table of content:
Understanding Graphs, Automatic Differentiation, and Autograd
Building Your First Neural Network
Going Deep with PyTorch
Memory Management and Using Multiple GPUs
PyTorch is one of the foremost python deep learning libraries out there. It’s the go-to choice for deep learning research, and as each day passes by, more and more companies and research labs are adopting this library.
In this series of tutorials, we will be introducing you to PyTorch, and how to make the best use of the libraries as well as the ecosystem of tools built around it. We’ll first cover the basic building blocks and then move on to how you can quickly prototype custom architectures. We will finally conclude with a couple of posts on how to scale your code, and how to debug your code if things go awry.
You can get all the code in this post, (and other posts as well) in the Github repo here.
A lot of tutorial series on PyTorch would start begin with a rudimentary discussion of what the basic structures are. However, I’d like to instead start by discussing automatic differentiation first.
Automatic Differentiation is a building block of not only PyTorch but every DL library out there. In my opinion, PyTorch’s automatic differentiation engine, called Autograd is a brilliant tool to understand how automatic differentiation works. This will not only help you understand PyTorch better but also other DL libraries.
Modern neural network architectures can have millions of learnable parameters. From a computational point of view, training a neural network consists of two phases:
The forward pass is pretty straightforward. The output of one layer is the input to the next and so forth.
Backward pass is a bit more complicated since it requires us to use the chain rule to compute the gradients of weights w.r.t to the loss function.
Let us take a very simple neural network consisting of just 5 neurons. Our neural network looks like the following.
The following equations describe our neural network.
Let us compute the gradients for each of the learnable parameters w.
All these gradients have been computed by applying the chain rule. Note that all the individual gradients on the right-hand side of the equations mentioned above can be computed directly since the numerators of the gradients are explicit functions of the denominators.
We could manually compute the gradients of our network as it was very simple. Imagine, what if you had a network with 152 layers. Or, if the network had multiple branches.
When we design software to implement neural networks, we want to come up with a way that can allow us to seamlessly compute the gradients, regardless of the architecture type so that the programmer doesn’t have to manually compute gradients when changes are made to the network.
We galvanize this idea in form of a data structure called a Computation graph. A computation graph looks very similar to the diagram of the graph that we made in the image above. However, the nodes in a computation graph are basically operators. These operators are basically the mathematical operators except for one case, where we need to represent the creation of a user-defined variable.
Notice that we have also denoted the leaf variables a,w1,w2,w3,w4a,w1,w2,w3,w4 in the graph for sake of clarity. However, it should be noted that they are not a part of the computation graph. What they represent in our graph is the special case for user-defined variables which we just covered as an exception.
The variables, b,c, and d are created as a result of mathematical operations, whereas variables a, w1, w2, w3, and w4 are initialized by the user itself. Since they are not created by any mathematical operator, nodes corresponding to their creation are represented by their name themselves. This is true for all the leaf nodes in the graph.
Now, we are ready to describe how we will compute gradients using a computation graph.
Each node of the computation graph, with the exception of leaf nodes, can be considered as a function that takes some inputs and produces an output. Consider the node of the graph which produces variable d from w4cw4c and w3bw3b. Therefore we can write,
Now, we can easily compute the gradient of the ff with respect to its inputs, ∂f/∂w3b and ∂f/∂w4c (which are both 1). Now, label the edges coming into the nodes with their respective gradients like the following image.
We do it for the entire graph. The graph looks like this.
Following we describe the algorithm for computing the derivative of any node in this graph with respect to the loss, L. Let’s say we want to compute the derivative, ∂f/∂w4.
If you see, the product is precisely the same expression we derived using the chain rule. If there is more than one path to a variable from L then, we multiply the edges along each path and then add them together. For example, ∂L/∂a is computed as
Now we get what a computational graph is, let’s get back to PyTorch and understand how the above is implemented in PyTorch.
Tensor is a data structure that is a fundamental building block of PyTorch. Tensors are pretty much like NumPy arrays, except that, unlike NumPy, tensors are designed to take advantage of the parallel computation capabilities of a GPU. A lot of Tensor syntax is similar to that of NumPy arrays.
import torch tsr = torch.Tensor(3,5) print(tsr) tensor([[ 0.0000e+00, 0.0000e+00, 8.4452e-29, -1.0842e-19, 1.2413e-35], [ 1.4013e-45, 1.2416e-35, 1.4013e-45, 2.3331e-35, 1.4013e-45], [ 1.0108e-36, 1.4013e-45, 8.3641e-37, 1.4013e-45, 1.0040e-36]])
One its own,
Tensor is just like a NumPy
ndarray. A data structure that can let you do fast linear algebra options. If you want PyTorch to create a graph corresponding to these operations, you will have to set the
requires_grad attribute of the
Tensor to True.
The API can be a bit confusing here. There are multiple ways to initialize tensors in PyTorch. While some ways can let you explicitly define the requires_grad in the constructor itself, others require you to set it manually after the creation of the Tensor.
t1 = torch.randn((3,3), requires_grad = True) t2 = torch.FloatTensor(3,3) # No way to specify requires_grad while initiating t2.requires_grad = True
requires_grad is contagious. It means that when a
Tensor is created by operating on other
requires_grad of the resultant
Tensor would be set
True given at least one of the tensors used for creation has it’s
requires_grad set to
Tensor has a something an attribute called
grad_fn, which refers to the mathematical operator that creates the variable. If
requires_grad is set to False,
grad_fn would be None.
In our example where, d=f(w3b,w4c)d=f(w3b,w4c), d‘s grad function would be the addition operator, since f adds it’s to input together. Notice that the addition operator is also the node in our graph that outputs d. If our Tensor is a leaf node (initialized by the user), then the
grad_fn is also None.
import torch a = torch.randn((3,3), requires_grad = True) w1 = torch.randn((3,3), requires_grad = True) w2 = torch.randn((3,3), requires_grad = True) w3 = torch.randn((3,3), requires_grad = True) w4 = torch.randn((3,3), requires_grad = True) b = w1*a c = w2*a d = w3*b + w4*c L = 10 - d print("The grad fn for a is", a.grad_fn) print("The grad fn for d is", d.grad_fn)
If you run the code above, you get the following output.
The grad fn for a is None The grad fn for d is <AddBackward0 object at 0x1033afe48>
One can use the member function
is_leaf to determine whether a variable is a leaf
Tensor or not.
All mathematical operations in PyTorch are implemented by the torch.nn.Autograd.Function class. This class has two important member functions we need to look at.
The first is its forward function, which simply computes the output using its inputs.
backward function takes the incoming gradient coming from the part of the network in front of it. As you can see, the gradient to be backpropagated from a function f is basically the gradient that is backpropagated to f from the layers in front of it multiplied by the local gradient of the output of f with respect to its inputs. This is exactly what the
backward function does.
Let’s again understand with our example of:
Algorithmically, here’s how backpropagation happens with a computation graph. (Not the actual implementation, only representative)
def backward (incoming_gradients): self.Tensor.grad = incoming_gradients for inp in self.inputs: if inp.grad_fn is not None: new_incoming_gradients = // incoming_gradient * local_grad(self.Tensor, inp) inp.grad_fn.backward(new_incoming_gradients) else: pass
self.Tensor is basically the
Tensor created by Autograd.Function, which was d in our example.
Incoming gradients and local gradients have been described above.
In order to compute derivatives in our neural network, we generally call
backward on the
Tensor representing our loss. Then, we backtrack through the graph starting from the node representing the
grad_fn of our loss.
As described above, the
backward function is recursively called throughout the graph as we backtrack. Once, we reach a leaf node, since the
grad_fn is None, but stop backtracking through that path.
One thing to note here is that PyTorch gives an error if you call
backward() on vector-valued Tensor. This means you can only call
backward on a scalar-valued Tensor. In our example, if we assume a to be a vector-valued Tensor, and call
backward on L, it will throw up an error.
import torch a = torch.randn((3,3), requires_grad = True) w1 = torch.randn((3,3), requires_grad = True) w2 = torch.randn((3,3), requires_grad = True) w3 = torch.randn((3,3), requires_grad = True) w4 = torch.randn((3,3), requires_grad = True) b = w1*a c = w2*a d = w3*b + w4*c L = (10 - d) L.backward()
Running the above snippet results in the following error.
RuntimeError: grad can be implicitly created only for scalar outputs
This is because gradients can be computed with respect to scalar values by definition. You can’t exactly differentiate a vector with respect to another vector. The mathematical entity used for such cases is called a Jacobian, the discussion of which is beyond the scope of this article.
There are two ways to overcome this.
If you just make a small change in the above code set
L to be the sum of all the errors, our problem will be solved.
import torch a = torch.randn((3,3), requires_grad = True) w1 = torch.randn((3,3), requires_grad = True) w2 = torch.randn((3,3), requires_grad = True) w3 = torch.randn((3,3), requires_grad = True) w4 = torch.randn((3,3), requires_grad = True) b = w1*a c = w2*a d = w3*b + w4*c # Replace L = (10 - d) by L = (10 -d).sum() L.backward()
Once that’s done, you can access the gradients by calling the
grad attribute of
The second way is, for some reason have to absolutely call
backward on a vector function, you can pass a
torch.ones of the size of shape of the tensor you are trying to call backward with.
# Replace L.backward() with L.backward(torch.ones(L.shape))
backward used to take incoming gradients as its input. Doing the above makes the backward thinking that incoming gradients are just tensors of ones of the same size as L, and it’s able to backpropagate.
In this way, we can have gradients for every
Tensor , and we can update them using the Optimisation algorithm of our choice.
w1 = w1 - learning_rate * w1.grad
And so on.
PyTorch creates something called a Dynamic Computation Graph, which means that the graph is generated on the fly.
forward function of a Variable is called, there exists no node for the
grad_fn) in the graph.
a = torch.randn((3,3), requires_grad = True) #No graph yet, as a is a leaf w1 = torch.randn((3,3), requires_grad = True) #Same logic as above b = w1*a #Graph with node `mulBackward` is created.
The graph is created as a result of
forward function of many Tensors being invoked. Only then, the buffers for the non-leaf nodes allocated for the graph and intermediate values (used for computing gradients later. When you call, as the gradients are computed, these buffers (for non-leaf variables) are essentially freed, and the graph is destroyed ( In a sense, you can’t backpropagate through it since the buffers holding values to compute the gradients are gone).
Next time, you will call
forward on the same set of tensors, the leaf node buffers from the previous run will be shared, while the non-leaf node buffers will be created again.
If you call
backward more than once on a graph with non-leaf nodes, you’ll be met with the following error.
RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.
This is because the non-leaf buffers get destroyed the first time backward() is called and hence, there’s no path to navigate to the leaves when backward is invoked the second time. You can undo this non-leaf buffer destroying behavior by adding
retain_graph = True argument to the
loss.backward(retain_graph = True)
If you do the above, you will be able to backpropagate again through the same graph and the gradients will be accumulated, i.e. the next you backpropagate, the gradients will be added to those already stored in the previous back pass.
This is in contrast to the Static Computation Graphs, used by TensorFlow where the graph is declared before running the program. Then the graph is “run” by feeding values to the predefined graph. The dynamic graph paradigm allows you to make changes to your network architecture during runtime, as a graph is created only when a piece of code is run. This means a graph may be redefined during the lifetime of a program since you don’t have to define it beforehand. This, however, is not possible with static graphs where graphs are created before running the program, and merely executed later. Dynamic graphs also make debugging way easier since it’s easier to locate the source of your error.
This is an attribute of the
Tensor class. By default, it’s False. It comes in handy when you have to freeze some layers and stop them from updating parameters while training. You can simply set the
requires_grad to False, and these
Tensors won’t participate in the computation graph.
Thus, no gradient would be propagated to them, or to those layers which depend upon these layers for gradient flow
requires_grad. When set to True,
requires_grad is contagious meaning even if one operand of an operation has
requires_grad set to True, so will the result.
When we are computing gradients, we need to cache input values, and intermediate features as they may be required to compute the gradient later.
The gradient of b=w1∗a w.r.t it’s inputs w1 and a is a and w1, respectively. We need to store these values for gradient computation during the backward pass. This affects the memory footprint of the network.
While we are performing inference, we don’t compute gradients, and thus, don’t need to store these values. In fact, no graph needs to be created during inference as it will lead to the useless consumption of memory.
PyTorch offers a context manager, called
torch.no_grad for this purpose.
with torch.no_grad: inference code goes here
No graph is defined for operations executed under this context manager.
Understanding how Autograd and computation graphs work can make life with PyTorch a whole lot easier.