Note: This is an R port of the official tutorial available here. All credits goes to Soumith Chintala.

Neural networks can be constructed using the nn functionality.

Now that you had a glimpse of autograd, nn depends on autograd to define models and differentiate them. An nn.Module contains layers, and a method forward(input) that returns the output.

For example, look at this network that classifies digit images:

Convnet for mnist classification

It is a simple feed-forward network. It takes the input, feeds it through several layers one after the other, and then finally gives the output.

A typical training procedure for a neural network is as follows:

  • Define the neural network that has some learnable parameters (or weights)
  • Iterate over a dataset of inputs
  • Process input through the network
  • Compute the loss (how far is the output from being correct)
  • Propagate gradients back into the network’s parameters
  • Update the weights of the network, typically using a simple update rule: weight = weight - learning_rate * gradient.

Define the network

Let’s define this network:

Net <- nn_module(
  initialize = function() {
    self$conv1 = nn_conv2d(1, 6, 3)
    self$conv2 = nn_conv2d(6, 16, 3)
    # an affine operation: y = Wx + b
    self$fc1 = nn_linear(16 * 6 * 6, 120)  # 6*6 from image dimension
    self$fc2 = nn_linear(120, 84)
    self$fc3 = nn_linear(84, 10)
  forward = function(x) {
    x %>% 
      self$conv1() %>% 
      nnf_relu() %>% 
      nnf_max_pool2d(c(2,2)) %>% 
      self$conv2() %>% 
      nnf_relu() %>% 
      nnf_max_pool2d(c(2,2)) %>% 
      torch_flatten(start_dim = 2) %>% 
      self$fc1() %>% 
      nnf_relu() %>% 
      self$fc2() %>% 
      nnf_relu() %>% 

net <- Net()

You just have to define the forward function, and the backward function (where gradients are computed) is automatically defined for you using autograd. You can use any of the Tensor operations in the forward function.

The learnable parameters of a model are returned by net$parameters.

#> List of 10
#>  $ conv1.weight:Float [1:6, 1:1, 1:3, 1:3]
#>  $ conv1.bias  :Float [1:6]
#>  $ conv2.weight:Float [1:16, 1:6, 1:3, 1:3]
#>  $ conv2.bias  :Float [1:16]
#>  $ fc1.weight  :Float [1:120, 1:576]
#>  $ fc1.bias    :Float [1:120]
#>  $ fc2.weight  :Float [1:84, 1:120]
#>  $ fc2.bias    :Float [1:84]
#>  $ fc3.weight  :Float [1:10, 1:84]
#>  $ fc3.bias    :Float [1:10]

Let’s try a random 32x32 input. Note: expected input size of this net (LeNet) is 32x32. To use this net on the MNIST dataset, please resize the images from the dataset to 32x32.

input <- torch_randn(1, 1, 32, 32)
out <- net(input)
#> torch_tensor
#> -0.1401  0.0395 -0.0822 -0.1264 -0.0146 -0.0360  0.0296  0.0146  0.0830 -0.0447
#> [ CPUFloatType{1,10} ]

Zero the gradient buffers of all parameters and backprops with random gradients:

out$backward(torch_randn(1, 10))

Note: nn only supports mini-batches. The entire torch.nn package only supports inputs that are a mini-batch of samples, and not a single sample. For example, nn_conv2d will take in a 4D Tensor of nSamples x nChannels x Height x Width. If you have a single sample, just use input$unsqueeze(1) to add a fake batch dimension.

Before proceeding further, let’s recap all the classes you’ve seen so far.


  • torch_tensor - A multi-dimensional array with support for autograd operations like backward(). Also holds the gradient w.r.t. the tensor.

  • nn_module - Neural network module. Convenient way of encapsulating parameters, with helpers for moving them to GPU, exporting, loading, etc.

  • nn_parameter - A kind of Tensor, that is automatically registered as a parameter when assigned as an attribute to a Module.

  • autograd_function - Implements forward and backward definitions of an autograd operation. Every Tensor operation creates at least a single Function node that connects to functions that created a Tensor and encodes its history.

At this point, we covered

  • Defining a neural network
  • Processing inputs and calling backward

Still left

  • Computing the loss
  • Updating the weights of the network

Loss function

A loss function takes the (output, target) pair of inputs, and computes a value that estimates how far away the output is from the target.

There are several different loss functions under the nn package . A simple loss is: nnf_mse_loss which computes the mean-squared error between the input and the target.

For example:

output <- net(input)
target <- torch_randn(10)  # a dummy target, for example
target <- target$view(c(1, -1))  # make it the same shape as output

loss <- nnf_mse_loss(output, target)
#> torch_tensor
#> 1.29631
#> [ CPUFloatType{} ]

Now, if you follow loss in the backward direction, using its $grad_fn attribute, you will see a graph of computations that looks like this:

input -> conv2d -> relu -> maxpool2d -> conv2d -> relu -> maxpool2d
      -> view -> linear -> relu -> linear -> relu -> linear
      -> MSELoss
      -> loss

So, when we call loss$backward(), the whole graph is differentiated w.r.t. the loss, and all Tensors in the graph that has requires_grad=True will have their #grad Tensor accumulated with the gradient.

For illustration, let us follow a few steps backward:

#> MseLossBackward
#> AddmmBackward
#> torch::autograd::AccumulateGrad


To backpropagate the error all we have to do is to loss$backward(). You need to clear the existing gradients though, else gradients will be accumulated to existing gradients.

Now we shall call loss$backward(), and have a look at conv1’s bias gradients before and after the backward.

net$zero_grad()     # zeroes the gradient buffers of all parameters

# conv1.bias.grad before backward
#> torch_tensor
#>  0
#>  0
#>  0
#>  0
#>  0
#>  0
#> [ CPUFloatType{6} ]


# conv1.bias.grad after backward
#> torch_tensor
#> 0.01 *
#>  0.6251
#>  0.1677
#>  0.7964
#>  1.0600
#>  0.3815
#> -0.2729
#> [ CPUFloatType{6} ]

Now, we have seen how to use loss functions.

Update the weights

The simplest update rule used in practice is the Stochastic Gradient Descent (SGD):

\[weight = weight - learning_rate * gradient\]

We can implement this using simple R code:

learning_rate <- 0.01
for (f in net$parameters) {
    f$sub_(f$grad * learning_rate)

Note: Weight updates here is wraped around with_no_grad as we don’t the updates to be tracked by the autograd engine.

However, as you use neural networks, you want to use various different update rules such as SGD, Nesterov-SGD, Adam, RMSProp, etc.

# create your optimizer
optimizer <- optim_sgd(net$parameters, lr = 0.01)

# in your training loop:
optimizer$zero_grad()   # zero the gradient buffers
output <- net(input)
loss <- nnf_mse_loss(output, target)
optimizer$step()    # Does the update

Note: Observe how gradient buffers had to be manually set to zero using optimizer$zero_grad(). This is because gradients are accumulated as explained in the Backprop section.