Convolutional neural networks (CNNs) are nice – they’re capable of detect options in a picture irrespective of the place. Properly, not precisely. They’re not detached to only any form of motion. Shifting up or down, or left or proper, is ok; rotating round an axis shouldn’t be. That’s due to how convolution works: traverse by row, then traverse by column (or the opposite approach spherical). If we wish “extra” (e.g., profitable detection of an upside-down object), we have to lengthen convolution to an operation that’s rotation-equivariant. An operation that’s equivariant to some sort of motion is not going to solely register the moved function per se, but additionally, hold monitor of which concrete motion made it seem the place it’s.

That is the second submit in a collection that introduces group-equivariant CNNs (GCNNs). The first was a high-level introduction to why we’d need them, and the way they work. There, we launched the important thing participant, the symmetry group, which specifies what sorts of transformations are to be handled equivariantly. Should you haven’t, please check out that submit first, since right here I’ll make use of terminology and ideas it launched.

At present, we code a easy GCNN from scratch. Code and presentation tightly comply with a pocket book supplied as a part of College of Amsterdam’s 2022 Deep Studying Course. They’ll’t be thanked sufficient for making accessible such wonderful studying supplies.

In what follows, my intent is to clarify the overall considering, and the way the ensuing structure is constructed up from smaller modules, every of which is assigned a transparent goal. For that cause, I received’t reproduce all of the code right here; as a substitute, I’ll make use of the bundle gcnn. Its strategies are closely annotated; so to see some particulars, don’t hesitate to take a look at the code.

As of at this time, gcnn implements one symmetry group: (C_4), the one which serves as a operating instance all through submit one. It’s straightforwardly extensible, although, making use of sophistication hierarchies all through.

Step 1: The symmetry group (C_4)

In coding a GCNN, the very first thing we have to present is an implementation of the symmetry group we’d like to make use of. Right here, it’s (C_4), the four-element group that rotates by 90 levels.

We are able to ask gcnn to create one for us, and examine its components.

# remotes::install_github("skeydan/gcnn")
library(gcnn)
library(torch)

C_4 <- CyclicGroup(order = 4)
elems <- C_4$components()
elems

torch_tensor
 0.0000
 1.5708
 3.1416
 4.7124
[ CPUFloatType{4} ]

Components are represented by their respective rotation angles: (0), (frac{pi}{2}), (pi), and (frac{3 pi}{2}).

Teams are conscious of the id, and know the best way to assemble a component’s inverse:

C_4$id

g1 <- elems[2]
C_4$inverse(g1)

torch_tensor
 0
[ CPUFloatType{1} ]

torch_tensor
4.71239
[ CPUFloatType{} ]

Right here, what we care about most is the group components’ motion. Implementation-wise, we have to distinguish between them appearing on one another, and their motion on the vector area (mathbb{R}^2), the place our enter pictures reside. The previous half is the straightforward one: It might merely be applied by including angles. The truth is, that is what gcnn does after we ask it to let g1 act on g2:

g2 <- elems[3]

# in C_4$left_action_on_H(), H stands for the symmetry group
C_4$left_action_on_H(torch_tensor(g1)$unsqueeze(1), torch_tensor(g2)$unsqueeze(1))

torch_tensor
 4.7124
[ CPUFloatType{1,1} ]

What’s with the unsqueeze()s? Since (C_4)’s final raison d’être is to be a part of a neural community, left_action_on_H() works with batches of components, not scalar tensors.

Issues are a bit much less simple the place the group motion on (mathbb{R}^2) is worried. Right here, we want the idea of a group illustration. That is an concerned subject, which we received’t go into right here. In our present context, it really works about like this: We have now an enter sign, a tensor we’d prefer to function on ultimately. (That “a way” might be convolution, as we’ll see quickly.) To render that operation group-equivariant, we first have the illustration apply the inverse group motion to the enter. That completed, we go on with the operation as if nothing had occurred.

To provide a concrete instance, let’s say the operation is a measurement. Think about a runner, standing on the foot of some mountain path, able to run up the climb. We’d prefer to file their top. One possibility we’ve is to take the measurement, then allow them to run up. Our measurement might be as legitimate up the mountain because it was down right here. Alternatively, we is likely to be well mannered and never make them wait. As soon as they’re up there, we ask them to return down, and after they’re again, we measure their top. The consequence is similar: Physique top is equivariant (greater than that: invariant, even) to the motion of operating up or down. (After all, top is a reasonably boring measure. However one thing extra fascinating, reminiscent of coronary heart price, wouldn’t have labored so properly on this instance.)

Returning to the implementation, it seems that group actions are encoded as matrices. There’s one matrix for every group component. For (C_4), the so-called commonplace illustration is a rotation matrix:

[
begin{bmatrix} cos(theta) & -sin(theta) sin(theta) & cos(theta) end{bmatrix}
]

In gcnn, the operate making use of that matrix is left_action_on_R2(). Like its sibling, it’s designed to work with batches (of group components in addition to (mathbb{R}^2) vectors). Technically, what it does is rotate the grid the picture is outlined on, after which, re-sample the picture. To make this extra concrete, that methodology’s code seems to be about as follows.

Here’s a goat.

img_path <- system.file("imgs", "z.jpg", bundle = "gcnn")
img <- torchvision::base_loader(img_path) |> torchvision::transform_to_tensor()
img$permute(c(2, 3, 1)) |> as.array() |> as.raster() |> plot()

First, we name C_4$left_action_on_R2() to rotate the grid.

# Grid form is [2, 1024, 1024], for a 2nd, 1024 x 1024 picture.
img_grid_R2 <- torch::torch_stack(torch::torch_meshgrid(
    checklist(
      torch::torch_linspace(-1, 1, dim(img)[2]),
      torch::torch_linspace(-1, 1, dim(img)[3])
    )
))

# Remodel the picture grid with the matrix illustration of some group component.
transformed_grid <- C_4$left_action_on_R2(C_4$inverse(g1)$unsqueeze(1), img_grid_R2)

Second, we re-sample the picture on the reworked grid. The goat now seems to be as much as the sky.

transformed_img <- torch::nnf_grid_sample(
  img$unsqueeze(1), transformed_grid,
  align_corners = TRUE, mode = "bilinear", padding_mode = "zeros"
)

transformed_img[1,..]$permute(c(2, 3, 1)) |> as.array() |> as.raster() |> plot()

Step 2: The lifting convolution

We wish to make use of present, environment friendly torch performance as a lot as potential. Concretely, we wish to use nn_conv2d(). What we want, although, is a convolution kernel that’s equivariant not simply to translation, but additionally to the motion of (C_4). This may be achieved by having one kernel for every potential rotation.

Implementing that concept is precisely what LiftingConvolution does. The precept is similar as earlier than: First, the grid is rotated, after which, the kernel (weight matrix) is re-sampled to the reworked grid.

Why, although, name this a lifting convolution? The same old convolution kernel operates on (mathbb{R}^2); whereas our prolonged model operates on mixtures of (mathbb{R}^2) and (C_4). In math communicate, it has been lifted to the semi-direct product (mathbb{R}^2rtimes C_4).

lifting_conv <- LiftingConvolution(
    group = CyclicGroup(order = 4),
    kernel_size = 5,
    in_channels = 3,
    out_channels = 8
  )

x <- torch::torch_randn(c(2, 3, 32, 32))
y <- lifting_conv(x)
y$form

[1]  2  8  4 28 28

Since, internally, LiftingConvolution makes use of a further dimension to understand the product of translations and rotations, the output shouldn’t be four-, however five-dimensional.

Step 3: Group convolutions

Now that we’re in “group-extended area”, we are able to chain quite a lot of layers the place each enter and output are group convolution layers. For instance:

group_conv <- GroupConvolution(
  group = CyclicGroup(order = 4),
    kernel_size = 5,
    in_channels = 8,
    out_channels = 16
)

z <- group_conv(y)
z$form

[1]  2 16  4 24 24

All that is still to be carried out is bundle this up. That’s what gcnn::GroupEquivariantCNN() does.

Step 4: Group-equivariant CNN

We are able to name GroupEquivariantCNN() like so.

cnn <- GroupEquivariantCNN(
    group = CyclicGroup(order = 4),
    kernel_size = 5,
    in_channels = 1,
    out_channels = 1,
    num_hidden = 2, # variety of group convolutions
    hidden_channels = 16 # variety of channels per group conv layer
)

img <- torch::torch_randn(c(4, 1, 32, 32))
cnn(img)$form

[1] 4 1

At informal look, this GroupEquivariantCNN seems to be like every previous CNN … weren’t it for the group argument.

Now, after we examine its output, we see that the extra dimension is gone. That’s as a result of after a sequence of group-to-group convolution layers, the module tasks right down to a illustration that, for every batch merchandise, retains channels solely. It thus averages not simply over places – as we usually do – however over the group dimension as properly. A remaining linear layer will then present the requested classifier output (of dimension out_channels).

And there we’ve the whole structure. It’s time for a real-world(ish) check.

Rotated digits!

The concept is to coach two convnets, a “regular” CNN and a group-equivariant one, on the standard MNIST coaching set. Then, each are evaluated on an augmented check set the place every picture is randomly rotated by a steady rotation between 0 and 360 levels. We don’t anticipate GroupEquivariantCNN to be “good” – not if we equip with (C_4) as a symmetry group. Strictly, with (C_4), equivariance extends over 4 positions solely. However we do hope it would carry out considerably higher than the shift-equivariant-only commonplace structure.

First, we put together the information; specifically, the augmented check set.

dir <- "/tmp/mnist"

train_ds <- torchvision::mnist_dataset(
  dir,
  obtain = TRUE,
  remodel = torchvision::transform_to_tensor
)

test_ds <- torchvision::mnist_dataset(
  dir,
  prepare = FALSE,
  remodel = operate(x) >
      torchvision::transform_to_tensor() 
)

train_dl <- dataloader(train_ds, batch_size = 128, shuffle = TRUE)
test_dl <- dataloader(test_ds, batch_size = 128)

How does it look?

test_images <- coro::acquire(
  test_dl, 1
)[[1]]$x[1:32, 1, , ] |> as.array()

par(mfrow = c(4, 8), mar = rep(0, 4), mai = rep(0, 4))
test_images |>
  purrr::array_tree(1) |>
  purrr::map(as.raster) |>
  purrr::iwalk(~ {
    plot(.x)
  })

We first outline and prepare a standard CNN. It’s as much like GroupEquivariantCNN(), architecture-wise, as potential, and is given twice the variety of hidden channels, in order to have comparable capability total.

 default_cnn <- nn_module(
   "default_cnn",
   initialize = operate(kernel_size, in_channels, out_channels, num_hidden, hidden_channels) {
     self$conv1 <- torch::nn_conv2d(in_channels, hidden_channels, kernel_size)
     self$convs <- torch::nn_module_list()
     for (i in 1:num_hidden) {
       self$convs$append(torch::nn_conv2d(hidden_channels, hidden_channels, kernel_size))
     }
     self$avg_pool <- torch::nn_adaptive_avg_pool2d(1)
     self$final_linear <- torch::nn_linear(hidden_channels, out_channels)
   },
   ahead = operate(x) >
       ((.) torch::nnf_layer_norm(., .$form[2:4]))() 
 )

fitted <- default_cnn |>
    luz::setup(
      loss = torch::nn_cross_entropy_loss(),
      optimizer = torch::optim_adam,
      metrics = checklist(
        luz::luz_metric_accuracy()
      )
    ) |>
    luz::set_hparams(
      kernel_size = 5,
      in_channels = 1,
      out_channels = 10,
      num_hidden = 4,
      hidden_channels = 32
    ) %>%
    luz::set_opt_hparams(lr = 1e-2, weight_decay = 1e-4) |>
    luz::match(train_dl, epochs = 10, valid_data = test_dl) 

Practice metrics: Loss: 0.0498 - Acc: 0.9843
Legitimate metrics: Loss: 3.2445 - Acc: 0.4479

Unsurprisingly, accuracy on the check set shouldn’t be that nice.

Subsequent, we prepare the group-equivariant model.

fitted <- GroupEquivariantCNN |>
  luz::setup(
    loss = torch::nn_cross_entropy_loss(),
    optimizer = torch::optim_adam,
    metrics = checklist(
      luz::luz_metric_accuracy()
    )
  ) |>
  luz::set_hparams(
    group = CyclicGroup(order = 4),
    kernel_size = 5,
    in_channels = 1,
    out_channels = 10,
    num_hidden = 4,
    hidden_channels = 16
  ) |>
  luz::set_opt_hparams(lr = 1e-2, weight_decay = 1e-4) |>
  luz::match(train_dl, epochs = 10, valid_data = test_dl)

Practice metrics: Loss: 0.1102 - Acc: 0.9667
Legitimate metrics: Loss: 0.4969 - Acc: 0.8549

For the group-equivariant CNN, accuracies on check and coaching units are so much nearer. That could be a good consequence! Let’s wrap up at this time’s exploit resuming a thought from the primary, extra high-level submit.

A problem

Going again to the augmented check set, or moderately, the samples of digits displayed, we discover an issue. In row two, column 4, there’s a digit that “beneath regular circumstances”, ought to be a 9, however, most likely, is an upside-down 6. (To a human, what suggests that is the squiggle-like factor that appears to be discovered extra usually with sixes than with nines.) Nonetheless, you could possibly ask: does this have to be an issue? Perhaps the community simply must be taught the subtleties, the sorts of issues a human would spot?

The best way I view it, all of it is dependent upon the context: What actually ought to be completed, and the way an software goes for use. With digits on a letter, I’d see no cause why a single digit ought to seem upside-down; accordingly, full rotation equivariance can be counter-productive. In a nutshell, we arrive on the similar canonical crucial advocates of truthful, simply machine studying hold reminding us of:

At all times consider the way in which an software goes for use!

In our case, although, there’s one other side to this, a technical one. gcnn::GroupEquivariantCNN() is an easy wrapper, in that its layers all make use of the identical symmetry group. In precept, there isn’t any want to do that. With extra coding effort, totally different teams can be utilized relying on a layer’s place within the feature-detection hierarchy.

Right here, let me lastly let you know why I selected the goat image. The goat is seen via a red-and-white fence, a sample – barely rotated, as a result of viewing angle – made up of squares (or edges, when you like). Now, for such a fence, kinds of rotation equivariance reminiscent of that encoded by (C_4) make lots of sense. The goat itself, although, we’d moderately not have look as much as the sky, the way in which I illustrated (C_4) motion earlier than. Thus, what we’d do in a real-world image-classification activity is use moderately versatile layers on the backside, and more and more restrained layers on the prime of the hierarchy.

Thanks for studying!

Photograph by Marjan Blan | @marjanblan on Unsplash