Posit AI Weblog: torch for tabular knowledge

Machine studying on image-like knowledge might be many issues: enjoyable (canines vs. cats), societally helpful (medical imaging), or societally dangerous (surveillance). As compared, tabular knowledge – the bread and butter of knowledge science – could seem extra mundane.

What’s extra, when you’re significantly excited about deep studying (DL), and on the lookout for the additional advantages to be gained from massive knowledge, massive architectures, and large compute, you’re more likely to construct a formidable showcase on the previous as an alternative of the latter.

So for tabular knowledge, why not simply go together with random forests, or gradient boosting, or different classical strategies? I can consider not less than just a few causes to study DL for tabular knowledge:

  • Even when all of your options are interval-scale or ordinal, thus requiring “simply” some type of (not essentially linear) regression, making use of DL could end in efficiency advantages as a consequence of subtle optimization algorithms, activation features, layer depth, and extra (plus interactions of all of those).

  • If, as well as, there are categorical options, DL fashions could revenue from embedding these in steady area, discovering similarities and relationships that go unnoticed in one-hot encoded representations.

  • What if most options are numeric or categorical, however there’s additionally textual content in column F and a picture in column G? With DL, completely different modalities might be labored on by completely different modules that feed their outputs into a typical module, to take over from there.


On this introductory submit, we preserve the structure simple. We don’t experiment with fancy optimizers or nonlinearities. Nor can we add in textual content or picture processing. Nonetheless, we do make use of embeddings, and fairly prominently at that. Thus from the above bullet record, we’ll shed a light-weight on the second, whereas leaving the opposite two for future posts.

In a nutshell, what we’ll see is

  • Easy methods to create a customized dataset, tailor-made to the particular knowledge you could have.

  • Easy methods to deal with a mixture of numeric and categorical knowledge.

  • Easy methods to extract continuous-space representations from the embedding modules.


The dataset, Mushrooms, was chosen for its abundance of categorical columns. It’s an uncommon dataset to make use of in DL: It was designed for machine studying fashions to deduce logical guidelines, as in: IF a AND NOT b OR c […], then it’s an x.

Mushrooms are categorised into two teams: edible and non-edible. The dataset description lists 5 doable guidelines with their ensuing accuracies. Whereas the least we need to go into right here is the hotly debated subject of whether or not DL is suited to, or the way it could possibly be made extra suited to rule studying, we’ll enable ourselves some curiosity and take a look at what occurs if we successively take away all columns used to assemble these 5 guidelines.

Oh, and earlier than you begin copy-pasting: Right here is the instance in a Google Colaboratory notebook.


  destfile = "agaricus-lepiota.knowledge"

mushroom_data <- read_csv(
  col_names = c(
  col_types = rep("c", 23) %>% paste(collapse = "")
) %>%
  # can as nicely take away as a result of there's simply 1 distinctive worth

In torch, dataset() creates an R6 class. As with most R6 lessons, there’ll often be a necessity for an initialize() methodology. Beneath, we use initialize() to preprocess the info and retailer it in handy items. Extra on that in a minute. Previous to that, please notice the 2 different strategies a dataset has to implement:

  • .getitem(i) . That is the entire function of a dataset: Retrieve and return the commentary situated at some index it’s requested for. Which index? That’s to be determined by the caller, a dataloader. Throughout coaching, often we need to permute the order during which observations are used, whereas not caring about order in case of validation or take a look at knowledge.

  • .size(). This methodology, once more to be used of a dataloader, signifies what number of observations there are.

In our instance, each strategies are simple to implement. .getitem(i) straight makes use of its argument to index into the info, and .size() returns the variety of observations:

mushroom_dataset <- dataset(
  title = "mushroom_dataset",

  initialize = perform(indices) {
    knowledge <- self$prepare_mushroom_data(mushroom_data[indices, ])
    self$xcat <- knowledge[[1]][[1]]
    self$xnum <- knowledge[[1]][[2]]
    self$y <- knowledge[[2]]

  .getitem = perform(i) {
    xcat <- self$xcat[i, ]
    xnum <- self$xnum[i, ]
    y <- self$y[i, ]
    list(x = list(xcat, xnum), y = y)
  .size = perform() {
  prepare_mushroom_data = perform(enter) {
    enter <- enter %>%
      mutate(across(.fns = as.issue)) 
    target_col <- enter$toxic %>% 
      as.integer() %>%
      `-`(1) %>%
    categorical_cols <- enter %>% 
      select(-toxic) %>%
      select(where(perform(x) nlevels(x) != 2)) %>%
      mutate(across(.fns = as.integer)) %>%

    numerical_cols <- enter %>%
      select(-toxic) %>%
      select(where(perform(x) nlevels(x) == 2)) %>%
      mutate(across(.fns = as.integer)) %>%
    list(list(torch_tensor(categorical_cols), torch_tensor(numerical_cols)),

As for knowledge storage, there’s a subject for the goal, self$y, however as an alternative of the anticipated self$x we see separate fields for numerical options (self$xnum) and categorical ones (self$xcat). That is only for comfort: The latter shall be handed into embedding modules, which require its inputs to be of sort torch_long(), versus most different modules that, by default, work with torch_float().

Accordingly, then, all prepare_mushroom_data() does is break aside the info into these three elements.

Indispensable apart: On this dataset, actually all options occur to be categorical – it’s simply that for some, there are however two varieties. Technically, we might simply have handled them the identical because the non-binary options. However since usually in DL, we simply go away binary options the way in which they’re, we use this as an event to point out learn how to deal with a mixture of numerous knowledge varieties.

Our customized dataset outlined, we create situations for coaching and validation; every will get its companion dataloader:

train_indices <- sample(1:nrow(mushroom_data), dimension = floor(0.8 * nrow(mushroom_data)))
valid_indices <- setdiff(1:nrow(mushroom_data), train_indices)

train_ds <- mushroom_dataset(train_indices)
train_dl <- train_ds %>% dataloader(batch_size = 256, shuffle = TRUE)

valid_ds <- mushroom_dataset(valid_indices)
valid_dl <- valid_ds %>% dataloader(batch_size = 256, shuffle = FALSE)


In torch, how a lot you modularize your fashions is as much as you. Typically, excessive levels of modularization improve readability and assist with troubleshooting.

Right here we issue out the embedding performance. An embedding_module, to be handed the specific options solely, will name torch’s nn_embedding() on every of them:

embedding_module <- nn_module(
  initialize = perform(cardinalities) {
    self$embeddings = nn_module_list(lapply(cardinalities, perform(x) nn_embedding(num_embeddings = x, embedding_dim = ceiling(x/2))))
  ahead = perform(x) {
    embedded <- vector(mode = "record", size = length(self$embeddings))
    for (i in 1:length(self$embeddings)) {
      embedded[[i]] <- self$embeddings[[i]](x[ , i])
    torch_cat(embedded, dim = 2)

The principle mannequin, when known as, begins by embedding the specific options, then appends the numerical enter and continues processing:

web <- nn_module(

  initialize = perform(cardinalities,
                        fc2_dim) {
    self$embedder <- embedding_module(cardinalities)
    self$fc1 <- nn_linear(sum(map(cardinalities, perform(x) ceiling(x/2)) %>% unlist()) + num_numerical, fc1_dim)
    self$fc2 <- nn_linear(fc1_dim, fc2_dim)
    self$output <- nn_linear(fc2_dim, 1)

  ahead = perform(xcat, xnum) {
    embedded <- self$embedder(xcat)
    all <- torch_cat(list(embedded, xnum$to(dtype = torch_float())), dim = 2)
    all %>% self$fc1() %>%
      nnf_relu() %>%
      self$fc2() %>%
      self$output() %>%

Now instantiate this mannequin, passing in, on the one hand, output sizes for the linear layers, and on the opposite, characteristic cardinalities. The latter shall be utilized by the embedding modules to find out their output sizes, following a easy rule “embed into an area of dimension half the variety of enter values”:

cardinalities <- map(
  mushroom_data[ , 2:ncol(mushroom_data)], compose(nlevels, as.issue)) %>%
  keep(perform(x) x > 2) %>%
  unlist() %>%

num_numerical <- ncol(mushroom_data) - length(cardinalities) - 1

fc1_dim <- 16
fc2_dim <- 16

mannequin <- web(

gadget <- if (cuda_is_available()) torch_device("cuda:0") else "cpu"

mannequin <- mannequin$to(gadget = gadget)


The coaching loop now’s “enterprise as traditional”:

optimizer <- optim_adam(mannequin$parameters, lr = 0.1)

for (epoch in 1:20) {

  train_losses <- c()  

  coro::loop(for (b in train_dl) {
    output <- mannequin(b$x[[1]]$to(gadget = gadget), b$x[[2]]$to(gadget = gadget))
    loss <- nnf_binary_cross_entropy(output, b$y$to(dtype = torch_float(), gadget = gadget))
    train_losses <- c(train_losses, loss$merchandise())

  valid_losses <- c()

  coro::loop(for (b in valid_dl) {
    output <- mannequin(b$x[[1]]$to(gadget = gadget), b$x[[2]]$to(gadget = gadget))
    loss <- nnf_binary_cross_entropy(output, b$y$to(dtype = torch_float(), gadget = gadget))
    valid_losses <- c(valid_losses, loss$merchandise())

  cat(sprintf("Loss at epoch %d: coaching: %3f, validation: %3fn", epoch, mean(train_losses), mean(valid_losses)))
Loss at epoch 1: coaching: 0.274634, validation: 0.111689
Loss at epoch 2: coaching: 0.057177, validation: 0.036074
Loss at epoch 3: coaching: 0.025018, validation: 0.016698
Loss at epoch 4: coaching: 0.010819, validation: 0.010996
Loss at epoch 5: coaching: 0.005467, validation: 0.002849
Loss at epoch 6: coaching: 0.002026, validation: 0.000959
Loss at epoch 7: coaching: 0.000458, validation: 0.000282
Loss at epoch 8: coaching: 0.000231, validation: 0.000190
Loss at epoch 9: coaching: 0.000172, validation: 0.000144
Loss at epoch 10: coaching: 0.000120, validation: 0.000110
Loss at epoch 11: coaching: 0.000098, validation: 0.000090
Loss at epoch 12: coaching: 0.000079, validation: 0.000074
Loss at epoch 13: coaching: 0.000066, validation: 0.000064
Loss at epoch 14: coaching: 0.000058, validation: 0.000055
Loss at epoch 15: coaching: 0.000052, validation: 0.000048
Loss at epoch 16: coaching: 0.000043, validation: 0.000042
Loss at epoch 17: coaching: 0.000038, validation: 0.000038
Loss at epoch 18: coaching: 0.000034, validation: 0.000034
Loss at epoch 19: coaching: 0.000032, validation: 0.000031
Loss at epoch 20: coaching: 0.000028, validation: 0.000027

Whereas loss on the validation set remains to be lowering, we’ll quickly see that the community has discovered sufficient to acquire an accuracy of 100%.


To verify classification accuracy, we re-use the validation set, seeing how we haven’t employed it for tuning anyway.


test_dl <- valid_ds %>% dataloader(batch_size = valid_ds$.size(), shuffle = FALSE)
iter <- test_dl$.iter()
b <- iter$.subsequent()

output <- mannequin(b$x[[1]]$to(gadget = gadget), b$x[[2]]$to(gadget = gadget))
preds <- output$to(gadget = "cpu") %>% as.array()
preds <- ifelse(preds > 0.5, 1, 0)

comp_df <- data.frame(preds = preds, y = b[[2]] %>% as_array())
num_correct <- sum(comp_df$preds == comp_df$y)
num_total <- nrow(comp_df)
accuracy <- num_correct/num_total

Phew. No embarrassing failure for the DL strategy on a process the place simple guidelines are adequate. Plus, we’ve actually been parsimonious as to community dimension.

Earlier than concluding with an inspection of the discovered embeddings, let’s have some enjoyable obscuring issues.

Making the duty more durable

The next guidelines (with accompanying accuracies) are reported within the dataset description.

Disjunctive guidelines for toxic mushrooms, from most normal
    to most particular:

    P_1) odor=NOT(almond.OR.anise.OR.none)
         120 toxic circumstances missed, 98.52% accuracy

    P_2) spore-print-color=inexperienced
         48 circumstances missed, 99.41% accuracy
    P_3) odor=none.AND.stalk-surface-below-ring=scaly.AND.
         8 circumstances missed, 99.90% accuracy
    P_4) habitat=leaves.AND.cap-color=white
             100% accuracy     

    Rule P_4) may additionally be

    P_4') inhabitants=clustered.AND.cap_color=white

    These rule contain 6 attributes (out of twenty-two). 

Evidently, there’s no distinction being made between coaching and take a look at units; however we’ll stick with our 80:20 break up anyway. We’ll successively take away all talked about attributes, beginning with the three that enabled 100% accuracy, and persevering with our manner up. Listed here are the outcomes I obtained seeding the random quantity generator like so:

cap-color, inhabitants, habitat 0.9938
cap-color, inhabitants, habitat, stalk-surface-below-ring, stalk-color-above-ring 1
cap-color, inhabitants, habitat, stalk-surface-below-ring, stalk-color-above-ring, spore-print-color 0.9994
cap-color, inhabitants, habitat, stalk-surface-below-ring, stalk-color-above-ring, spore-print-color, odor 0.9526

Nonetheless 95% right … Whereas experiments like this are enjoyable, it appears like they’ll additionally inform us one thing critical: Think about the case of so-called “debiasing” by eradicating options like race, gender, or earnings. What number of proxy variables should be left that enable for inferring the masked attributes?

A have a look at the hidden representations

Wanting on the weight matrix of an embedding module, what we see are the discovered representations of a characteristic’s values. The primary categorical column was cap-shape; let’s extract its corresponding embeddings:

embedding_weights <- vector(mode = "record")
for (i in 1: length(mannequin$embedder$embeddings)) {
  embedding_weights[[i]] <- mannequin$embedder$embeddings[[i]]$parameters$weight$to(gadget = "cpu")

cap_shape_repr <- embedding_weights[[1]]
-0.0025 -0.1271  1.8077
-0.2367 -2.6165 -0.3363
-0.5264 -0.9455 -0.6702
 0.3057 -1.8139  0.3762
-0.8583 -0.7752  1.0954
 0.2740 -0.7513  0.4879
[ CPUFloatType{6,3} ]

The variety of columns is three, since that’s what we selected when creating the embedding layer. The variety of rows is six, matching the variety of out there classes. We could lookup per-feature classes within the dataset description (agaricus-lepiota.names):

cap_shapes <- c("bell", "conical", "convex", "flat", "knobbed", "sunken")

For visualization, it’s handy to do principal elements evaluation (however there are different choices, like t-SNE). Listed here are the six cap shapes in two-dimensional area:

pca <- prcomp(cap_shape_repr, middle = TRUE, scale. = TRUE, rank = 2)$x[, c("PC1", "PC2")]

pca %>% %>%
  mutate(class = cap_shapes) %>%
  ggplot(aes(x = PC1, y = PC2)) +
  geom_point() +
  geom_label_repel(aes(label = class)) + 
  coord_cartesian(xlim = c(-2, 2), ylim = c(-2, 2)) +
  theme(facet.ratio = 1) +

Naturally, how fascinating you discover the outcomes relies on how a lot you care in regards to the hidden illustration of a variable. Analyses like these could shortly flip into an exercise the place excessive warning is to be utilized, as any biases within the knowledge will instantly translate into biased representations. Furthermore, discount to two-dimensional area could or might not be ample.

This concludes our introduction to torch for tabular knowledge. Whereas the conceptual focus was on categorical options, and learn how to make use of them together with numerical ones, we’ve taken care to additionally present background on one thing that can come up again and again: defining a dataset tailor-made to the duty at hand.

Thanks for studying!

Mind picture segmentation with torch

IID: Which means and Interpretation for Rookies