Text classification from scratch • keras3

Introduction

This example shows how to do text classification starting from raw text (as a set of text files on disk). We demonstrate the workflow on the IMDB sentiment classification dataset (unprocessed version). We use layer_text_vectorization() for word splitting & indexing.

Setup

library(tensorflow, exclude = c("shape", "set_random_seed"))
library(tfdatasets, exclude = "shape")
library(keras3)
use_virtualenv("r-keras")

Load the data: IMDB movie review sentiment classification

Let’s download the data and inspect its structure.

if (!dir.exists("datasets/aclImdb")) {
  dir.create("datasets")
  download.file(
    "https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz",
    "datasets/aclImdb_v1.tar.gz"
  )
  untar("datasets/aclImdb_v1.tar.gz", exdir = "datasets")
  unlink("datasets/aclImdb/train/unsup", recursive = TRUE)
}

The aclImdb folder contains a train and test subfolder:

head(list.files("datasets/aclImdb/test"))

## [1] "labeledBow.feat" "neg"             "pos"             "urls_neg.txt"
## [5] "urls_pos.txt"

head(list.files("datasets/aclImdb/train"))

## [1] "labeledBow.feat" "neg"             "pos"             "unsupBow.feat"
## [5] "urls_neg.txt"    "urls_pos.txt"

The aclImdb/train/pos and aclImdb/train/neg folders contain text files, each of which represents one review (either positive or negative):

writeLines(strwrap(readLines("datasets/aclImdb/train/pos/4229_10.txt")))

## Don't waste time reading my review. Go out and see this
## astonishingly good episode, which may very well be the best Columbo
## ever written! Ruth Gordon is perfectly cast as the scheming yet
## charming mystery writer who murders her son-in-law to avenge his
## murder of her daughter. Columbo is his usual rumpled, befuddled and
## far-cleverer-than-he-seems self, and this particular installment
## features fantastic chemistry between Gordon and Falk. Ironically,
## this was not written by heralded creators Levinson or Link yet is
## possibly the densest, most thoroughly original and twist-laden
## Columbo plot ever. Utterly satisfying in nearly every department
## and overflowing with droll and witty dialogue and thinking. Truly
## unexpected and inventive climax tops all. 10/10...seek this one out
## on Netflix!

We are only interested in the pos and neg subfolders, so let’s delete the other subfolder that has text files in it:

unlink("datasets/aclImdb/train/unsup", recursive = TRUE)

You can use the utility text_dataset_from_directory() to generate a labeled tf_dataset object from a set of text files on disk filed into class-specific folders.

Let’s use it to generate the training, validation, and test datasets. The validation and training datasets are generated from two subsets of the train directory, with 20% of samples going to the validation dataset and 80% going to the training dataset.

Having a validation dataset in addition to the test dataset is useful for tuning hyperparameters, such as the model architecture, for which the test dataset should not be used.

Before putting the model out into the real world however, it should be retrained using all available training data (without creating a validation dataset), so its performance is maximized.

When using the validation_split and subset arguments, make sure to either specify a random seed, or to pass shuffle=FALSE, so that the validation & training splits you get have no overlap.

batch_size <- 32

raw_train_ds <- text_dataset_from_directory(
  "datasets/aclImdb/train",
  batch_size = batch_size,
  validation_split = 0.2,
  subset = "training",
  seed = 1337
)

## Found 25000 files belonging to 2 classes.
## Using 20000 files for training.

raw_val_ds <- text_dataset_from_directory(
  "datasets/aclImdb/train",
  batch_size = batch_size,
  validation_split = 0.2,
  subset = "validation",
  seed = 1337
)

## Found 25000 files belonging to 2 classes.
## Using 5000 files for validation.

raw_test_ds <- text_dataset_from_directory(
  "datasets/aclImdb/test",
  batch_size = batch_size
)

## Found 25000 files belonging to 2 classes.

cat("Number of batches in raw_train_ds:", length(raw_train_ds), "\n")

## Number of batches in raw_train_ds: 625

cat("Number of batches in raw_val_ds:", length(raw_val_ds), "\n")

## Number of batches in raw_val_ds: 157

cat("Number of batches in raw_test_ds:", length(raw_test_ds), "\n")

## Number of batches in raw_test_ds: 782

Let’s preview a few samples:

# It's important to take a look at your raw data to ensure your normalization
# and tokenization will work as expected. We can do that by taking a few
# examples from the training set and looking at them.
# This is one of the places where eager execution shines:
# we can just evaluate these tensors using .numpy()
# instead of needing to evaluate them in a Session/Graph context.
batch <- iter_next(as_iterator(raw_train_ds))
str(batch)

## List of 2
##  $ :<tf.Tensor: shape=(32), dtype=string, numpy=…>
##  $ :<tf.Tensor: shape=(32), dtype=int32, numpy=…>

c(text_batch, label_batch) %<-% batch
for (i in 1:3) {
  print(text_batch[i])
  print(label_batch[i])
}

## tf.Tensor(b"I have read the novel Reaper of Ben Mezrich a fews years ago and last night I accidentally came to see this adaption.<br /><br />Although it's been years since I read the story the first time, the differences between the novel and the movie are humongous. Very important elements, which made the whole thing plausible are just written out or changed to bad.<br /><br />If the plot sounds interesting to you: go and get the novel. Its much, much, much better.<br /><br />Still 4 out of 10 since it was hard to stop watching because of the great basic plot by Ben Mezrich.", shape=(), dtype=string)
## tf.Tensor(0, shape=(), dtype=int32)
## tf.Tensor(b'After seeing all the Jesse James, Quantrill, jayhawkers,etc films in the fifties, it is quite a thrill to see this film with a new perspective by director Ang Lee. The scene of the attack of Lawrence, Kansas is awesome. The romantic relationship between Jewel and Toby Mcguire turns out to be one of the best parts and Jonathan Rhys-Meyers is outstanding as the bad guy. All the time this film makes you feel the horror of war, and the desperate situation of the main characters who do not know if they are going to survive the next hours. Definitely worth seeing.', shape=(), dtype=string)
## tf.Tensor(1, shape=(), dtype=int32)
## tf.Tensor(b'AG was an excellent presentation of drama, suspense and thriller that is so rare to American TV. Sheriff Lucas gave many a viewer the willies. We rooted for Caleb as he strove to resist the overtures of Sheriff Lucas. We became engrossed and fearful upon learning of the unthinkable connection between these two characters. The manipulations which weekly gave cause to fear what Lucas would do next were truly surprising. This show lived up to the "Gothic" moniker in ways American entertainment has so seldom attempted, much less mastered. The suits definitely made a big mistake in not supporting this show. This show puts shame to the current glut of "reality" shows- which are so less than satisfying viewing.The call for a DVD box set is well based. This show is quality viewing for a discerning market hungry for quality viewing. A public that is tiring of over-saturation of mind-numbing reality fare will welcome this gem of real storytelling. Bring on the DVD box set!!', shape=(), dtype=string)
## tf.Tensor(1, shape=(), dtype=int32)

Prepare the data

In particular, we remove <br /> tags.

# Having looked at our data above, we see that the raw text contains HTML break
# tags of the form '<br />'. These tags will not be removed by the default
# standardizer (which doesn't strip HTML). Because of this, we will need to
# create a custom standardization function.
custom_standardization_fn <- function(string_tensor) {
  string_tensor |>
    tf$strings$lower() |> # convert to all lowercase
    tf$strings$regex_replace("<br />", " ") |> # remove '<br />' HTML tag
    tf$strings$regex_replace("[[:punct:]]", "") # remove punctuation
}


# Model constants.
max_features <- 20000
embedding_dim <- 128
sequence_length <- 500

# Now that we have our custom standardization, we can instantiate our text
# vectorization layer. We are using this layer to normalize, split, and map
# strings to integers, so we set our 'output_mode' to 'int'.
# Note that we're using the default split function,
# and the custom standardization defined above.
# We also set an explicit maximum sequence length, since the CNNs later in our
# model won't support ragged sequences.
vectorize_layer <- layer_text_vectorization(
  standardize = custom_standardization_fn,
  max_tokens = max_features,
  output_mode = "int",
  output_sequence_length = sequence_length,
)

# Now that the vectorize_layer has been created, call `adapt` on a text-only
# dataset to create the vocabulary. You don't have to batch, but for very large
# datasets this means you're not keeping spare copies of the dataset in memory.

# Let's make a text-only dataset (no labels):
text_ds <- raw_train_ds |>
  dataset_map(\(x, y) x)
# Let's call `adapt`:
vectorize_layer |> adapt(text_ds)

Two options to vectorize the data

There are 2 ways we can use our text vectorization layer:

Option 1: Make it part of the model, so as to obtain a model that processes raw strings, like this:

text_input <- keras_input(shape = c(1L), dtype = "string", name = 'text')
x <- text_input |>
  vectorize_layer() |>
  layer_embedding(max_features + 1, embedding_dim)

Option 2: Apply it to the text dataset to obtain a dataset of word indices, then feed it into a model that expects integer sequences as inputs.

An important difference between the two is that option 2 enables you to do asynchronous CPU processing and buffering of your data when training on GPU. So if you’re training the model on GPU, you probably want to go with this option to get the best performance. This is what we will do below.

If we were to export our model to production, we’d ship a model that accepts raw strings as input, like in the code snippet for option 1 above. This can be done after training. We do this in the last section.

vectorize_text <- function(text, label) {
  text <- text |>
    op_expand_dims(-1) |>
    vectorize_layer()
  list(text, label)
}

# Vectorize the data.
train_ds <- raw_train_ds |> dataset_map(vectorize_text)
val_ds   <- raw_val_ds   |> dataset_map(vectorize_text)
test_ds  <- raw_test_ds  |> dataset_map(vectorize_text)

# Do async prefetching / buffering of the data for best performance on GPU.
train_ds <- train_ds |>
  dataset_cache() |>
  dataset_prefetch(buffer_size = 10)
val_ds <- val_ds |>
  dataset_cache() |>
  dataset_prefetch(buffer_size = 10)
test_ds <- test_ds |>
  dataset_cache() |>
  dataset_prefetch(buffer_size = 10)

Build a model

We choose a simple 1D convnet starting with an Embedding layer.

# A integer input for vocab indices.
inputs <- keras_input(shape = c(NA), dtype = "int64")

predictions <- inputs |>
  # Next, we add a layer to map those vocab indices into a space of dimensionality
  # 'embedding_dim'.
  layer_embedding(max_features, embedding_dim) |>
  layer_dropout(0.5) |>
  # Conv1D + global max pooling
  layer_conv_1d(128, 7, padding = "valid", activation = "relu", strides = 3) |>
  layer_conv_1d(128, 7, padding = "valid", activation = "relu", strides = 3) |>
  layer_global_max_pooling_1d() |>
  # We add a vanilla hidden layer:
  layer_dense(128, activation = "relu") |>
  layer_dropout(0.5) |>
  # We project onto a single unit output layer, and squash it with a sigmoid:
  layer_dense(1, activation = "sigmoid", name = "predictions")

model <- keras_model(inputs, predictions)

summary(model)

## Model: "functional"
## ┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
## ┃ Layer (type)                    ┃ Output Shape           ┃       Param # ┃
## ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
## │ input_layer (InputLayer)        │ (None, None)           │             0 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ embedding_1 (Embedding)         │ (None, None, 128)      │     2,560,000 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout (Dropout)               │ (None, None, 128)      │             0 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ conv1d (Conv1D)                 │ (None, None, 128)      │       114,816 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ conv1d_1 (Conv1D)               │ (None, None, 128)      │       114,816 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ global_max_pooling1d            │ (None, 128)            │             0 │
## │ (GlobalMaxPooling1D)            │                        │               │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dense (Dense)                   │ (None, 128)            │        16,512 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ dropout_1 (Dropout)             │ (None, 128)            │             0 │
## ├─────────────────────────────────┼────────────────────────┼───────────────┤
## │ predictions (Dense)             │ (None, 1)              │           129 │
## └─────────────────────────────────┴────────────────────────┴───────────────┘
##  Total params: 2,806,273 (10.71 MB)
##  Trainable params: 2,806,273 (10.71 MB)
##  Non-trainable params: 0 (0.00 B)

# Compile the model with binary crossentropy loss and an adam optimizer.
model |> compile(loss = "binary_crossentropy",
                 optimizer = "adam",
                 metrics = "accuracy")

Train the model

epochs <- 3

# Fit the model using the train and test datasets.
model |> fit(train_ds, validation_data = val_ds, epochs = epochs)

## Epoch 1/3
## 625/625 - 6s - 10ms/step - accuracy: 0.6953 - loss: 0.5231 - val_accuracy: 0.8618 - val_loss: 0.3205
## Epoch 2/3
## 625/625 - 1s - 2ms/step - accuracy: 0.9032 - loss: 0.2390 - val_accuracy: 0.8742 - val_loss: 0.3113
## Epoch 3/3
## 625/625 - 2s - 2ms/step - accuracy: 0.9553 - loss: 0.1211 - val_accuracy: 0.8666 - val_loss: 0.3531

Evaluate the model on the test set

model |> evaluate(test_ds)

## 782/782 - 1s - 2ms/step - accuracy: 0.8520 - loss: 0.3960

## $accuracy
## [1] 0.85204
##
## $loss
## [1] 0.3959631

Make an end-to-end model

If you want to obtain a model capable of processing raw strings, you can simply create a new model (using the weights we just trained):

# A string input
inputs <- keras_input(shape = c(1), dtype = "string")
# Turn strings into vocab indices
indices <- vectorize_layer(inputs)
# Turn vocab indices into predictions
outputs <- model(indices)

# Our end to end model
end_to_end_model <- keras_model(inputs, outputs)
end_to_end_model |> compile(
  loss = "binary_crossentropy",
  optimizer = "adam",
  metrics = c("accuracy")
)

# Test it with `raw_test_ds`, which yields raw strings
end_to_end_model |> evaluate(raw_test_ds)

## 782/782 - 3s - 4ms/step - accuracy: 0.8520 - loss: 0.0000e+00

## $accuracy
## [1] 0.85204
##
## $loss
## [1] 0