如何在 tf 数据集上应用文本矢量化后创建滑动 window？

Question

I am reading a large text file using TensorFlow's TextLineDataset . I want to tokenize the dataset and create a sliding window and separate the tokenized text into two parts - input and label. If the text file has the following texts:

Lorem ipsum dolor sit amet...

then I want to create sequences of a specified length pre-padded with 0's. I want to iterate over the text and use all but the last as input and the last one as the label. So, my target is to first tokenize the texts as something like this:

Lorem: 1,
ipsum: 2,
dolor: 3,
sit: 4,
amet: 5,
...

Then create a sequence of let's say a length of 5 like this to train a model:

X_train = [[0, 0, 0, 0, 1], [0, 0, 0, 1, 2], [0, 0, 1, 2, 3], ...]
y_train = [2, 3, 4, ...] # next word of the sequence in X_train

I am using TextVectorization to tokenize but cannot figure out an efficient way to create the inputs and labels for a large dataset.

vectorize_layer = tf.keras.layers.TextVectorization(output_mode='int',
                                                    max_tokens=MAX_WORDS,
                                                    output_sequence_length=MAX_SEQUENCE_LENGTH)
vectorize_layer.adapt(train_data)
train_data = train_data.map(vectorize_layer)

Using a for loop over the dataset would make the device run out of memory trying to allocate a large amount of memory. What is the best way to do this?

Answer 1

You could use the sliding window function from tensorflow-text ; however, the TextVectorization layer seems to only apply post-padding:

import tensorflow as tf
import tensorflow_text as tft

with open('data.txt', 'w') as f:
  f.write('Lorem ipsum dolor sit amet, consectetur adipiscing elit. Aliquam efficitur viverra lacus?\n')

train_data = tf.data.TextLineDataset(['/content/data.txt'])

vectorize_layer = tf.keras.layers.TextVectorization(output_mode='int', max_tokens=50, pad_to_max_tokens=True)
vectorize_layer.adapt(train_data)

window_size = 5

def sliding_window(x):
  encoded = vectorize_layer(x)
  x = tft.sliding_window(encoded, width=window_size, axis=0)
  y = tft.sliding_window(encoded, width=window_size + 1, axis=0)[:, -1]
  return x[:tf.shape(y)[0],:], y

train_data = train_data.map(sliding_window)


vocab = tf.constant(vectorize_layer.get_vocabulary())
keys = tf.cast(tf.range(vocab.shape[0]), tf.int64)
table = tf.lookup.StaticHashTable(
    tf.lookup.KeyValueTensorInitializer(keys, vocab),
    default_value="")

train_data = tf.data.Dataset.zip((train_data.map(lambda x, y: x).flat_map(tf.data.Dataset.from_tensor_slices),
                                 train_data.map(lambda x, y: y).flat_map(tf.data.Dataset.from_tensor_slices)))

for x, y in train_data:
  print('x -->', x, 'y -->', y)
  print('x -->', table.lookup(x), 'y -->', table.lookup(y), '\n')

x --> tf.Tensor([ 4  6  9  3 11], shape=(5,), dtype=int64) y --> tf.Tensor(10, shape=(), dtype=int64)
x --> tf.Tensor([b'lorem' b'ipsum' b'dolor' b'sit' b'amet'], shape=(5,), dtype=string) y --> tf.Tensor(b'consectetur', shape=(), dtype=string) 

x --> tf.Tensor([ 6  9  3 11 10], shape=(5,), dtype=int64) y --> tf.Tensor(13, shape=(), dtype=int64)
x --> tf.Tensor([b'ipsum' b'dolor' b'sit' b'amet' b'consectetur'], shape=(5,), dtype=string) y --> tf.Tensor(b'adipiscing', shape=(), dtype=string) 

x --> tf.Tensor([ 9  3 11 10 13], shape=(5,), dtype=int64) y --> tf.Tensor(7, shape=(), dtype=int64)
x --> tf.Tensor([b'dolor' b'sit' b'amet' b'consectetur' b'adipiscing'], shape=(5,), dtype=string) y --> tf.Tensor(b'elit', shape=(), dtype=string) 

x --> tf.Tensor([ 3 11 10 13  7], shape=(5,), dtype=int64) y --> tf.Tensor(12, shape=(), dtype=int64)
x --> tf.Tensor([b'sit' b'amet' b'consectetur' b'adipiscing' b'elit'], shape=(5,), dtype=string) y --> tf.Tensor(b'aliquam', shape=(), dtype=string) 

x --> tf.Tensor([11 10 13  7 12], shape=(5,), dtype=int64) y --> tf.Tensor(8, shape=(), dtype=int64)
x --> tf.Tensor([b'amet' b'consectetur' b'adipiscing' b'elit' b'aliquam'], shape=(5,), dtype=string) y --> tf.Tensor(b'efficitur', shape=(), dtype=string) 

x --> tf.Tensor([10 13  7 12  8], shape=(5,), dtype=int64) y --> tf.Tensor(2, shape=(), dtype=int64)
x --> tf.Tensor([b'consectetur' b'adipiscing' b'elit' b'aliquam' b'efficitur'], shape=(5,), dtype=string) y --> tf.Tensor(b'viverra', shape=(), dtype=string) 

x --> tf.Tensor([13  7 12  8  2], shape=(5,), dtype=int64) y --> tf.Tensor(5, shape=(), dtype=int64)
x --> tf.Tensor([b'adipiscing' b'elit' b'aliquam' b'efficitur' b'viverra'], shape=(5,), dtype=string) y --> tf.Tensor(b'lacus', shape=(), dtype=string) 

Note that sequences that do not have a corresponding label are discarded with the line x[:tf.shape(y)[0],:] . Also, the lookup table is only for demonstration purposes and not needed to achieve what you want. You can look at tft.pad_along_dimension if you want to apply pre-padding.