Torchtext 预处理自定义文本数据集

Author: Anupam Sharma

This tutorial illustrates the usage of torchtext on a dataset that is not built-in. In the tutorial, we will preprocess a dataset that can be further utilized to train a sequence-to-sequence model for machine translation (something like, in this tutorial: Sequence to Sequence Learning with Neural Networks) but without using legacy version of torchtext.

In this tutorial, we will learn how to:

• Read a dataset

• Tokenize sentence

• Apply transforms to sentence

• Perform bucket batching

Let us assume that we need to prepare a dataset to train a model that can perform English to German translation. We will use a tab-delimited German - English sentence pairs provided by the Tatoeba Project which can be downloaded from this link.

Sentence pairs for other languages can be found in this link.

Setup

First, download the dataset, extract the zip, and note the path to the file deu.txt.

Ensure that following packages are installed:

• Torchdata 0.6.0 (Installation instructions )

• Torchtext 0.15.0 (Installation instructions )

• Spacy

Here, we are using Spacy to tokenize text. In simple words tokenization means to convert a sentence to list of words. Spacy is a python package used for various Natural Language Processing (NLP) tasks.

Download the English and German models from Spacy as shown below:

python -m spacy download en_core_web_sm
python -m spacy download de_core_news_sm

Let us start by importing required modules:

import torchdata.datapipes as dp
import torchtext.transforms as T
import spacy
from torchtext.vocab import build_vocab_from_iterator
eng = spacy.load("en_core_web_sm") # Load the English model to tokenize English text
de = spacy.load("de_core_news_sm") # Load the German model to tokenize German text

Now we will load the dataset

FILE_PATH = 'data/deu.txt'
data_pipe = dp.iter.IterableWrapper([FILE_PATH])
data_pipe = dp.iter.FileOpener(data_pipe, mode='rb')
data_pipe = data_pipe.parse_csv(skip_lines=0, delimiter='\t', as_tuple=True)

In the above code block, we are doing following things:

1. At line 2, we are creating an iterable of filenames

2. At line 3, we pass the iterable to FileOpener which then opens the file in read mode

3. At line 4, we call a function to parse the file, which again returns an iterable of tuples representing each rows of the tab-delimited file

DataPipes can be thought of something like a dataset object, on which we can perform various operations. Check this tutorial for more details on DataPipes.

We can verify if the iterable has the pair of sentences as shown below:

for sample in data_pipe:
    print(sample)
    break

Note that we also have attribution details along with pair of sentences. We will write a small function to remove the attribution details:

def removeAttribution(row):
    """
    Function to keep the first two elements in a tuple
    """
    return row[:2]
data_pipe = data_pipe.map(removeAttribution)

The map function at line 6 in above code block can be used to apply some function on each elements of data_pipe. Now, we can verify that the data_pipe only contains pair of sentences.

for sample in data_pipe:
    print(sample)
    break

Now, let us define few functions to perform tokenization:

def engTokenize(text):
    """
    Tokenize an English text and return a list of tokens
    """
    return [token.text for token in eng.tokenizer(text)]

def deTokenize(text):
    """
    Tokenize a German text and return a list of tokens
    """
    return [token.text for token in de.tokenizer(text)]

Above function accepts a text and returns a list of words as shown below:

print(engTokenize("Have a good day!!!"))
print(deTokenize("Haben Sie einen guten Tag!!!"))

Building the vocabulary

Let us consider an English sentence as the source and a German sentence as the target.

Vocabulary can be considered as the set of unique words we have in the dataset. We will build vocabulary for both our source and target now.

Let us define a function to get tokens from elements of tuples in the iterator.

def getTokens(data_iter, place):
    """
    Function to yield tokens from an iterator. Since, our iterator contains
    tuple of sentences (source and target), `place` parameters defines for which
    index to return the tokens for. `place=0` for source and `place=1` for target
    """
    for english, german in data_iter:
        if place == 0:
            yield engTokenize(english)
        else:
            yield deTokenize(german)

Now, we will build vocabulary for source:

source_vocab = build_vocab_from_iterator(
    getTokens(data_pipe,0),
    min_freq=2,
    specials= ['<pad>', '<sos>', '<eos>', '<unk>'],
    special_first=True
)
source_vocab.set_default_index(source_vocab['<unk>'])

The code above, builds the vocabulary from the iterator. In the above code block:

• At line 2, we call the getTokens() function with place=0 as we need vocabulary for source sentences.

• At line 3, we set min_freq=2. This means, the function will skip those words that occurs less than 2 times.

• At line 4, we specify some special tokens:

  • <sos> for start of sentence

  • <eos> for end of sentence

  • <unk> for unknown words. An example of unknown word is the one skipped because of min_freq=2.

  • <pad> is the padding token. While training, a model we mostly train in batches. In a batch, there can be sentences of different length. So, we pad the shorter sentences with <pad> token to make length of all sequences in the batch equal.

• At line 5, we set special_first=True. Which means <pad> will get index 0, <sos> index 1, <eos> index 2, and <unk> will get index 3 in the vocabulary.

• At line 7, we set default index as index of <unk>. That means if some word is not in vocabulary, we will use <unk> instead of that unknown word.

Similarly, we will build vocabulary for target sentences:

target_vocab = build_vocab_from_iterator(
    getTokens(data_pipe,1),
    min_freq=2,
    specials= ['<pad>', '<sos>', '<eos>', '<unk>'],
    special_first=True
)
target_vocab.set_default_index(target_vocab['<unk>'])

Note that the example above shows how can we add special tokens to our vocabulary. The special tokens may change based on the requirements.

Now, we can verify that special tokens are placed at the beginning and then other words. In the below code, source_vocab.get_itos() returns a list with tokens at index based on vocabulary.

print(source_vocab.get_itos()[:9])

Numericalize sentences using vocabulary

After building the vocabulary, we need to convert our sentences to corresponding indices. Let us define some functions for this:

def getTransform(vocab):
    """
    Create transforms based on given vocabulary. The returned transform is applied to sequence
    of tokens.
    """
    text_tranform = T.Sequential(
        ## converts the sentences to indices based on given vocabulary
        T.VocabTransform(vocab=vocab),
        ## Add <sos> at beginning of each sentence. 1 because the index for <sos> in vocabulary is
        # 1 as seen in previous section
        T.AddToken(1, begin=True),
        ## Add <eos> at beginning of each sentence. 2 because the index for <eos> in vocabulary is
        # 2 as seen in previous section
        T.AddToken(2, begin=False)
    )
    return text_tranform

Now, let us see how to use the above function. The function returns an object of Transforms which we will use on our sentence. Let us take a random sentence and check how the transform works.

temp_list = list(data_pipe)
some_sentence = temp_list[798][0]
print("Some sentence=", end="")
print(some_sentence)
transformed_sentence = getTransform(source_vocab)(engTokenize(some_sentence))
print("Transformed sentence=", end="")
print(transformed_sentence)
index_to_string = source_vocab.get_itos()
for index in transformed_sentence:
    print(index_to_string[index], end=" ")

In the above code,:

• At line 2, we take a source sentence from list that we created from data_pipe at line 1

• At line 5, we get a transform based on a source vocabulary and apply it to a tokenized sentence. Note that transforms take list of words and not a sentence.

• At line 8, we get the mapping of index to string and then use it get the transformed sentence

Now we will use DataPipe functions to apply transform to all our sentences. Let us define some more functions for this.

def applyTransform(sequence_pair):
    """
    Apply transforms to sequence of tokens in a sequence pair
    """

    return (
        getTransform(source_vocab)(engTokenize(sequence_pair[0])),
        getTransform(target_vocab)(deTokenize(sequence_pair[1]))
    )
data_pipe = data_pipe.map(applyTransform) ## Apply the function to each element in the iterator
temp_list = list(data_pipe)
print(temp_list[0])

Make batches (with bucket batch)

Generally, we train models in batches. While working for sequence to sequence models, it is recommended to keep the length of sequences in a batch similar. For that we will use bucketbatch function of data_pipe.

Let us define some functions that will be used by the bucketbatch function.

def sortBucket(bucket):
    """
    Function to sort a given bucket. Here, we want to sort based on the length of
    source and target sequence.
    """
    return sorted(bucket, key=lambda x: (len(x[0]), len(x[1])))

Now, we will apply the bucketbatch function:

data_pipe = data_pipe.bucketbatch(
    batch_size = 4, batch_num=5,  bucket_num=1,
    use_in_batch_shuffle=False, sort_key=sortBucket
)

In the above code block:

• We keep batch size = 4.

• batch_num is the number of batches to keep in a bucket

• bucket_num is the number of buckets to keep in a pool for shuffling

• sort_key specifies the function that takes a bucket and sorts it

Now, let us consider a batch of source sentences as X and a batch of target sentences as y. Generally, while training a model, we predict on a batch of X and compare the result with y. But, a batch in our data_pipe is of the form [(X_1,y_1), (X_2,y_2), (X_3,y_3), (X_4,y_4)]:

print(list(data_pipe)[0])

So, we will now convert them into the form: ((X_1,X_2,X_3,X_4), (y_1,y_2,y_3,y_4)). For this we will write a small function:

def separateSourceTarget(sequence_pairs):
    """
    input of form: `[(X_1,y_1), (X_2,y_2), (X_3,y_3), (X_4,y_4)]`
    output of form: `((X_1,X_2,X_3,X_4), (y_1,y_2,y_3,y_4))`
    """
    sources,targets = zip(*sequence_pairs)
    return sources,targets

## Apply the function to each element in the iterator
data_pipe = data_pipe.map(separateSourceTarget)
print(list(data_pipe)[0])

Now, we have the data as desired.

Padding

As discussed earlier while building vocabulary, we need to pad shorter sentences in a batch to make all the sequences in a batch of equal length. We can perform padding as follows:

def applyPadding(pair_of_sequences):
    """
    Convert sequences to tensors and apply padding
    """
    return (T.ToTensor(0)(list(pair_of_sequences[0])), T.ToTensor(0)(list(pair_of_sequences[1])))
## `T.ToTensor(0)` returns a transform that converts the sequence to `torch.tensor` and also applies
# padding. Here, `0` is passed to the constructor to specify the index of the `<pad>` token in the
# vocabulary.
data_pipe = data_pipe.map(applyPadding)

Now, we can use the index to string mapping to see how the sequence would look with tokens instead of indices:

source_index_to_string = source_vocab.get_itos()
target_index_to_string = target_vocab.get_itos()

def showSomeTransformedSentences(data_pipe):
    """
    Function to show how the sentences look like after applying all transforms.
    Here we try to print actual words instead of corresponding index
    """
    for sources,targets in data_pipe:
        if sources[0][-1] != 0:
            continue # Just to visualize padding of shorter sentences
        for i in range(4):
            source = ""
            for token in sources[i]:
                source += " " + source_index_to_string[token]
            target = ""
            for token in targets[i]:
                target += " " + target_index_to_string[token]
            print(f"Source: {source}")
            print(f"Traget: {target}")
        break

showSomeTransformedSentences(data_pipe)

In the above output we can observe that the shorter sentences are padded with <pad>. Now, we can use data_pipe while writing our training function.

Some parts of this tutorial was inspired from this article.