Loading keras imdb dataset

Loading Keras IMDB dataset

import numpy as np
import pandas as pd

from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer

This downloads 25.000 movie reviews from IMDB with the label positive/negative.

Each review is encoded as a list of indexes (integers)

Words are encoded by overall frequency in the dataset. Integer 3 encodes the 3rd most frequent word in the data.

Create data

# set number of features (use top 1000 words)
num_words = 1000
index_from = 3

# load data
(x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=num_words, index_from=index_from)

Convert word to index (and reverse)

Note that by default indexing starts from 3 in the default settings as follows.

start_char=1, oov_char=2, index_from=3,

This means that:

  • 0 is the padding char
  • 1 is the start char
  • 2 is oov
  • 3 is unknown

So we need to load in the mapping and then shift it 3 to the right, very convenient.

# Get word index (word => index)
word_to_id = imdb.get_word_index()
word_to_id = {k:(v+index_from) for k,v in word_to_id.items()}
word_to_id["<PAD>"] = 0
word_to_id["<START>"] = 1
word_to_id["<UNK>"] = 2
word_to_id["<UNUSED>"] = 3

# Reverse word index (index => word)
id_to_word = dict([(value, key) for (key, value) in word_to_id.items()])

View data

" ".join([id_to_word[i] for i in x_train[1000]])
"<START> although i had seen <UNK> in a theater way back in <UNK> i couldn't remember anything of the plot except for <UNK> <UNK> of <UNK> <UNK> running and fighting against a <UNK> of <UNK> <UNK> and <UNK> <UNK> the ending br br after reading some of the other reviews i <UNK> up a <UNK> of the <UNK> released dvd to once again <UNK> the world of <UNK> br br it turns out this is one of those films <UNK> during the <UNK> that would go <UNK> to video today the film stars <UNK> <UNK> <UNK> <UNK> as <UNK> <UNK> <UNK> out of the <UNK> to <UNK> the <UNK> of <UNK> to <UNK> and <UNK> <UNK> the game a <UNK> <UNK> <UNK> by the <UNK> who <UNK> his people by <UNK> what sounds like <UNK> power the <UNK> of the <UNK> <UNK> the star <UNK> <UNK> <UNK> <UNK> is <UNK> in the <UNK> <UNK> by <UNK> <UNK> who never <UNK> or leaves the house once <UNK> tries to <UNK> in with the <UNK> by <UNK> a <UNK> red <UNK> with <UNK> of <UNK> and white <UNK> to say <UNK> finds himself running and fighting for his life along the <UNK> <UNK> of <UNK> on his way to a <UNK> with <UNK> and the game br br star <UNK> <UNK> was <UNK> <UNK> by director robert <UNK> who it looks like was never on the set the so called script is just this side of <UNK> see other reviews for the many <UNK> throughout the town of <UNK> has a few good moments but is <UNK> <UNK> by bad editing the ending <UNK> still there's the <UNK> of a good action <UNK> here a <UNK> <UNK> version with more <UNK> action and <UNK> <UNK> might even be pretty good"


# Convert list of integers to one-hot matrix
tokenizer = Tokenizer(num_words=num_words)
# Convert list of integers into matrix with max length of dict
train_features = tokenizer.sequences_to_matrix(x_train, mode='binary')
test_features = tokenizer.sequences_to_matrix(x_test, mode='binary')

print(f"train features shape: {train_features.shape}, test features shape {test_features.shape}") 
train features shape: (25000, 1000), test features shape (25000, 1000)

Now you have:

  • (25000, 1000) train features ohe matrix
  • (25000, 1000) test features ohe matrix
  • (25000,) train labels
  • (25000,) test labels