In [0]:
# Based on
# https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/6.2-understanding-recurrent-neural-networks.ipynb
In [0]:
import warnings
warnings.filterwarnings('ignore')
In [3]:
%matplotlib inline
%pylab inline
In [5]:
import tensorflow as tf
tf.logging.set_verbosity(tf.logging.ERROR)
print(tf.__version__)
In [6]:
# let's see what compute devices we have available, hopefully a GPU
sess = tf.Session()
devices = sess.list_devices()
for d in devices:
print(d.name)
In [7]:
# a small sanity check, does tf seem to work ok?
hello = tf.constant('Hello TF!')
print(sess.run(hello))
In [8]:
from tensorflow import keras
print(keras.__version__)
In [9]:
# https://keras.io/datasets/#imdb-movie-reviews-sentiment-classification
max_features = 10000 # number of words to consider as features
maxlen = 500 # cut texts after this number of words (among top max_features most common words)
# each review is encoded as a sequence of word indexes
# indexed by overall frequency in the dataset
# output is 0 (negative) or 1 (positive)
imdb = tf.keras.datasets.imdb.load_data(num_words=max_features)
(raw_input_train, y_train), (raw_input_test, y_test) = imdb
In [0]:
# https://www.tensorflow.org/api_docs/python/tf/keras/preprocessing/sequence/pad_sequences
input_train = tf.keras.preprocessing.sequence.pad_sequences(raw_input_train, maxlen=maxlen)
input_test = tf.keras.preprocessing.sequence.pad_sequences(raw_input_test, maxlen=maxlen)
In [11]:
input_train.shape, input_test.shape, y_train.shape, y_test.shape
Out[11]:
In [0]:
# tf.keras.layers.SimpleRNN?
# tf.keras.layers.Embedding?
In [13]:
embedding_dim = 32
model = tf.keras.Sequential()
# Parameters: max_features (10000) * 8 = 80000
model.add(tf.keras.layers.Embedding(name='embedding', input_dim=max_features, output_dim=embedding_dim, input_length=maxlen))
# model.add(tf.keras.layers.Embedding(max_features, 32, input_length=maxlen))
# model.add(tf.keras.layers.SimpleRNN(32, return_sequences=True))
# model.add(tf.keras.layers.SimpleRNN(32, return_sequences=True))
# https://arxiv.org/ftp/arxiv/papers/1701/1701.05923.pdf
# n = output dimension
# m = input dimension
# Total number of parameters for
# RNN = n**2 + nm (like fc) + n (bias)
# n = 1, m =32: 1 + 32 + 1 = 34
# model.add(tf.keras.layers.SimpleRNN(name='rnn', units=1))
# n = 32, m =32: 1024 + 1024 + 32 = 2080
model.add(tf.keras.layers.SimpleRNN(name='rnn1', units=32, return_sequences=True))
model.add(tf.keras.layers.SimpleRNN(name='rnn2', units=32))
# Input format: maxlen (500) * dimension of embedding (8)
# Output: 4000
# model.add(tf.keras.layers.Flatten())
# binary classifier
model.add(tf.keras.layers.Dense(name='fc', units=32, activation='relu'))
model.add(tf.keras.layers.Dense(name='classifier', units=1, activation='sigmoid'))
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
model.summary()
In [14]:
batch_size = 1000
%time history = model.fit(input_train, y_train, epochs=10, batch_size=batch_size, validation_split=0.2)
In [15]:
train_loss, train_accuracy = model.evaluate(input_train, y_train, batch_size=batch_size)
train_accuracy
Out[15]:
In [16]:
test_loss, test_accuracy = model.evaluate(input_test, y_test, batch_size=batch_size)
test_accuracy
Out[16]:
In [17]:
# precition
model.predict(input_test[0:5])
Out[17]:
In [18]:
# ground truth
y_test[0:5]
Out[18]:
In [0]: