High-level RNN Keras (CNTK) Example¶

In [1]:

import os
import sys
import numpy as np
os.environ['KERAS_BACKEND'] = "cntk"
import keras as K
import cntk
from keras.models import Sequential
from keras.layers import Dense, Embedding, GRU, CuDNNGRU
from common.params_lstm import *
from common.utils import *

Using CNTK backend

In [2]:

print("OS: ", sys.platform)
print("Python: ", sys.version)
print("Keras: ", K.__version__)
print("Numpy: ", np.__version__)
print("CNTK: ", cntk.__version__)
print(K.backend.backend())

OS:  linux
Python:  3.5.2 |Anaconda custom (64-bit)| (default, Jul  2 2016, 17:53:06) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]
Keras:  2.1.1
Numpy:  1.13.3
CNTK:  2.2
cntk

In [3]:

def create_symbol(CUDNN=True):
    model = Sequential()
    model.add(Embedding(MAXFEATURES, EMBEDSIZE, input_length=MAXLEN))
    # Only return last output
    if not CUDNN:
        model.add(GRU(NUMHIDDEN, return_sequences=False, return_state=False))
    else:
        model.add(CuDNNGRU(NUMHIDDEN, return_sequences=False, return_state=False))
    model.add(Dense(2, activation='softmax'))
    return model

In [4]:

def init_model(m):
    m.compile(
        loss = "categorical_crossentropy",
        optimizer = K.optimizers.Adam(LR, BETA_1, BETA_2, EPS),
        metrics = ['accuracy'])
    return m

In [5]:

%%time
# Data into format for library
x_train, x_test, y_train, y_test = imdb_for_library(seq_len=MAXLEN, max_features=MAXFEATURES, one_hot=True)
print(x_train.shape, x_test.shape, y_train.shape, y_test.shape)
print(x_train.dtype, x_test.dtype, y_train.dtype, y_test.dtype)

Preparing train set...
Preparing test set...
Trimming to 30000 max-features
Padding to length 150
(25000, 150) (25000, 150) (25000, 2) (25000, 2)
int32 int32 int32 int32
CPU times: user 5.3 s, sys: 278 ms, total: 5.58 s
Wall time: 5.58 s

In [7]:

%%time
# Load symbol
# CuDNN RNNs are only available with the TensorFlow backend.
sym = create_symbol(CUDNN=False)

CPU times: user 137 ms, sys: 107 ms, total: 244 ms
Wall time: 111 ms

In [8]:

%%time
# Initialise model
model = init_model(sym)

CPU times: user 7.83 ms, sys: 0 ns, total: 7.83 ms
Wall time: 7.8 ms

In [9]:

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_2 (Embedding)      (None, 150, 125)          3750000   
_________________________________________________________________
gru_1 (GRU)                  (None, 100)               67800     
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 202       
=================================================================
Total params: 3,818,002
Trainable params: 3,818,002
Non-trainable params: 0
_________________________________________________________________

In [10]:

%%time
# 86s
# Train model
model.fit(x_train,
          y_train,
          batch_size=BATCHSIZE,
          epochs=EPOCHS,
          verbose=1)

Epoch 1/3
25000/25000 [==============================] - 29s 1ms/step - loss: 0.4392 - acc: 0.7803
Epoch 2/3
25000/25000 [==============================] - 29s 1ms/step - loss: 0.2057 - acc: 0.9239
Epoch 3/3
25000/25000 [==============================] - 29s 1ms/step - loss: 0.1104 - acc: 0.9631
CPU times: user 1min 23s, sys: 3.78 s, total: 1min 27s
Wall time: 1min 26s

Out[10]:

<keras.callbacks.History at 0x7f33482c9550>

In [11]:

%%time
y_guess = model.predict(x_test, batch_size=BATCHSIZE)
y_guess = np.argmax(y_guess, axis=-1)
y_truth = np.argmax(y_test, axis=-1)

CPU times: user 10.2 s, sys: 236 ms, total: 10.4 s
Wall time: 10.4 s

In [12]:

print("Accuracy: ", sum(y_guess == y_truth)/len(y_guess))

Accuracy:  0.86092