In [1]:

    

%load_ext autoreload
%autoreload 2


    

In [2]:

    

import numpy as np
import tensorflow as tf

from __future__ import print_function


    

In [3]:

    

def create_examples(N, batch_size):
    A = np.random.binomial(n=1, p=0.5, size=(batch_size, N))
    B = np.random.binomial(n=1, p=0.5, size=(batch_size, N,))

    X = np.zeros((batch_size, 2 *N,), dtype=np.float32)
    X[:,:N], X[:,N:] = A, B

    Y = (A ^ B).astype(np.float32)
    return X,Y


    

In [4]:

    

from tf_rl.models import MLP


    

In [5]:

    

tf.ops.reset_default_graph()
sess = tf.InteractiveSession()


    

In [6]:

    

N = 5
# we add a single hidden layer of size 12
# otherwise code is similar to above
HIDDEN_SIZE = 12

x = tf.placeholder(tf.float32, (None, 2 * N), name="x")
y_golden = tf.placeholder(tf.float32, (None, N), name="y")

mlp = MLP(2 * N, [HIDDEN_SIZE, N], [tf.tanh, tf.sigmoid])
y = mlp(x)

cost = tf.reduce_mean(tf.square(y - y_golden))

optimizer = tf.train.AdagradOptimizer(learning_rate=0.3)
train_op = optimizer.minimize(cost)
sess.run(tf.initialize_all_variables())


    

In [7]:

    

mlp2 = mlp.copy(sess)


    

In [8]:

    

for t in range(5000):
    example_x, example_y = create_examples(N, 10)
    cost_t, _ = sess.run([cost, train_op], {x: example_x, y_golden: example_y})
    if t % 500 == 0: 
        print(cost_t.mean())


    

    




0.25368
0.24415
0.170396
0.0975643
0.0823987
0.0314237
0.0417636
0.0364044
0.039408
0.026788

In [9]:

    

N_EXAMPLES = 1000
example_x, example_y = create_examples(N, N_EXAMPLES)
is_correct = tf.less_equal(tf.abs(y - y_golden), tf.constant(0.5))
accuracy = tf.reduce_mean(tf.cast(is_correct, "float"))

acc_result = sess.run(accuracy, {x: example_x, y_golden: example_y})
print("Accuracy over %d examples: %.0f %%" % (N_EXAMPLES, 100.0 * acc_result))


    

    




Accuracy over 1000 examples: 99 %

In [10]:

    

# If copy works accuracy should be around 50% for this one
N_EXAMPLES = 1000
example_x, example_y = create_examples(N, N_EXAMPLES)
is_correct = tf.less_equal(tf.abs(mlp2(x) - y_golden), tf.constant(0.5))
accuracy = tf.reduce_mean(tf.cast(is_correct, "float"))

acc_result = sess.run(accuracy, {x: example_x, y_golden: example_y})
print("Accuracy over %d examples: %.0f %%" % (N_EXAMPLES, 100.0 * acc_result))


    

    




Accuracy over 1000 examples: 52 %

In [ ]: