def test_dnc_optimization(self):
batch_size = 7
time_steps = 15
input_size = 30
controller_config = {
"hidden_size": 64,
}
memory_config = {
'read_heads_num': 10,
'word_size': 9,
'words_num': 27,
}
output_size = 36
dnc = DNC(controller_config,
memory_config,
output_size,
classic_dnc_output=False)
dnc_initial_state = dnc.initial_state(batch_size)
inputs = tf.random_normal([time_steps, batch_size, input_size])
dnc_output_op, _ = rnn.dynamic_rnn(cell=dnc,
inputs=inputs,
initial_state=dnc_initial_state,
time_major=True)
targets = np.random.rand(time_steps, batch_size, output_size)
loss = tf.reduce_mean(tf.square(dnc_output_op - targets))
optimizier_op = tf.train.GradientDescentOptimizer(5).minimize(loss)
init_op = tf.global_variables_initializer()
with self.test_session():
init_op.run()
optimizier_op.run()
def test_dnc_output_shape(self):
batch_size = 3
controller_config = {
"hidden_size": 64,
}
memory_config = {
'read_heads_num': 7,
'word_size': 5,
'words_num': 16,
}
output_size = 10
for input_size in [10, 17, 49]:
dnc = DNC(controller_config,
memory_config,
output_size,
classic_dnc_output=False)
initial_state = dnc.initial_state(batch_size)
input_shape = dnc._W * dnc._R + input_size
test_input = np.random.uniform(
-3, 3, (batch_size, input_shape)).astype(np.float32)
example_output_op, _ = dnc(
tf.convert_to_tensor(test_input),
initial_state,
)
init = tf.global_variables_initializer()
with self.test_session() as sess:
init.run()
example_output = sess.run(example_output_op)
self.assertEqual(example_output.shape, (batch_size, output_size))
