def test_static_dropout_rnn_cell():
sess = tf.Session()
x = np.random.randn(1, 10, 50).astype(np.float32)
with sess.graph.as_default():
with tf.variable_scope("DropoutIsOn"):
rnn_drop_cell = rnn_cell_w_dropout(100,
0.9999999999,
'gru',
training=True)
rnn_drop, _ = tf.nn.dynamic_rnn(rnn_drop_cell,
x,
sequence_length=np.array(
[10], dtype=np.int),
dtype=tf.float32)
with tf.variable_scope("DropoutIsOff"):
rnn_no_drop_cell = rnn_cell_w_dropout(100,
0.9999999999,
'gru',
training=False)
rnn_no_drop, _ = tf.nn.dynamic_rnn(rnn_no_drop_cell,
x,
sequence_length=np.array(
[10], dtype=np.int),
dtype=tf.float32)
sess.run(tf.global_variables_initializer())
out_ten = sess.run(rnn_drop)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) < 20
out_ten = sess.run(rnn_no_drop)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) > 20
def test_static_dropout_rnn_cell():
with tf.device('/cpu:0'):
sess = tf.Session()
x = np.random.randn(1, 10, 50).astype(np.float32)
with sess.graph.as_default():
with tf.variable_scope("DropoutIsOn"):
rnn_drop_cell = rnn_cell_w_dropout(100, 0.9999999999, 'gru', training=True)
rnn_drop, _ = tf.nn.dynamic_rnn(rnn_drop_cell, x, sequence_length=np.array([10], dtype=np.int), dtype=tf.float32)
with tf.variable_scope("DropoutIsOff"):
rnn_no_drop_cell = rnn_cell_w_dropout(100, 0.9999999999, 'gru', training=False)
rnn_no_drop, _ = tf.nn.dynamic_rnn(rnn_no_drop_cell, x, sequence_length=np.array([10], dtype=np.int), dtype=tf.float32)
sess.run(tf.global_variables_initializer())
out_ten = sess.run(rnn_drop)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) < 20
out_ten = sess.run(rnn_no_drop)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) > 20
def test_placeholder_dropout_rnn_cell():
sess = tf.Session()
x = np.random.randn(1, 10, 50).astype(np.float32)
with sess.graph.as_default():
train_flag = tf.placeholder_with_default(False, shape=(), name='TEST_TRAIN_FLAG')
with tf.variable_scope("DropoutMightBeOn"):
rnn_cell = rnn_cell_w_dropout(100, 0.9999999999, 'gru', training=train_flag)
rnn, _ = tf.nn.dynamic_rnn(rnn_cell, x, sequence_length=np.array([10], dtype=np.int), dtype=tf.float32)
sess.run(tf.global_variables_initializer())
out_ten = sess.run(rnn, {train_flag: True})
assert len(out_ten[np.nonzero(out_ten)].squeeze()) < 20
out_ten = sess.run(rnn)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) > 20
def test_placeholder_dropout_rnn_cell():
with tf.device('/cpu:0'):
sess = tf.Session()
x = np.random.randn(1, 10, 50).astype(np.float32)
with sess.graph.as_default():
train_flag = tf.placeholder_with_default(False, shape=(), name='TEST_TRAIN_FLAG')
with tf.variable_scope("DropoutMightBeOn"):
rnn_cell = rnn_cell_w_dropout(100, 0.9999999999, 'gru', training=train_flag)
rnn, _ = tf.nn.dynamic_rnn(rnn_cell, x, sequence_length=np.array([10], dtype=np.int), dtype=tf.float32)
sess.run(tf.global_variables_initializer())
out_ten = sess.run(rnn, {train_flag: True})
assert len(out_ten[np.nonzero(out_ten)].squeeze()) < 20
out_ten = sess.run(rnn)
assert len(out_ten[np.nonzero(out_ten)].squeeze()) > 20