def __init__(self,
depth,
bias=LSTMBiasInit,
initializer=block_util.RsqrtInitializer(),
name=None):
super(LSTM, self).__init__([depth], name)
with self._BlockScope():
self._depth = depth
self._nn = blocks_std.NN(
4 * depth, bias=bias, act=None, initializer=initializer)
self._hidden_linear = blocks_std.Linear(
4 * depth, initializer=initializer)
def testLinearShared(self):
# Create a linear map which is applied twice on different inputs
# (i.e. the weights of the map are shared).
linear_map = blocks_std.Linear(6)
x1 = tf.random_normal(shape=[1, 5])
x2 = tf.random_normal(shape=[1, 5])
xs = x1 + x2
# Apply the transform with the same weights.
y1 = linear_map(x1)
y2 = linear_map(x2)
ys = linear_map(xs)
with self.test_session() as sess:
# Initialize all the variables of the graph.
tf.global_variables_initializer().run()
y1_res, y2_res, ys_res = sess.run([y1, y2, ys])
self.assertAllClose(y1_res + y2_res, ys_res)
def testLinear(self): m = blocks_std.Linear(10) x = tf.placeholder(dtype=tf.float32, shape=[8, 9]) y = m(x) self.assertEqual(m._matrix.get_shape(), [9, 10]) self.assertIs(x, self.CheckLinear(y, m))