def testGain(self):
shape = (10, 10)
for dtype in [dtypes.float32, dtypes.float64]:
init_default = init_ops.identity_initializer(dtype=dtype)
init_custom = init_ops.identity_initializer(gain=0.9, dtype=dtype)
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init_default(shape).eval(), np.eye(*shape))
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init_custom(shape).eval(), np.eye(*shape) * 0.9)
def testGain(self):
shape = (10, 10)
for dtype in [dtypes.float32, dtypes.float64]:
init_default = init_ops.identity_initializer(dtype=dtype)
init_custom = init_ops.identity_initializer(gain=0.9, dtype=dtype)
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init_default(shape).eval(), np.eye(*shape))
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init_custom(shape).eval(), np.eye(*shape) * 0.9)
def build(self, inputs_shape):
self._kernel_w = self.add_variable(
"{}_w".format(_WEIGHTS_VARIABLE_NAME),
[self.output_size, self.output_size],
dtype=self.dtype,
initializer=init_ops.identity_initializer(gain=.05))
self._kernel_c = self.add_variable(
"{}_c".format(_WEIGHTS_VARIABLE_NAME),
[inputs_shape[1], self.output_size],
dtype=self.dtype,
initializer=init_ops.identity_initializer(gain=.05))
if self._use_bias:
self._bias_c = self.add_variable(
"{}_c".format(_BIAS_VARIABLE_NAME), [self.output_size],
dtype=self.dtype)
def test_cell_call(self, *args, **kwargs):
"""Test that the state transition works properly on hidden-to-hidden states."""
batch_size = 2
num_units = 3
with self.test_session() as sess:
cell_0layer = stlstm.STLSTMCell(num_units, st_num_layers=0)
cell_1layer = stlstm.STLSTMCell(
num_units,
st_num_layers=1,
st_kernel_initializer=init_ops.identity_initializer())
init_state0 = cell_0layer.zero_state(batch_size,
dtype=dtypes.float32)
init_state1 = cell_1layer.zero_state(batch_size,
dtype=dtypes.float32)
inputs = random_ops.random_uniform([batch_size, 5])
h0, st0 = cell_0layer(inputs, init_state0)
h1, st1 = cell_1layer(inputs, init_state1)
variables.global_variables_initializer().run()
out0, newst0 = sess.run([h0, st0])
out1, newst1 = sess.run([h1, st1])
# NOTE: assert that a stlstm with 0 st layers output is the same as the new hidden state
self.assertAllEqual(out0, newst0[1])
# NOTE: assert the same for 1 layer st with an identity kernel and zero biases
self.assertAllEqual(out1, newst1[1])
def testPartitions(self):
shape = (10, 10)
init = init_ops.identity_initializer()
partitioner = partitioned_variables.variable_axis_size_partitioner(1)
with self.test_session(graph=ops.Graph(), use_gpu=True):
with variable_scope.variable_scope(
"foo", partitioner=partitioner, initializer=init):
v = array_ops.identity(variable_scope.get_variable("bar", shape=shape))
variables.global_variables_initializer().run()
self.assertAllClose(v.eval(), np.eye(*shape))
def testPartitions(self):
shape = (10, 10)
init = init_ops.identity_initializer()
partitioner = partitioned_variables.variable_axis_size_partitioner(1)
with self.test_session(graph=ops.Graph(), use_gpu=True):
with variable_scope.variable_scope(
"foo", partitioner=partitioner, initializer=init):
v = array_ops.identity(variable_scope.get_variable("bar", shape=shape))
variables.global_variables_initializer().run()
self.assertAllClose(v.eval(), np.eye(*shape))
def test_cell_residual(self, *args, **kwargs):
num_units = 3
# test residual and bias
# NOTE: test that 2 layer identity kernel with ones biases, new hidden state = output + output + 2
with self.test_session() as sess:
cell = stlstm.STLSTMCell(
num_units,
st_num_layers=2,
st_kernel_initializer=init_ops.identity_initializer(),
st_bias_initializer=init_ops.constant_initializer(1.),
st_residual=True)
init_state = cell.zero_state(2, dtype=dtypes.float32)
inputs = random_ops.random_uniform([2, 5])
h, st = cell(inputs, init_state)
variables.global_variables_initializer().run()
out, newst = sess.run([h, st])
self.assertAllClose(2. * out + 2, newst[1])
def test_cell_build(self, *args, **kwargs):
"""Check that state transition variables are created properly."""
num_units = 5
num_layers = 2
inputs_shape = [2, 8]
st_kernel_initializer = init_ops.identity_initializer()
# NOTE: test building with num_layers=0
with self.test_session():
cell = stlstm.STLSTMCell(num_units, st_num_layers=0)
self.assertIsNone(cell._st_kernels)
self.assertIsNone(cell._st_biases)
with self.test_session():
cell = stlstm.STLSTMCell(
num_units,
st_kernel_initializer=st_kernel_initializer,
st_num_layers=num_layers)
cell.build(inputs_shape)
variables.global_variables_initializer().run()
# check cell._st_kernels/biases
# NOTE: check length of new variable arrays
self.assertEqual(num_layers, len(cell._st_kernels))
self.assertEqual(len(cell._st_kernels), len(cell._st_biases))
# NOTE: check sizes of new variable arrays
for layer in range(num_layers):
self.assertAllEqual(cell._st_kernels[layer].shape,
[num_units, num_units])
self.assertEqual(cell._st_biases[layer].shape[0], num_units)
# NOTE: check variables are initialied correctly
for layer in range(num_layers):
self.assertAllEqual(cell._st_kernels[layer].eval(),
np.eye(num_units))
self.assertAllEqual(cell._st_biases[layer].eval(),
np.zeros(num_units))
def identity():
return init_ops.identity_initializer()
def testNonSquare(self):
init = init_ops.identity_initializer()
shape = (10, 5)
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init(shape).eval(), np.eye(*shape))
def testInvalidShape(self):
init = init_ops.identity_initializer()
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertRaises(ValueError, init, shape=[5, 7, 7])
self.assertRaises(ValueError, init, shape=[5])
self.assertRaises(ValueError, init, shape=[])
def testNonSquare(self):
init = init_ops.identity_initializer()
shape = (10, 5)
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertAllClose(init(shape).eval(), np.eye(*shape))
def testInvalidShape(self):
init = init_ops.identity_initializer()
with self.test_session(graph=ops.Graph(), use_gpu=True):
self.assertRaises(ValueError, init, shape=[5, 7, 7])
self.assertRaises(ValueError, init, shape=[5])
self.assertRaises(ValueError, init, shape=[])
