def test_evaluate_is_batch_normalized(self):
Y1 = np.ones((4, 1, 2))
T1 = create_targets_object(np.ones((4, 1, 2)) * 2)
Y2 = np.ones((4, 10, 2))
T2 = create_targets_object(np.ones((4, 10, 2)) * 2)
for err in self.error_funcs:
e1, d1 = err(Y1, T1)
e2, d2 = err(Y2, T2)
self.assertAlmostEqual(e1, e2)
def check_deltas(net,
X=None,
T=None,
n_timesteps=3,
n_batches=5,
rnd=np.random.RandomState()):
if X is None:
X = rnd.randn(n_timesteps, n_batches, net.get_input_size())
if T is None:
T = rnd.randn(n_timesteps, n_batches, net.get_output_size())
# normalize targets to sum to one
T = T / T.sum(2).reshape(n_timesteps, n_batches, 1)
T = create_targets_object(T)
######### calculate gradient ##########
net.forward_pass(X)
delta_calc = net.backward_pass(T).flatten()
######### estimate gradient ##########
def f(X):
net.forward_pass(X.reshape(n_timesteps, n_batches, -1))
return net.calculate_error(T)
delta_approx = approx_fprime(X.copy().flatten(), f, 1e-7)
return np.sum(
(delta_approx - delta_calc)**2) / n_batches, delta_calc, delta_approx
def check_gradient(net,
X=None,
T=None,
n_timesteps=3,
n_batches=5,
rnd=np.random.RandomState()):
if X is None:
X = rnd.randn(n_timesteps, n_batches, net.get_input_size())
if T is None:
T = rnd.randn(n_timesteps, n_batches, net.get_output_size())
# normalize targets to sum to one
T = T / T.sum(2).reshape(n_timesteps, n_batches, 1)
T = create_targets_object(T)
weights = net.param_buffer.copy()
######### calculate gradient ##########
net.forward_pass(X)
net.backward_pass(T)
grad_calc = net.calc_gradient().squeeze().copy()
######### estimate gradient ##########
def f(W):
net.param_buffer = W
net.forward_pass(X)
return net.calculate_error(T).copy()
grad_approx = approx_fprime(weights.copy(), f, 1e-7)
return np.sum(
(grad_approx - grad_calc)**2) / n_batches, grad_calc, grad_approx
def test_shuffle_is_random(self):
X = np.arange(10).reshape(1, -1, 1)
T = create_targets_object(np.arange(10).reshape(1, -1, 1))
global_rnd.set_seed(1)
_, _, s1 = shuffle_data(X, T)
_, _, s2 = shuffle_data(X, T)
self.assertFalse(np.all(s1 == s2))
def test_shuffle_seed_overwrites_global_seed(self):
X = np.arange(10).reshape(1, -1, 1)
T = create_targets_object(np.arange(10).reshape(1, -1, 1))
global_rnd.set_seed(1)
_, _, s1 = shuffle_data(X, T, seed=1)
global_rnd.set_seed(1)
_, _, s2 = shuffle_data(X, T, seed=2)
self.assertFalse(np.all(s1 == s2))
def test_shuffle_depends_on_global_seed(self):
X = np.arange(10).reshape(1, -1, 1)
T = create_targets_object(np.arange(10).reshape(1, -1, 1))
global_rnd.set_seed(1)
_, _, s1 = shuffle_data(X, T)
global_rnd.set_seed(1)
_, _, s2 = shuffle_data(X, T)
np.testing.assert_array_equal(s1.flat, s2.flat)
def test_finite_differences(self):
Y = np.zeros((4, 3, 2)) + 0.5
T = create_targets_object(np.ones((4, 3, 2)))
for err in self.error_funcs:
def f(X):
return err(X.reshape(*Y.shape), T)[0]
delta_approx = approx_fprime(Y.flatten().copy(), f, 1e-7)
delta_calc = err(Y, T)[1].flatten()
np.testing.assert_array_almost_equal(delta_approx, delta_calc)
def read_data(candidates, targets='T'):
input_data = load_data(get_files_containing(candidates, 'X'))
targets_data = load_data(get_files_containing(candidates, targets))
mask_candidates = get_files_containing(candidates, 'M')
if mask_candidates:
mask = load_data(mask_candidates)
else:
mask = None
targets = create_targets_object(targets_data, mask)
return input_data, targets
def transform_ds_to_nsp(ds):
"""
Takes a dataset dictionary like the one returned from load_dataset
and transforms it into a next-step-prediction task.
"""
ds_nsp = {}
for use in ['training', 'validation', 'test']:
if use not in ds or ds[use] is None:
continue
nsp_targets = create_targets_object(ds[use][0][1:])
ds_nsp[use] = (ds[use][0][:-1, :, :],
nsp_targets)
return ds_nsp
def _labeling_binarizing(outputs, targets):
# TODO: use mask to mask deltas
time_size, batch_size, label_count = outputs.shape
deltas = np.zeros((time_size, batch_size, label_count))
deltas[:] = float('-inf')
errors = np.zeros(batch_size)
targets = create_targets_object(targets)
for b, (y, t) in enumerate(Online(outputs, targets, verbose=False)()):
err, delt = ctcpp(y, list(t.data[0]))
errors[b] = err
deltas[:y.shape[0], b:b+1, :] = delt.as_array()
return np.mean(errors), -deltas / batch_size
def setUp(self):
self.input_size = 2
self.output_size = 3
self.layer_types = [ForwardLayer, RnnLayer, MrnnLayer, LstmLayer,
Lstm97Layer]
self.activation_functions = ["linear", "tanh", "tanhx2", "sigmoid",
"softmax"]
n_timesteps = 5
n_batches = 6
self.input_data = rnd.randn(n_timesteps, n_batches, self.input_size)
self.targets = rnd.randn(n_timesteps, n_batches, self.output_size)
self.targets = self.targets / self.targets.sum(2).reshape(n_timesteps,
n_batches,
1)
self.targets = create_targets_object(self.targets)
def setUp(self):
self.input_size = 2
self.output_size = 3
self.layer_types = [
ForwardLayer, RnnLayer, MrnnLayer, LstmLayer, Lstm97Layer
]
self.activation_functions = [
"linear", "tanh", "tanhx2", "sigmoid", "softmax"
]
n_timesteps = 5
n_batches = 6
self.input_data = rnd.randn(n_timesteps, n_batches, self.input_size)
self.targets = rnd.randn(n_timesteps, n_batches, self.output_size)
self.targets = self.targets / self.targets.sum(2).reshape(
n_timesteps, n_batches, 1)
self.targets = create_targets_object(self.targets)
def check_deltas(net, X=None, T=None, n_timesteps=3, n_batches=5,
rnd=np.random.RandomState()):
if X is None:
X = rnd.randn(n_timesteps, n_batches, net.get_input_size())
if T is None:
T = rnd.randn(n_timesteps, n_batches, net.get_output_size())
# normalize targets to sum to one
T = T / T.sum(2).reshape(n_timesteps, n_batches, 1)
T = create_targets_object(T)
######### calculate gradient ##########
net.forward_pass(X)
delta_calc = net.backward_pass(T).flatten()
######### estimate gradient ##########
def f(X):
net.forward_pass(X.reshape(n_timesteps, n_batches, -1))
return net.calculate_error(T)
delta_approx = approx_fprime(X.copy().flatten(), f, 1e-7)
return np.sum((delta_approx - delta_calc) ** 2) / n_batches, delta_calc, delta_approx
def check_gradient(net, X=None, T=None, n_timesteps=3, n_batches=5,
rnd=np.random.RandomState()):
if X is None:
X = rnd.randn(n_timesteps, n_batches, net.get_input_size())
if T is None:
T = rnd.randn(n_timesteps, n_batches, net.get_output_size())
# normalize targets to sum to one
T = T / T.sum(2).reshape(n_timesteps, n_batches, 1)
T = create_targets_object(T)
weights = net.param_buffer.copy()
######### calculate gradient ##########
net.forward_pass(X)
net.backward_pass(T)
grad_calc = net.calc_gradient().squeeze().copy()
######### estimate gradient ##########
def f(W):
net.param_buffer = W
net.forward_pass(X)
return net.calculate_error(T).copy()
grad_approx = approx_fprime(weights.copy(), f, 1e-7)
return np.sum((grad_approx - grad_calc) ** 2) / n_batches, grad_calc, grad_approx
def test_deriv_shape(self):
Y = np.ones((4, 3, 2))
T = create_targets_object(np.ones((4, 3, 2)) * 2)
for err in self.error_funcs:
e, d = err(Y, T)
self.assertEqual(d.shape, Y.shape)
def test_evaluate_returns_scalar(self):
Y = np.ones((4, 3, 2))
T = create_targets_object(np.ones((4, 3, 2)) * 2)
for err in self.error_funcs:
e, d = err(Y, T)
self.assertIsInstance(e, float)
