def test_weight_minimization(self):
base_network = reproducible_network_train()
decay_network = reproducible_network_train(
decay_rate=0.1, addons=[algorithms.WeightDecay])
iter_networks = zip(base_network.layers, decay_network.layers)
for net_layer, decay_layer in iter_networks:
self.assertGreater(
np.linalg.norm(net_layer.weight.get_value()),
np.linalg.norm(decay_layer.weight.get_value()),
)
def test_weight_minimization(self):
base_network = reproducible_network_train()
decay_network = reproducible_network_train(
decay_rate=0.1,
addons=[algorithms.WeightDecay]
)
iter_networks = zip(base_network.layers[1:-1],
decay_network.layers[1:-1])
for net_layer, decay_layer in iter_networks:
self.assertGreater(
np.linalg.norm(net_layer.weight.get_value()),
np.linalg.norm(decay_layer.weight.get_value()),
)
def test_with_step_minimization_alg(self):
default_step = 0.3
net1 = reproducible_network_train(step=default_step)
net2 = reproducible_network_train(
step=default_step,
decay_rate=0.25,
zero_weight=10,
addons=[algorithms.WeightElimination]
)
net3 = reproducible_network_train(
step=default_step,
decay_rate=0.25,
zero_weight=10,
addons=[algorithms.WeightElimination,
algorithms.StepDecay]
)
# Check that step is valid for each network
StepCase = namedtuple('StepCase', 'network expected_step')
step_test_cases = (
StepCase(network=net1, expected_step=default_step),
StepCase(network=net2, expected_step=default_step),
StepCase(network=net3, expected_step=default_step / 6.),
)
for case in step_test_cases:
step = case.network.variables.step
self.assertAlmostEqual(step.get_value(), case.expected_step,
places=5)
# Compare weight norm between networks
WeightNormCase = namedtuple('WeightNormCase',
'with_smaller_norm with_bigger_norm')
norm_test_cases = (
WeightNormCase(with_smaller_norm=net2, with_bigger_norm=net1),
WeightNormCase(with_smaller_norm=net3, with_bigger_norm=net1),
)
for case in norm_test_cases:
network_layers = zip(case.with_smaller_norm.layers[1:-1],
case.with_bigger_norm.layers[1:-1])
for smaller_norm, bigger_norm in network_layers:
weight_smaller_norm = smaller_norm.weight.get_value()
weight_bigger_norm = bigger_norm.weight.get_value()
self.assertGreater(
np.linalg.norm(weight_bigger_norm),
np.linalg.norm(weight_smaller_norm)
)
def test_with_step_minimization_alg(self):
default_step = 0.3
net1 = reproducible_network_train(step=default_step)
net2 = reproducible_network_train(
step=default_step,
decay_rate=0.25,
zero_weight=10,
addons=[algorithms.WeightElimination]
)
net3 = reproducible_network_train(
step=default_step,
decay_rate=0.25,
zero_weight=10,
addons=[algorithms.WeightElimination,
algorithms.StepDecay]
)
# Check that step is valid for each network
StepCase = namedtuple('StepCase', 'network expected_step')
step_test_cases = (
StepCase(network=net1, expected_step=default_step),
StepCase(network=net2, expected_step=default_step),
StepCase(network=net3, expected_step=default_step / 6.),
)
for case in step_test_cases:
step = case.network.variables.step
self.assertAlmostEqual(step.get_value(), case.expected_step,
places=2)
# Compare weight norm between networks
WeightNormCase = namedtuple('WeightNormCase',
'with_smaller_norm with_bigger_norm')
norm_test_cases = (
WeightNormCase(with_smaller_norm=net2, with_bigger_norm=net1),
WeightNormCase(with_smaller_norm=net3, with_bigger_norm=net1),
)
for case in norm_test_cases:
network_layers = zip(case.with_smaller_norm.layers[1:-1],
case.with_bigger_norm.layers[1:-1])
for smaller_norm, bigger_norm in network_layers:
weight_smaller_norm = smaller_norm.weight.get_value()
weight_bigger_norm = bigger_norm.weight.get_value()
self.assertGreater(
np.linalg.norm(weight_bigger_norm),
np.linalg.norm(weight_smaller_norm)
)
def test_log_scale(self):
original_image_name = format_image_name("log_scale.png")
original_image = os.path.join(IMGDIR, original_image_name)
with image_comparison(original_image) as fig:
ax = fig.add_subplot(1, 1, 1)
network = reproducible_network_train(step=0.3)
network.plot_errors(logx=True, ax=ax, show=False)
def test_log_scale(self):
original_image_name = format_image_name("log_scale.png")
original_image = os.path.join(IMGDIR, original_image_name)
with image_comparison(original_image) as fig:
ax = fig.add_subplot(1, 1, 1)
network = reproducible_network_train(step=0.3)
plots.error_plot(network, logx=True, ax=ax, show=False)
def test_simple_plot(self):
original_image_name = format_image_name("simple_plot.png")
original_image = os.path.join(IMGDIR, original_image_name)
with image_comparison(original_image) as fig:
ax = fig.add_subplot(1, 1, 1)
network = reproducible_network_train(step=0.3)
plots.error_plot(network, ax=ax, show=False)
def test_simple_plot(self):
original_image_name = format_image_name("simple_plot.png")
original_image = os.path.join(IMGDIR, original_image_name)
with image_comparison(original_image) as fig:
ax = fig.add_subplot(1, 1, 1)
network = reproducible_network_train(step=0.3)
network.plot_errors(ax=ax, show=False)
def test_weight_minimization(self):
base_network = reproducible_network_train()
decay_network = reproducible_network_train(
decay_rate=0.3,
zero_weight=5,
addons=[algorithms.WeightElimination])
iter_networks = zip(
base_network.layers[1:-1],
decay_network.layers[1:-1],
)
for net_layer, decay_layer in iter_networks:
self.assertGreater(
np.linalg.norm(self.eval(net_layer.weight)),
np.linalg.norm(self.eval(decay_layer.weight)),
)
def test_with_step_minimization_alg(self):
default_step = 0.3
net1 = reproducible_network_train(step=default_step)
net2 = reproducible_network_train(step=default_step,
addons=[algorithms.WeightDecay])
net3 = reproducible_network_train(step=default_step,
addons=[
algorithms.WeightDecay,
algorithms.StepDecay,
])
# Check that step is valid for each network
StepCase = namedtuple('StepCase', 'network expected_step')
step_test_cases = (
StepCase(network=net1, expected_step=default_step),
StepCase(network=net2, expected_step=default_step),
StepCase(network=net3, expected_step=default_step / 6.),
)
for case_index, case in enumerate(step_test_cases, start=1):
step = case.network.variables.step
self.assertAlmostEqual(self.eval(step),
case.expected_step,
places=5,
msg="Test case #{}".format(case_index))
# Compare weight norm between networks
WeightNormCase = namedtuple('WeightNormCase',
'with_smaller_norm with_bigger_norm')
norm_test_cases = (
WeightNormCase(with_smaller_norm=net2, with_bigger_norm=net1),
WeightNormCase(with_smaller_norm=net3, with_bigger_norm=net1),
)
for case in norm_test_cases:
network_layers = zip(
case.with_smaller_norm.layers[1:-1],
case.with_bigger_norm.layers[1:-1],
)
for smaller_norm, bigger_norm in network_layers:
weight_smaller_norm = self.eval(smaller_norm.weight)
weight_bigger_norm = self.eval(bigger_norm.weight)
self.assertGreater(np.linalg.norm(weight_bigger_norm),
np.linalg.norm(weight_smaller_norm))
def test_error_plot_show_image(self):
def mock_plt_show():
pass
# Test suppose not to fail
real_plt_show = plt.show
plt.show = mock_plt_show
network = reproducible_network_train(step=0.3)
plots.error_plot(network, show=True)
plt.show = real_plt_show
def test_error_plot_show_image(self):
def mock_plt_show():
pass
# Test suppose not to fail
real_plt_show = plt.show
plt.show = mock_plt_show
network = reproducible_network_train(step=0.3)
plots.error_plot(network, show=True)
plt.show = real_plt_show
