def test_teleport(network: nn.Module, input_shape: Tuple = (1, 1, 28, 28), verbose: bool = False,
atol: float = 1e-5, model_name: str = None):
"""
Return mean of the difference between the weights of network and a random teleportation, and checks if
teleportation has the same network function
Args:
network (nn.Module): Network to be tested
input_shape (tuple): Input shape of network
verbose (bool): Flag to print comparision between network and a teleportation
atol (float): Absolute tolerance allowed between outputs to pass the test
model_name (str): The name or label assigned to differentiate the model
Returns:
float with the average of the difference between the weights of the network and a teleportation
"""
model_name = model_name or network.__class__.__name__
model = NeuralTeleportationModel(network=network, input_shape=input_shape)
model.eval() # model must be set to eval because of dropout
x = torch.rand(input_shape)
pred1 = model(x).detach().numpy()
w1 = model.get_weights().detach().numpy()
model.random_teleport()
pred2 = model(x).detach().numpy()
w2 = model.get_weights().detach().numpy()
diff_average = np.mean(np.abs((pred1 - pred2)))
if verbose:
print("Sample outputs: ")
print("Pre teleportation: ", pred1.flatten()[:10])
print("Post teleportation: ", pred2.flatten()[:10])
print("Diff weight average: ", np.mean(np.abs((w1 - w2))))
print("Diff prediction average: ", diff_average)
assert not np.allclose(w1, w2)
assert np.allclose(pred1, pred2, atol=atol), "Teleporation did not work for model {}. Average difference: {}". \
format(model_name, diff_average)
print("Teleportation successful for " + model_name + " model.")
return diff_average
def test_multiple_teleport(network: nn.Module, input_shape: Tuple = (1, 1, 28, 28), verbose: bool = False,
atol: float = 1e-5, model_name: str = None):
"""
Test multiple successive teleporations.
Args:
network (nn.Module): Network to be tested
input_shape (tuple): Input shape of network
verbose (bool): Flag to print comparision between network and a teleportation
atol (float): Absolute tolerance allowed between outputs to pass the test
model_name (str): The name or label assigned to differentiate the model
"""
model_name = model_name or network.__class__.__name__
model = NeuralTeleportationModel(network=network, input_shape=input_shape)
model.eval() # model must be set to eval because of dropout
x = torch.rand(input_shape)
pred1 = model(x).detach().numpy()
for _ in range(10):
model.random_teleport(cob_range=10, sampling_type='inter_landscape')
pred2 = model(x).detach().numpy()
diff_average = np.mean(np.abs((pred1 - pred2)))
assert np.allclose(pred1, pred2,
atol=atol), "Multiple Teleporation did not work for model {}. Average difference: {}".format(
model_name, diff_average)
for _ in range(10):
model.random_teleport(cob_range=10, sampling_type='inter_landscape', reset_teleportation=False)
pred2 = model(x).detach().numpy()
diff_average = np.mean(np.abs((pred1 - pred2)))
assert np.allclose(pred1, pred2,
atol=atol), "Multiple Teleporation did not work for model {}. Average difference: {}".format(
model_name, diff_average)
print("Multiple Teleportations successful for " + model_name + " model.")