def mutate_arch(parent_arch):
# Choose one of the three mutations
mutation = np.random.choice(
['identity', 'hidden_state_mutation', 'op_mutation'])
adjacency_matrix, node_list = copy.deepcopy(
parent_arch.adjacency_matrix), copy.deepcopy(parent_arch.node_list)
if mutation == 'identity':
return Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
elif mutation == 'hidden_state_mutation':
# Pick one of the intermediate nodes in the graph (neither input nor output)
if type(search_space) == SearchSpace1:
# Node 1 has now choice for node 2 as it always has 2 parents
low = 3
else:
low = 2
random_node = np.random.randint(low=low,
high=adjacency_matrix.shape[-1])
# Pick one of the parents of the node
# parent_of_node_to_modify = np.random.choice(adjacency_matrix[:random_node, random_node].nonzero()[0])
parents = adjacency_matrix[:random_node, random_node].nonzero()[0]
#print('matrix', adjacency_matrix)
#print('parents', parents)
if parents.any():
parent_of_node_to_modify = np.random.choice(parents)
# Select a new parent for this node, (needs to be different from previous parent)
new_parent_of_node = np.random.choice(
np.argwhere(adjacency_matrix[:random_node,
random_node] == 0).flatten())
# Remove old parent from child
adjacency_matrix[parent_of_node_to_modify, random_node] = 0
# Add new parent to child
adjacency_matrix[new_parent_of_node, random_node] = 1
# Create new child config
return Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
else:
# don't make any changes
return Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
else: # op_mutation
OPS = [CONV3X3, CONV1X1, MAXPOOL3X3]
op_idx_to_change = np.random.randint(len(node_list))
# Remove current op on selected idx (because we want a new op)
OPS.remove(node_list[op_idx_to_change])
# Select one of the remaining ops
new_op = np.random.choice(OPS)
node_list[op_idx_to_change] = new_op
return Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
def get_neighborhood(arch):
neighbors = []
for i in range(len(arch.node_list)):
adjacency_matrix, node_list = copy.deepcopy(
arch.adjacency_matrix), copy.deepcopy(arch.node_list)
OPS = [CONV3X3, CONV1X1, MAXPOOL3X3]
OPS.remove(node_list[i])
new_op = np.random.choice(OPS)
node_list[i] = new_op
neighbors.append(
Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list))
if type(search_space) == SearchSpace1:
# Node 1 has now choice for node 2 as it always has 2 parents
low = 3
else:
low = 2
for i in range(low, arch.adjacency_matrix.shape[-1]):
parents = arch.adjacency_matrix[:i, i].nonzero()[0]
for parent in parents:
for new_parent in np.argwhere(
arch.adjacency_matrix[:i, i] == 0).flatten():
# Select a new parent for this node, (needs to be different from previous parent)
adjacency_matrix = copy.deepcopy(arch.adjacency_matrix)
node_list = copy.deepcopy(arch.node_list)
adjacency_matrix[parent, i] = 0
# Add new parent to child
adjacency_matrix[new_parent, i] = 1
# Create new child config
neighbors.append(
Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list))
return neighbors
def objective_function(self, nasbench, config, budget=108):
adjacency_matrix, node_list = super(
SearchSpace2, self).convert_config_to_nasbench_format(config)
# adjacency_matrix = upscale_to_nasbench_format(adjacency_matrix)
node_list = [INPUT, *node_list, CONV1X1, OUTPUT]
adjacency_list = adjacency_matrix.astype(np.int).tolist()
model_spec = api.ModelSpec(matrix=adjacency_list, ops=node_list)
nasbench_data = nasbench.query(model_spec, epochs=budget)
# record the data to history
architecture = Model()
arch = Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
architecture.update_data(arch, nasbench_data, budget)
self.run_history.append(architecture)
return nasbench_data['validation_accuracy'], nasbench_data[
'training_time']
def random_architecture():
adjacency_matrix, node_list = search_space.sample_with_loose_ends()
architecture = Architecture(adjacency_matrix=adjacency_matrix,
node_list=node_list)
return architecture