def mutate(self, rollout, node_mutate_prob=0.5):
new_arch = copy.deepcopy(rollout.arch)
# randomly select a cell gorup to modify
mutate_i_cg = np.random.randint(0, self.num_cell_groups)
num_prims = self._num_primitives \
if not self.cellwise_primitives else self._num_primitives_list[mutate_i_cg]
_num_step = self.get_num_steps(mutate_i_cg)
if np.random.random() < node_mutate_prob:
# mutate connection
# if #cell init nodes is 1, no need to mutate the connection to node 1
start = int(self.num_init_nodes == 1) * self.num_node_inputs
node_mutate_idx = np.random.randint(start,
len(new_arch[mutate_i_cg][0]))
i_out = node_mutate_idx // self.num_node_inputs
out_node = i_out + self.num_init_nodes
offset = np.random.randint(1, out_node)
new_arch[mutate_i_cg][0][node_mutate_idx] = \
(new_arch[mutate_i_cg][0][node_mutate_idx] + offset) % out_node
else:
# mutate op
op_mutate_idx = np.random.randint(0, len(new_arch[mutate_i_cg][1]))
offset = np.random.randint(1, num_prims)
new_arch[mutate_i_cg][0][op_mutate_idx] = \
(new_arch[mutate_i_cg][0][op_mutate_idx] + offset) % num_prims
return Rollout(new_arch, info={}, search_space=self)
def random_sample(self):
"""
Random sample a discrete architecture.
"""
return Rollout(Rollout.random_sample_arch(
self.num_cell_groups, self.num_steps, self.num_init_nodes,
self.num_node_inputs, self._num_primitives
if not self.cellwise_primitives else self._num_primitives_list),
info={},
search_space=self)
def rollout_from_genotype(self, genotype):
genotype_list = list(genotype._asdict().values())
assert len(genotype_list) == 2 * self.num_cell_groups
arch = []
for i in range(self.num_cell_groups):
cell_geno = genotype_list[2 * i]
nodes, ops = [], []
for i_out in range(self.num_steps):
for i_in in range(self.num_node_inputs):
conn = cell_geno[i_out * self.num_node_inputs + i_in]
ops.append(self.shared_primitives.index(conn[0]))
nodes.append(conn[1])
cg_arch = [nodes, ops]
arch.append(cg_arch)
return Rollout(arch, {}, self)
def rollout_from_genotype(self, genotype):
genotype_list = list(genotype._asdict().values())
assert len(genotype_list) == 2 * self.num_cell_groups
conn_list = genotype_list[:self.num_cell_groups]
arch = []
for i_cg, cell_geno in enumerate(conn_list):
nodes, ops = [], []
num_steps = self.get_num_steps(i_cg)
for i_out in range(num_steps):
for i_in in range(self.num_node_inputs):
conn = cell_geno[i_out * self.num_node_inputs + i_in]
ops.append(self.cell_shared_primitives[i_cg].index(conn[0]))
nodes.append(conn[1])
cg_arch = [nodes, ops]
arch.append(cg_arch)
return Rollout(arch, {}, self)
def mutate(self, rollout, node_mutate_prob=0.5):
new_arch = copy.deepcopy(rollout.arch)
mutate_i_cg = np.random.randint(0, self.num_cell_groups)
num_prims = self._num_primitives if not self.cellwise_primitives else self._num_primitives_list[
mutate_i_cg]
_num_step = self.get_num_steps(mutate_i_cg)
if np.random.random() < node_mutate_prob:
node_mutate_idx = np.random.randint(0,
len(new_arch[mutate_i_cg][0]))
i_out = node_mutate_idx // self.num_init_nodes
new_node = np.random.randint(0, i_out + self.num_init_nodes)
while new_node == new_arch[mutate_i_cg][0][node_mutate_idx]:
new_node = np.random.randint(0, i_out + self.num_init_nodes)
new_arch[mutate_i_cg][0][node_mutate_idx] = new_node
else:
op_mutate_idx = np.random.randint(0, len(new_arch[mutate_i_cg][1]))
new_prim = np.random.randint(0, num_prims)
while new_prim == new_arch[mutate_i_cg][0][op_mutate_idx]:
new_prim = np.random.randint(0, num_prims)
new_arch[mutate_i_cg][0][op_mutate_idx] = new_prim
return Rollout(new_arch, info={}, search_space=self)