class BayesianSearcher(Searcher):
def __init__(self,
n_classes,
input_shape,
path,
verbose,
trainer_args=None,
default_model_len=constant.MODEL_LEN,
default_model_width=constant.MODEL_WIDTH,
beta=constant.BETA,
kernel_lambda=constant.KERNEL_LAMBDA,
t_min=constant.T_MIN):
super().__init__(n_classes, input_shape, path, verbose, trainer_args,
default_model_len, default_model_width)
self.gpr = IncrementalGaussianProcess(kernel_lambda)
self.search_tree = SearchTree()
self.init_search_queue = None
self.init_gpr_x = []
self.init_gpr_y = []
self.beta = beta
self.t_min = t_min
def search(self, x_train, y_train, x_test, y_test):
if not self.history:
model = DefaultClassifierGenerator(self.n_classes,
self.input_shape).generate(
self.default_model_len,
self.default_model_width)
history_item = self.add_model(model, x_train, y_train, x_test,
y_test)
self.search_tree.add_child(-1, history_item['model_id'])
graph = Graph(model)
self.init_search_queue = []
# for child_graph in transform(graph):
# self.init_search_queue.append((child_graph, history_item['model_id']))
self.init_gpr_x.append(graph.extract_descriptor())
self.init_gpr_y.append(history_item['accuracy'])
pickle_to_file(self, os.path.join(self.path, 'searcher'))
return
if self.init_search_queue:
graph, father_id = self.init_search_queue.pop()
model = graph.produce_model()
history_item = self.add_model(model, x_train, y_train, x_test,
y_test)
self.search_tree.add_child(father_id, history_item['model_id'])
self.init_gpr_x.append(graph.extract_descriptor())
self.init_gpr_y.append(history_item['accuracy'])
pickle_to_file(self, os.path.join(self.path, 'searcher'))
return
if not self.init_search_queue and not self.gpr.first_fitted:
self.gpr.first_fit(self.init_gpr_x, self.init_gpr_y)
new_model, father_id = self.maximize_acq()
history_item = self.add_model(new_model, x_train, y_train, x_test,
y_test)
self.search_tree.add_child(father_id, history_item['model_id'])
self.gpr.incremental_fit(
Graph(new_model).extract_descriptor(), history_item['accuracy'])
pickle_to_file(self, os.path.join(self.path, 'searcher'))
def maximize_acq(self):
model_ids = self.search_tree.adj_list.keys()
target_graph = None
father_id = None
descriptors = self.descriptors
pq = PriorityQueue()
temp_list = []
for model_id in model_ids:
accuracy = self.get_accuracy_by_id(model_id)
temp_list.append((accuracy, model_id))
temp_list = sorted(temp_list)
if len(temp_list) > 5:
temp_list = temp_list[:-5]
for accuracy, model_id in temp_list:
model = self.load_model_by_id(model_id)
graph = Graph(model, False)
pq.put(Elem(accuracy, model_id, graph))
t = 1.0
t_min = self.t_min
alpha = 0.9
max_acq = -1
while not pq.empty() and t > t_min:
elem = pq.get()
ap = math.exp((elem.accuracy - max_acq) / t)
if ap > random.uniform(0, 1):
graphs = transform(elem.graph)
graphs = list(
filter(lambda x: x.extract_descriptor() not in descriptors,
graphs))
if not graphs:
continue
for temp_graph in graphs:
temp_acq_value = self.acq(temp_graph)
pq.put(Elem(temp_acq_value, elem.father_id, temp_graph))
descriptors[temp_graph.extract_descriptor()] = True
if temp_acq_value > max_acq:
max_acq = temp_acq_value
father_id = elem.father_id
target_graph = temp_graph
t *= alpha
model = self.load_model_by_id(father_id)
nm_graph = Graph(model, True)
if self.verbose:
print('Father ID: ', father_id)
print(target_graph.operation_history)
for args in target_graph.operation_history:
getattr(nm_graph, args[0])(*list(args[1:]))
return nm_graph.produce_model(), father_id
def acq(self, graph):
mean, std = self.gpr.predict(np.array([graph.extract_descriptor()]))
return mean + self.beta * std
class BayesianSearcher(Searcher):
def __init__(self, n_classes, input_shape, path, verbose):
super().__init__(n_classes, input_shape, path, verbose)
self.gpr = IncrementalGaussianProcess()
self.search_tree = SearchTree()
def search(self, x_train, y_train, x_test, y_test):
if not self.history:
model = DefaultClassifierGenerator(self.n_classes,
self.input_shape).generate()
history_item = self.add_model(model, x_train, y_train, x_test,
y_test)
self.search_tree.add_child(-1, history_item['model_id'])
self.gpr.first_fit(
Graph(model).extract_descriptor(), history_item['accuracy'])
pickle.dump(self, open(os.path.join(self.path, 'searcher'), 'wb'))
del model
backend.clear_session()
else:
model_ids = self.search_tree.get_leaves()
new_model, father_id = self.maximize_acq(model_ids)
history_item = self.add_model(new_model, x_train, y_train, x_test,
y_test)
self.search_tree.add_child(father_id, history_item['model_id'])
self.gpr.incremental_fit(
Graph(new_model).extract_descriptor(),
history_item['accuracy'])
pickle.dump(self, open(os.path.join(self.path, 'searcher'), 'wb'))
del new_model
backend.clear_session()
def maximize_acq(self, model_ids):
overall_max_acq_value = -1
father_id = None
target_graph = None
# exploration
for model_id in model_ids:
model = self.load_model_by_id(model_id)
graph = Graph(to_stub_model(model))
graph.clear_operation_history()
graphs = transform(graph)
for temp_graph in graphs:
temp_acq_value = self._acq(temp_graph)
if temp_acq_value > overall_max_acq_value:
overall_max_acq_value = temp_acq_value
father_id = model_id
target_graph = temp_graph
# exploitation
for i in range(constant.ACQ_EXPLOITATION_DEPTH):
graphs = transform(target_graph)
for temp_graph in graphs:
temp_acq_value = self._acq(temp_graph)
if temp_acq_value > overall_max_acq_value:
overall_max_acq_value = temp_acq_value
target_graph = temp_graph
model = self.load_model_by_id(father_id)
nm_graph = NetworkMorphismGraph(model)
for args in target_graph.operation_history:
getattr(nm_graph, args[0])(*list(args[1:]))
return nm_graph.produce_model(), father_id
def _acq(self, graph):
return self.gpr.predict(np.array([graph.extract_descriptor()]), )[0]