def build_model(self):
self.args.share_param = False
self.with_retrain = True
self.args.shared_initial_step = 0
if self.args.search_mode == "macro":
# generate model description in macro way (generate entire network description)
from graphnas.search_space import MacroSearchSpace
search_space_cls = MacroSearchSpace()
self.search_space = search_space_cls.get_search_space()
# layers_of_child_model is 2
self.action_list = search_space_cls.generate_action_list(
self.args.layers_of_child_model)
# build RNN controller
from graphnas.graphnas_controller import SimpleNASController
self.controller = SimpleNASController(
self.args,
action_list=self.action_list,
search_space=self.search_space,
cuda=self.args.cuda)
if self.args.dataset in ["cora", "citeseer", "pubmed"]:
# implements based on dgl
self.submodel_manager = CitationGNNManager(self.args)
if self.args.dataset in ["Cora", "Citeseer", "Pubmed"]:
# implements based on pyg
self.submodel_manager = GeoCitationManager(self.args)
if self.args.search_mode == "micro":
self.args.format = "micro"
self.args.predict_hyper = True
if not hasattr(self.args, "num_of_cell"):
self.args.num_of_cell = 2
from graphnas_variants.micro_graphnas.micro_search_space import IncrementSearchSpace
search_space_cls = IncrementSearchSpace()
search_space = search_space_cls.get_search_space()
from graphnas.graphnas_controller import SimpleNASController
from graphnas_variants.micro_graphnas.micro_model_manager import MicroCitationManager
self.submodel_manager = MicroCitationManager(self.args)
self.search_space = search_space
action_list = search_space_cls.generate_action_list(
cell=self.args.num_of_cell)
if hasattr(self.args, "predict_hyper") and self.args.predict_hyper:
self.action_list = action_list + [
"learning_rate", "dropout", "weight_decay", "hidden_unit"
]
else:
self.action_list = action_list
self.controller = SimpleNASController(
self.args,
action_list=self.action_list,
search_space=self.search_space,
cuda=self.args.cuda)
if self.cuda:
self.controller.cuda()
if self.cuda:
self.controller.cuda()
def eval(chromosome):
args = build_args()
dataset_list = ["Citeseer", "Pubmed", "cora"]
base_list = [
"pyg",
"pyg",
"dgl",
]
for dataset, actions, base in zip(dataset_list, chromosome, base_list):
# if dataset == "cora":
# continue
args.dataset = dataset
if base == "dgl":
manager = CitationGNNManager(args)
else:
manager = GeoCitationManager(args)
val_acc, test_acc = manager.evaluate(actions)
print(test_acc)
print("_" * 80)
return test_acc
def build_model(self):
if self.args.search_mode == "macro":
search_space_cls = MacroSearchSpace()
self.search_space = search_space_cls.get_search_space()
self.action_list = search_space_cls.generate_action_list(
self.args.layers_of_child_model)
# build RNN controller
if self.args.dataset in ["cora", "citeseer", "pubmed"]:
# implements based on dgl
self.submodel_manager = CitationGNNManager(self.args)
if self.args.dataset in ["Cora", "Citeseer", "Pubmed",
"CS", "Physics", "Computers", "Photo"]:
# implements based on pyg
self.submodel_manager = GeoCitationManager(self.args)
print("Search space:")
print(self.search_space)
print("Generated Action List: ")
print(self.action_list)
def build_model(self):
if self.args.search_mode == "macro":
# generate model description in macro way
# (generate entire network description)
search_space_cls = MacroSearchSpace()
self.search_space = search_space_cls.get_search_space()
self.action_list = search_space_cls.generate_action_list(
self.args.layers_of_child_model)
# build RNN controller
if self.args.dataset in ["cora", "citeseer", "pubmed"]:
# implements based on dgl
self.submodel_manager = CitationGNNManager(self.args)
if self.args.dataset in [
"Cora", "Citeseer", "Pubmed", "CS", "Physics", "Computers",
"Photo"
]:
# implements based on pyg
self.submodel_manager = GeoCitationManager(self.args)
if self.args.search_mode == "micro":
self.args.format = "micro"
self.args.predict_hyper = True
if not hasattr(self.args, "num_of_cell"):
self.args.num_of_cell = 2
search_space_cls = IncrementSearchSpace()
search_space = search_space_cls.get_search_space()
self.submodel_manager = MicroCitationManager(self.args)
self.search_space = search_space
action_list = search_space_cls.generate_action_list(
cell=self.args.num_of_cell)
if hasattr(self.args, "predict_hyper") and self.args.predict_hyper:
self.action_list = action_list + [
"learning_rate", "dropout", "weight_decay", "hidden_unit"
]
else:
self.action_list = action_list
print("Search space:")
print(self.search_space)
print("Generated Action List: ")
print(self.action_list)
if __name__ == "__main__":
args = build_args()
gnn_list = [
['gat', 'sum', 'linear', 4, 128, 'linear', 'sum', 'elu', 8, 6],
['gcn', 'sum', 'tanh', 6, 64, 'cos', 'sum', 'tanh', 6, 3],
['const', 'sum', 'relu6', 2, 128, 'gat', 'sum', 'linear', 2, 7],
]
dataset_list = ["Citeseer", "Pubmed", "cora"]
base_list = [
"pyg",
"pyg",
"dgl",
]
for dataset, actions, base in zip(dataset_list, gnn_list, base_list):
# if dataset == "cora":
# continue
args.dataset = dataset
if base == "dgl":
manager = CitationGNNManager(args)
else:
manager = GeoCitationManager(args)
test_scores_list = []
for i in range(100):
val_acc, test_acc = manager.evaluate(actions)
test_scores_list.append(test_acc)
print("_" * 80)
test_scores_list.sort()
print(dataset, np.mean(test_scores_list[5:-5]),
np.std(test_scores_list[5:-5]))