class HANSageHander(BaseHander):
"""
wrapper for SupervisedGraphSage model
"""
def __init__(self, num_class, data, args):
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.adj_matrix = data['adj_matrix']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.split_seed = args.split_seed
self.is_cuda = args.cuda
self.view = args.view
self.num_sample_tpl = args.num_sample_tpl
self.num_sample_permission = args.num_sample_permission
self.num_neighs_tpl = args.num_neighs_tpl
self.num_neighs_permission = args.num_neighs_permission
self.embed_dim = args.embed_dim
self.freeze = args.freeze
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists, args.split_seed, args.label_rate, self.is_cuda)
self.inst_generator = self.inputdata.gen_train_batch(batch_size=args.batch_size)
self.train_data_loader = self.inputdata.get_train_data_load(batch_size=args.batch_size, shuffle=True)
def build_model(self):
logger.info("define model.")
num_layers = 2
self.model = HANSage(self.num_class, self.num_nodes, self.feat_data, self.feat_dim, self.adj_lists, self.adj_matrix, self.is_cuda, self.num_sample_tpl, self.num_sample_permission, self.embed_dim, num_layers)
logger.info(self.model)
if self.is_cuda:
self.model.cuda()
self.custom_init(self.freeze)
self.optimizer = torch.optim.Adam(filter(lambda param: param.requires_grad, self.model.parameters()), lr=1e-3, weight_decay=1e-5)
unbalance_alpha = torch.Tensor([0.9934, 1])
if self.is_cuda:
unbalance_alpha = unbalance_alpha.cuda()
self.loss_func = nn.CrossEntropyLoss(weight=unbalance_alpha)
def custom_init(self, freeze=False):
logger.info("custom initialization. freeze={}".format(freeze))
from setting import model_path
import glob
checkpoint_tpl = torch.load(glob.glob(os.path.join(model_path, 'GraphSage', "*tpl*neigh{}".format(self.num_neighs_tpl)))[0])
tpl_state = checkpoint_tpl['state_dict']
self.model.encoder_tpl.load_state_dict(tpl_state, strict=False)
if freeze:
for param in self.model.encoder_tpl.parameters():
param.requires_grad = False
checkpoint_permission = torch.load(glob.glob(os.path.join(model_path, 'GraphSage', "*permission*neigh{}".format(self.num_neighs_permission)))[0])
permission_state = checkpoint_permission['state_dict']
self.model.encoder_permission.load_state_dict(permission_state, strict=False)
if freeze:
for param in self.model.encoder_permission.parameters():
param.requires_grad = False
def __init__(self, num_class, data, args):
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.is_cuda = args.cuda
self.view = args.view
self.num_sample = args.num_sample
self.embed_dim = args.embed_dim
self.freeze = args.freeze
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists,
args.label_rate, self.is_cuda)
self.train_data_loader = self.inputdata.get_train_data_load(
batch_size=args.batch_size, shuffle=True)
def preprocess_image(image_path):
class_id, sample_id = InputData.split_path(str(image_path))
high_resolution_image = cv2.imread(str(image_path), cv2.IMREAD_COLOR)
low_resolution, high_resolution_image = _reduce_resolution(
high_resolution_image)
return image_example(low_resolution, high_resolution_image, class_id,
sample_id)
class GraphSageHandler(BaseHander):
"""
wrapper for GraphSage model
"""
def __init__(self, num_class, data, args):
super(GraphSageHandler, self).__init__(args)
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.is_cuda = args.cuda
self.view = args.view
self.num_sample = args.num_sample
self.embed_dim = args.embed_dim
self.dropout = args.dropout
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists,
args.label_rate, self.is_cuda)
self.train_data_loader = self.inputdata.get_train_data_load(
batch_size=args.batch_size, shuffle=True)
def build_model(self):
logger.info("define model.")
self.num_layers = 2
self.model = GraphSage(self.adj_lists, self.feat_data, self.num_class,
self.embed_dim, self.num_sample,
self.num_layers, self.is_cuda)
logger.info('\n{}'.format(self.model))
if self.is_cuda:
self.model.cuda()
self.optimizer = torch.optim.Adam(self.model.parameters(),
lr=1e-3,
weight_decay=1e-5)
self.loss_func = nn.CrossEntropyLoss()
def __init__(self, num_class, data, args):
super(GraphSageHandler, self).__init__(args)
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.is_cuda = args.cuda
self.view = args.view
self.num_sample = args.num_sample
self.embed_dim = args.embed_dim
self.dropout = args.dropout
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists,
args.label_rate, self.is_cuda)
self.train_data_loader = self.inputdata.get_train_data_load(
batch_size=args.batch_size, shuffle=True)
self.ddc_target_data_loader = self.inputdata.get_ddc_target_data_load(
batch_size=args.batch_size, shuffle=True)
class GraphSageHandler(BaseHander):
"""
wrapper for GraphSage model
"""
def __init__(self, num_class, data, args):
super(GraphSageHandler, self).__init__(args)
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.adj_matrix = data['adj_matrix']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.split_seed = args.split_seed
self.is_cuda = args.cuda
self.view = args.view
self.num_neighs = args.num_neighs
self.num_sample = args.num_sample
self.embed_dim = args.embed_dim
self.dropout = args.dropout
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists,
args.split_seed, args.label_rate,
self.is_cuda)
self.inst_generator = self.inputdata.gen_train_batch(
batch_size=args.batch_size)
self.train_data_loader = self.inputdata.get_train_data_load(
batch_size=args.batch_size, shuffle=True)
def build_model(self):
logger.info("define model.")
num_layers = 2
self.model = GraphSage(self.num_class, self.num_nodes, self.feat_data,
self.feat_dim, self.adj_lists, self.adj_matrix,
self.is_cuda, self.num_sample, self.embed_dim,
num_layers)
logger.info(self.model)
if self.is_cuda:
self.model.cuda()
self.optimizer = torch.optim.Adam(self.model.parameters(),
lr=1e-3,
weight_decay=1e-5)
unbalance_alpha = torch.Tensor([0.9934, 1])
if self.is_cuda:
unbalance_alpha = unbalance_alpha.cuda()
self.loss_func = nn.CrossEntropyLoss(weight=unbalance_alpha)
def __init__(self, num_class, data, args):
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.adj_matrix = data['adj_matrix']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.split_seed = args.split_seed
self.is_cuda = args.cuda
self.view = args.view
self.num_sample_tpl = args.num_sample_tpl
self.num_sample_permission = args.num_sample_permission
self.num_neighs_tpl = args.num_neighs_tpl
self.num_neighs_permission = args.num_neighs_permission
self.embed_dim = args.embed_dim
self.freeze = args.freeze
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists, args.split_seed, args.label_rate, self.is_cuda)
self.inst_generator = self.inputdata.gen_train_batch(batch_size=args.batch_size)
self.train_data_loader = self.inputdata.get_train_data_load(batch_size=args.batch_size, shuffle=True)
class HANSageHander(BaseHander):
def __init__(self, num_class, data, args):
self.num_class = num_class
self.labels = data['labels']
self.adj_lists = data['adj_lists']
self.feat_data = data['feat_data']
self.num_nodes, self.feat_dim = self.feat_data.shape
self.is_cuda = args.cuda
self.view = args.view
self.num_sample = args.num_sample
self.embed_dim = args.embed_dim
self.freeze = args.freeze
self.inputdata = InputData(self.num_nodes, self.labels, self.adj_lists,
args.label_rate, self.is_cuda)
self.train_data_loader = self.inputdata.get_train_data_load(
batch_size=args.batch_size, shuffle=True)
def build_model(self):
logger.info("define model.")
self.num_layers = 2
self.model = HANSage(self.adj_lists, self.feat_data, self.num_class,
self.embed_dim, self.num_sample, self.num_layers,
self.is_cuda)
logger.info('\n{}'.format(self.model))
if self.is_cuda:
self.model.cuda()
self.custom_init(self.freeze)
self.optimizer = torch.optim.Adam(filter(
lambda param: param.requires_grad, self.model.parameters()),
lr=1e-3,
weight_decay=1e-5)
self.loss_func = nn.CrossEntropyLoss()
def custom_init(self, freeze=False):
logger.info("custom initialization. freeze={}".format(freeze))
from setting import model_path
import glob
all_views = [
'app_permission_app', 'app_url_app', 'app_component_app',
'app_tpl_app'
]
for i in range(len(all_views)):
checkpoint = torch.load(
glob.glob(os.path.join(model_path,
"*{}*".format(all_views[i])))[0])
state_dict = checkpoint['state_dict']
self.model.encoders[i].load_state_dict(state_dict, strict=False)
if freeze:
for param in self.model.encoders[i].parameters():
param.requires_grad = False