def train(self):
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
self.model_save_dir = os.path.join(self.config["model_save_dir"],
self.config["save_name"])
for epoch in range(config["max_epoch"]):
print(f"Epoch {epoch} starts !")
print("-" * 80)
if epoch > 0 and self.eval_engine.n_no_update == 0:
# previous epoch have already obtained better result
self.engine.save_checkpoint(model_dir=self.model_save_dir)
if self.eval_engine.n_no_update >= MAX_N_UPDATE:
print(
"Early stop criterion triggered, no performance update for {:} times"
.format(MAX_N_UPDATE))
break
train_loader = self.dataset
self.engine.train_an_epoch(epoch_id=epoch,
train_loader=train_loader)
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0],
self.engine.model, epoch)
self.config["run_time"] = self.monitor.stop()
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
if self.config["model"] == "ncf":
self.train_ncf()
elif self.config["model"] == "gcn":
self.train_gcn()
elif self.config["model"] == "mlp":
self.train_mlp()
elif self.config["model"] == "ncf_gcn":
self.train_gcn()
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
self.train_mlp()
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
self.train_ncf()
else:
raise ValueError(
"Model type error: Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn'."
)
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train(self):
"""Default train implementation
"""
assert hasattr(self,
"engine"), "Please specify the exact model engine !"
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
self.engine.data = self.dataset
print("Start training... ")
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print(f"Epoch {epoch} starts !")
print("-" * 80)
if self.check_early_stop(
engine=self.engine,
model_dir=os.path.join(self.config["model_save_dir"],
"model.cpk"),
epoch=epoch,
):
break
data_loader = self.build_data_loader()
self.engine.train_an_epoch(data_loader, epoch_id=epoch)
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0],
self.engine.model, epoch)
"""Sets the learning rate to the initial LR decayed by 10 every 10 epochs"""
lr = self.config["lr"] * (0.5**(epoch // 10))
for param_group in self.engine.optimizer.param_groups:
param_group["lr"] = lr
self.config["run_time"] = self.monitor.stop()
return self.eval_engine.best_valid_performance
def train(self):
""" Main training navigator
Returns:
"""
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
self.train_rescal()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
self.train_mf()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
return self.eval_engine.best_valid_performance
def train(self):
""" Main training navigator
Returns:
"""
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
if self.config["pre_train"] == 0:
self.train_compgcn()
else:
pass
self.train_lightgcn()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
class LightGCN_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(LightGCN_train, self).__init__(config)
self.load_dataset()
self.build_data_loader()
self.engine = LightGCNEngine(self.config)
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
plain_adj, norm_adj, mean_adj = self.dataset.get_adj_mat()
norm_adj = sparse_mx_to_torch_sparse_tensor(norm_adj)
self.config["norm_adj"] = norm_adj
self.config["num_batch"] = self.dataset.n_train // config["batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
def train(self):
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
self.model_save_dir = os.path.join(
self.config["model_save_dir"], self.config["save_name"]
)
for epoch in range(config["max_epoch"]):
print(f"Epoch {epoch} starts !")
print("-" * 80)
if epoch > 0 and self.eval_engine.n_no_update == 0:
# previous epoch have already obtained better result
self.engine.save_checkpoint(model_dir=self.model_save_dir)
if self.eval_engine.n_no_update >= MAX_N_UPDATE:
print(
"Early stop criterion triggered, no performance update for {:} times".format(
MAX_N_UPDATE
)
)
break
train_loader = self.dataset
self.engine.train_an_epoch(epoch_id=epoch, train_loader=train_loader)
self.eval_engine.train_eval(
self.dataset.valid[0], self.dataset.test[0], self.engine.model, epoch
)
self.config["run_time"] = self.monitor.stop()
def test(self):
self.engine.resume_checkpoint(model_dir=self.model_save_dir)
super(LightGCN_train, self).test()
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
print("train gnn ngcf")
if self.config["pre_train"] == 0:
self.train_gcn()
else:
pass
self.train_ngcf()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
class LightGCN_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(LightGCN_train, self).__init__(config)
self.load_dataset()
self.build_data_loader()
self.gpu_id, self.config["device_str"] = self.get_device()
def build_data_loader(self):
(
user_edge_list,
user_edge_type,
item_edge_list,
item_edge_type,
self.config["n_user_fea"],
self.config["n_item_fea"],
) = self.dataset.make_multi_graph()
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
self.config["user_edge_list"] = torch.LongTensor(user_edge_list)
self.config["user_edge_type"] = torch.LongTensor(user_edge_type)
self.config["item_edge_list"] = torch.LongTensor(item_edge_list)
self.config["item_edge_type"] = torch.LongTensor(item_edge_type)
self.config["num_batch"] = self.dataset.n_train // self.config[
"batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
plain_adj, norm_adj, mean_adj = self.dataset.get_adj_mat()
norm_adj = sparse_mx_to_torch_sparse_tensor(norm_adj)
self.config["norm_adj"] = norm_adj
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
if self.config["validate"]:
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0], engine.model,
epoch)
else:
self.eval_engine.train_eval(None, self.dataset.test[0],
engine.model, epoch)
def train(self):
""" Main training navigator
Returns:
"""
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
if self.config["pre_train"] == 0:
self.train_compgcn()
else:
pass
self.train_lightgcn()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train_lightgcn(self):
self.model_dir = os.path.join(self.config["model_save_dir"],
self.config["save_name"])
train_loader = self.dataset
self.engine = LightGCNEngine(self.config)
self._train(self.engine, train_loader, self.model_dir)
def train_compgcn(self):
""" Train RGCN
Returns:
None
"""
train_loader = self.dataset
# Train RGCN
self.engine = CompGCNEngine(self.config)
self.model_save_dir = os.path.join(
self.config["model_save_dir"],
self.config["compgcn_config"]["save_name"])
self._train(self.engine, train_loader, self.model_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
def train_rescal(self):
""" Train rescal
Returns:
None
"""
import sys
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
sys.path.append("../")
from scipy.io.matlab import loadmat
from scipy.sparse import lil_matrix
from rescal.rescal import als as rescal_als
def edge2Tensor(edge_list, edge_type, n_type):
r_tensor = np.zeros(
(edge_list.max() + 1, edge_list.max() + 1, n_type))
# print(np.count_nonzero(r_tensor))
for idx, e in enumerate(edge_type):
i = edge_list[0][idx]
j = edge_list[1][idx]
r_tensor[i][j][e] = 1
print(f"n values: {np.count_nonzero(r_tensor)}")
X = [
lil_matrix(r_tensor[:, :, k]) for k in range(r_tensor.shape[2])
]
return X
def get_emb(r_tensor, dim=64, lambda_A=10, lambda_R=10):
A, R, fit, itr, exectimes = rescal_als(r_tensor,
dim,
init="nvecs",
lambda_A=lambda_A,
lambda_R=lambda_R)
return A
(
user_edge_list,
user_edge_type,
item_edge_list,
item_edge_type,
n_user_fea,
n_item_fea,
) = self.dataset.make_multi_graph()
user_r = edge2Tensor(user_edge_list, user_edge_type, n_user_fea)
item_r = edge2Tensor(item_edge_list, item_edge_type, n_item_fea)
lambda_A = self.config["lambda_A"]
lambda_R = self.config["lambda_R"]
u_emb = get_emb(user_r,
dim=self.config["emb_dim"],
lambda_A=lambda_A,
lambda_R=lambda_R)
i_emb = get_emb(item_r,
dim=self.config["emb_dim"],
lambda_A=lambda_A,
lambda_R=lambda_R)
u_emb.astype(np.float64)
i_emb.astype(np.float64)
train_loader = self.dataset
self.engine = NGCFEngine(self.config)
self.engine.model.user_embedding.weight.data = torch.tensor(
u_emb.astype(np.float64)).to(self.engine.device)
self.engine.model.item_embedding.weight.data = torch.tensor(
i_emb.astype(np.float64)).to(self.engine.device)
self.model_save_dir = os.path.join(
self.config["model_save_dir"],
self.config["rescal_config"]["save_name"])
self._train(self.engine, train_loader, self.model_save_dir)
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
class NGCF_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(NGCF_train, self).__init__(self.config)
self.load_dataset()
self.build_data_loader()
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
(
user_edge_list,
user_edge_type,
item_edge_list,
item_edge_type,
self.config["n_user_fea"],
self.config["n_item_fea"],
) = self.dataset.make_multi_graph()
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
self.config["user_edge_list"] = torch.LongTensor(user_edge_list)
self.config["user_edge_type"] = torch.LongTensor(user_edge_type)
self.config["item_edge_list"] = torch.LongTensor(item_edge_list)
self.config["item_edge_type"] = torch.LongTensor(item_edge_type)
self.config["num_batch"] = self.dataset.n_train // self.config[
"batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
plain_adj, norm_adj, mean_adj = self.dataset.get_adj_mat()
norm_adj = sparse_mx_to_torch_sparse_tensor(norm_adj)
self.config["norm_adj"] = norm_adj
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
if self.config["validate"]:
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0], engine.model,
epoch)
else:
self.eval_engine.train_eval(None, self.dataset.test[0],
engine.model, epoch)
def train_rescal(self):
""" Train rescal
Returns:
None
"""
import sys
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
sys.path.append("../")
from scipy.io.matlab import loadmat
from scipy.sparse import lil_matrix
from rescal.rescal import als as rescal_als
def edge2Tensor(edge_list, edge_type, n_type):
r_tensor = np.zeros(
(edge_list.max() + 1, edge_list.max() + 1, n_type))
# print(np.count_nonzero(r_tensor))
for idx, e in enumerate(edge_type):
i = edge_list[0][idx]
j = edge_list[1][idx]
r_tensor[i][j][e] = 1
print(f"n values: {np.count_nonzero(r_tensor)}")
X = [
lil_matrix(r_tensor[:, :, k]) for k in range(r_tensor.shape[2])
]
return X
def get_emb(r_tensor, dim=64, lambda_A=10, lambda_R=10):
A, R, fit, itr, exectimes = rescal_als(r_tensor,
dim,
init="nvecs",
lambda_A=lambda_A,
lambda_R=lambda_R)
return A
(
user_edge_list,
user_edge_type,
item_edge_list,
item_edge_type,
n_user_fea,
n_item_fea,
) = self.dataset.make_multi_graph()
user_r = edge2Tensor(user_edge_list, user_edge_type, n_user_fea)
item_r = edge2Tensor(item_edge_list, item_edge_type, n_item_fea)
lambda_A = self.config["lambda_A"]
lambda_R = self.config["lambda_R"]
u_emb = get_emb(user_r,
dim=self.config["emb_dim"],
lambda_A=lambda_A,
lambda_R=lambda_R)
i_emb = get_emb(item_r,
dim=self.config["emb_dim"],
lambda_A=lambda_A,
lambda_R=lambda_R)
u_emb.astype(np.float64)
i_emb.astype(np.float64)
train_loader = self.dataset
self.engine = NGCFEngine(self.config)
self.engine.model.user_embedding.weight.data = torch.tensor(
u_emb.astype(np.float64)).to(self.engine.device)
self.engine.model.item_embedding.weight.data = torch.tensor(
i_emb.astype(np.float64)).to(self.engine.device)
self.model_save_dir = os.path.join(
self.config["model_save_dir"],
self.config["rescal_config"]["save_name"])
self._train(self.engine, train_loader, self.model_save_dir)
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train_compgcn(self):
""" Train RGCN
Returns:
None
"""
train_loader = self.dataset
# Train RGCN
self.engine = CompGCNEngine(self.config)
self.model_save_dir = os.path.join(
self.config["model_save_dir"],
self.config["compgcn_config"]["save_name"])
self._train(self.engine, train_loader, self.model_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
class LightGCN_train(TrainEngine):
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(LightGCN_train, self).__init__(self.config)
self.load_dataset()
self.build_data_loader()
self.gpu_id, self.config["device_str"] = self.get_device()
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
user_fea_norm_adj, item_fea_norm_adj = self.dataset.make_fea_sim_mat()
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
self.config["user_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
user_fea_norm_adj
)
self.config["item_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
item_fea_norm_adj
)
plain_adj, norm_adj, mean_adj = self.dataset.get_adj_mat()
norm_adj = sparse_mx_to_torch_sparse_tensor(norm_adj)
self.config["norm_adj"] = norm_adj
self.config["num_batch"] = self.dataset.n_train // self.config["batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
self.eval_engine.train_eval(
self.dataset.valid[0], self.dataset.test[0], engine.model, epoch
)
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
print("train gnn lightgcn")
if self.config["pre_train"] == 0:
self.train_gcn()
else:
pass
self.train_lightgcn()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train_lightgcn(self):
self.model_dir = os.path.join(
self.config["model_save_dir"], self.config["save_name"]
)
train_loader = self.dataset
self.engine = LightGCNEngine(self.config)
self._train(self.engine, train_loader, self.model_dir)
def train_gcn(self):
""" Train GCN
Returns:
None
"""
train_loader = self.dataset
# Train GCN
self.engine = GCN_SEngine(self.config)
self.gcn_save_dir = os.path.join(
self.config["model_save_dir"], self.config["gcn_config"]["save_name"]
)
print(self.gcn_save_dir)
self._train(self.engine, train_loader, self.gcn_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
def test(self):
self.engine.resume_checkpoint(model_dir=self.model_dir)
super(LightGCN_train, self).test()
class NCF_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(NCF_train, self).__init__(self.config)
self.load_dataset()
self.build_data_loader()
self.gpu_id, self.config["device_str"] = self.get_device()
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
user_fea_norm_adj, item_fea_norm_adj = self.dataset.make_fea_sim_mat()
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
self.config["user_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
user_fea_norm_adj
)
self.config["item_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
item_fea_norm_adj
)
self.config["num_batch"] = self.dataset.n_train // self.config["batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
if self.config["validate"]:
self.eval_engine.train_eval(
self.dataset.valid[0], self.dataset.test[0], engine.model, epoch
)
else:
self.eval_engine.train_eval(
None, self.dataset.test[0], engine.model, epoch
)
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
if self.config["model"] == "ncf":
self.train_ncf()
elif self.config["model"] == "gcn":
self.train_gcn()
elif self.config["model"] == "mlp":
self.train_mlp()
elif self.config["model"] == "ncf_gcn":
self.train_gcn()
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
self.train_mlp()
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
self.train_ncf()
else:
raise ValueError(
"Model type error: Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn'."
)
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train_ncf(self):
""" Train NeuMF
Returns:
None
"""
train_loader = self.sample_generator.instance_a_train_loader(
self.config["num_negative"], self.config["batch_size"]
)
self.engine = NeuMFEngine(self.config)
self.neumf_save_dir = os.path.join(
self.config["model_save_dir"], self.config["neumf_config"]["save_name"]
)
self._train(self.engine, train_loader, self.neumf_save_dir)
def train_gcn(self):
""" Train GCN
Returns:
None
"""
train_loader = self.dataset
# Train GCN
self.engine = GCN_SEngine(self.config)
self.gcn_save_dir = os.path.join(
self.config["model_save_dir"], self.config["gcn_config"]["save_name"]
)
self._train(self.engine, train_loader, self.gcn_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
def train_mlp(self):
""" Train MLP
Returns:
None
"""
# Train MLP
train_loader = self.sample_generator.instance_a_train_loader(
self.config["num_negative"], self.config["batch_size"]
)
self.engine = MLPEngine(self.config)
self.mlp_save_dir = os.path.join(
self.config["model_save_dir"], self.config["mlp_config"]["save_name"]
)
self._train(self.engine, train_loader, self.mlp_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
class MF_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(MF_train, self).__init__(self.config)
self.load_dataset()
self.build_data_loader()
self.gpu_id, self.config["device_str"] = self.get_device()
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
user_fea_norm_adj, item_fea_norm_adj = self.dataset.make_fea_sim_mat()
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
self.config["user_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
user_fea_norm_adj)
self.config["item_fea_norm_adj"] = sparse_mx_to_torch_sparse_tensor(
item_fea_norm_adj)
plain_adj, norm_adj, mean_adj = self.dataset.get_adj_mat()
norm_adj = sparse_mx_to_torch_sparse_tensor(norm_adj)
self.config["norm_adj"] = norm_adj
self.config["num_batch"] = self.dataset.n_train // self.config[
"batch_size"] + 1
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
if self.config["validate"]:
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0], engine.model,
epoch)
else:
self.eval_engine.train_eval(None, self.dataset.test[0],
engine.model, epoch)
def train(self):
""" Main training navigator
Returns:
"""
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
if self.config["pre_train"] == 0:
print(self.config["pre_train"])
self.train_gcn()
else:
pass
self.train_mf()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
def train_mf(self):
""" Train NeuMF
Returns:
None
"""
if self.config["loss"] == "bpr":
train_loader = self.sample_generator.pairwise_negative_train_loader(
self.config["batch_size"], self.config["device_str"])
elif self.config["loss"] == "bce":
train_loader = self.sample_generator.uniform_negative_train_loader(
self.config["num_negative"],
self.config["batch_size"],
self.config["device_str"],
)
else:
raise ValueError(
f"Unsupported loss type {self.config['loss']}, try other options: 'bpr' or 'bce'"
)
self.engine = MFEngine(self.config)
self.model_save_dir = os.path.join(self.config["model_save_dir"],
self.config["save_name"])
self._train(self.engine, train_loader, self.model_save_dir)
def train_gcn(self):
""" Train GCN
Returns:
None
"""
train_loader = self.dataset
# Train GCN
self.engine = GCN_SEngine(self.config)
self.gcn_save_dir = os.path.join(
self.config["model_save_dir"],
self.config["gcn_config"]["save_name"])
self._train(self.engine, train_loader, self.gcn_save_dir)
while self.eval_engine.n_worker:
print(f"Wait 15s for the complete of eval_engine.n_worker")
time.sleep(15) # wait the
class TrainEngine(object):
"""Training engine for all the models.
"""
def __init__(self, config):
"""Initialing
Args:
config (dict): Config dict received from command line. Should have the config["config_file"].
Attributes:
dataset (Dataset): A dataset containing DataFrame of train, validation and test.
train_data (DataLoader): Extracted training data or train DataLoader, need to be implement.
monitor (Monitor): An monitor object that monitor the computational resources.
engine (Model Engine)
"""
self.dataset = None
self.train_data = None
self.monitor = None
self.engine = None
self.config = prepare_env(config)
self.gpu_id, self.config["device_str"] = self.get_device()
self.eval_engine = EvalEngine(self.config)
def get_device(self):
"""
Get one gpu id that have the most available memory.
Returns:
(int, str): The gpu id (None if no available gpu) and the the device string (pytorch style).
"""
if "device" in self.config:
if self.config["device"] == "cpu":
return (None, "cpu")
elif "cuda" in self.config[
"device"]: # receive an string with "cuda:#"
return (
int(self.config["device"].replace("cuda:", "")),
self.config["device"],
)
elif len(self.config["device"]) < 1: # receive an int string
return (
int(self.config["device"]),
"cuda:" + self.config["device"],
)
gpu_id_list = GPUtil.getAvailable(
order="memory", limit=3) # get the fist gpu with the lowest load
if len(gpu_id_list) < 1:
gpu_id = None
device_str = "cpu"
else:
gpu_id = gpu_id_list[0]
# need to set 0 if ray only specify 1 gpu
if "CUDA_VISIBLE_DEVICES" in os.environ:
if len(os.environ["CUDA_VISIBLE_DEVICES"].split()) == 1:
# gpu_id = int(os.environ["CUDA_VISIBLE_DEVICES"])
gpu_id = 0
print("Find only one gpu with id: ", gpu_id)
device_str = "cuda:" + str(gpu_id)
# print(os.system("nvidia-smi"))
else:
print("Get a gpu with the most available memory :", gpu_id)
device_str = "cuda:" + str(gpu_id)
return gpu_id, device_str
def load_dataset(self):
""" Default implementation of building dataset
Returns:
None
"""
self.dataset = data_util.Dataset(self.config)
if "item_fea_type" in self.config or "user_fea_type" in self.config:
self.config["item_fea"] = self.dataset.item_feature
self.config["user_fea"] = self.dataset.user_feature
self.config["n_users"] = self.dataset.n_users
self.config["n_items"] = self.dataset.n_items
# noinspection PyTypeChecker
def build_data_loader(self):
""" Default data builder
Returns:
DataLoader
"""
return DataLoader(
torch.LongTensor(self.train_data.to_numpy()).to(
self.engine.device),
batch_size=self.config["batch_size"],
shuffle=True,
drop_last=True,
)
def check_early_stop(self, engine, model_dir, epoch):
""" Check if early stop criterion is triggered
Save model if previous epoch have already obtained better result
Args:
epoch (int): epoch num
Returns:
True: if early stop criterion is triggered
False: else
"""
if epoch > 0 and self.eval_engine.n_no_update == 0:
# save model if previous epoch have already obtained better result
engine.save_checkpoint(model_dir=model_dir)
if self.eval_engine.n_no_update >= MAX_N_UPDATE:
# stop training if early stop criterion is triggered
print(
"Early stop criterion triggered, no performance update for {:} times"
.format(MAX_N_UPDATE))
return True
return False
def train(self):
"""Default train implementation
"""
assert hasattr(self,
"engine"), "Please specify the exact model engine !"
self.monitor = Monitor(log_dir=self.config["run_dir"],
delay=1,
gpu_id=self.gpu_id)
self.engine.data = self.dataset
print("Start training... ")
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print(f"Epoch {epoch} starts !")
print("-" * 80)
if self.check_early_stop(
engine=self.engine,
model_dir=os.path.join(self.config["model_save_dir"],
"model.cpk"),
epoch=epoch,
):
break
data_loader = self.build_data_loader()
self.engine.train_an_epoch(data_loader, epoch_id=epoch)
self.eval_engine.train_eval(self.dataset.valid[0],
self.dataset.test[0],
self.engine.model, epoch)
"""Sets the learning rate to the initial LR decayed by 10 every 10 epochs"""
lr = self.config["lr"] * (0.5**(epoch // 10))
for param_group in self.engine.optimizer.param_groups:
param_group["lr"] = lr
self.config["run_time"] = self.monitor.stop()
return self.eval_engine.best_valid_performance
def test(self):
"""Evaluate the performance for the testing sets based on the final model.
"""
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
class MF_train(TrainEngine):
""" An instance class from the TrainEngine base class
"""
def __init__(self, config):
"""Constructor
Args:
config (dict): All the parameters for the model
"""
self.config = config
super(MF_train, self).__init__(self.config)
self.load_dataset()
self.build_data_loader()
self.gpu_id, self.config["device_str"] = self.get_device()
def build_data_loader(self):
# ToDo: Please define the directory to store the adjacent matrix
self.sample_generator = SampleGenerator(ratings=self.dataset.train)
def _train(self, engine, train_loader, save_dir):
self.eval_engine.flush()
epoch_bar = tqdm(range(self.config["max_epoch"]), file=sys.stdout)
for epoch in epoch_bar:
print("Epoch {} starts !".format(epoch))
print("-" * 80)
if self.check_early_stop(engine, save_dir, epoch):
break
engine.train_an_epoch(train_loader, epoch_id=epoch)
"""evaluate model on validation and test sets"""
if self.config["validate"]:
self.eval_engine.train_eval(
self.dataset.valid[0], self.dataset.test[0], engine.model, epoch
)
else:
self.eval_engine.train_eval(
None, self.dataset.test[0], engine.model, epoch
)
def train(self):
""" Main training navigator
Returns:
"""
# Options are: 'gcn', 'mlp', 'ncf', and 'ncf_gcn';
# Train NeuMF without pre-train
self.monitor = Monitor(
log_dir=self.config["run_dir"], delay=1, gpu_id=self.gpu_id
)
self.train_mf()
self.config["run_time"] = self.monitor.stop()
self.eval_engine.test_eval(self.dataset.test, self.engine.model)
return self.eval_engine.best_valid_performance
def train_mf(self):
""" Train NeuMF
Returns:
None
"""
if self.config["loss"] == "bpr":
train_loader = self.sample_generator.pairwise_negative_train_loader(
self.config["batch_size"], self.config["device_str"]
)
elif self.config["loss"] == "bce":
train_loader = self.sample_generator.uniform_negative_train_loader(
self.config["num_negative"],
self.config["batch_size"],
self.config["device_str"],
)
else:
raise ValueError(
f"Unsupported loss type {self.config['loss']}, try other options: 'bpr' or 'bce'"
)
self.engine = MFEngine(self.config)
self.model_save_dir = os.path.join(
self.config["model_save_dir"], self.config["save_name"]
)
print(self.model_save_dir)
self._train(self.engine, train_loader, self.model_save_dir)