args.num_layers,
dropout=args.dropout,
alpha=args.alpha,
lamda=args.lamda)
labels = labels.squeeze()
# set_seed(args.seed)
optimizer = th.optim.Adam([{
'params': model.params1,
'weight_decay': args.weight_decay1
}, {
'params': model.params2,
'weight_decay': args.weight_decay2
}],
lr=args.learn_rate)
early_stopping = EarlyStopping(args.patience, file_name='tmp')
device = th.device("cuda:0" if th.cuda.is_available() else "cpu")
graph = graph.to(device)
features = features.to(device)
labels = labels.to(device)
train_mask = train_mask.to(device)
val_mask = val_mask.to(device)
test_mask = test_mask.to(device)
print(model)
model = model.to(device)
dur = []
for epoch in range(args.num_epochs):
if epoch >= 3:
t0 = time.time()
model = VSGCNetMulti(num_feats,
num_classes,
k=args.k,
dropout=args.dropout,
propagation=args.propagation)
labels = labels.squeeze()
# set_seed(args.seed)
optimizer = th.optim.Adam(model.parameters(),
lr=args.learn_rate,
weight_decay=args.weight_decay)
early_stopping = EarlyStopping(args.patience,
file_name='{}_{}'.format(
args.filename, args.dataset))
device = th.device(
"cuda:{}".format(args.cuda) if th.cuda.is_available() else "cpu")
graph = graph.to(device)
features = features.to(device)
labels = labels.to(device)
train_mask = train_mask.to(device)
val_mask = val_mask.to(device)
test_mask = test_mask.to(device)
model = model.to(device)
dur = []
for epoch in range(args.num_epochs):
def train(model,
train_data,
valid_data,
device="cpu",
save_path=None,
early_stopping=False,
opti=torch.optim.Adam,
loss=torch.nn.CrossEntropyLoss(),
max_epoch=20,
static_opti_conf=None,
scheduled_opti_conf=None,
after_update=None,
after_epoch=None,
before_update=None,
accuracy_method=None,
verbose=False):
if static_opti_conf is None and scheduled_opti_conf is None:
raise RuntimeError("No opti conf given !")
history_acc_train = list()
history_acc_valid = list()
# Move model to good device
model = model.to(device)
best_acc = 0.0
# Instance optimizer and loss
init_conf = static_opti_conf if not static_opti_conf is None else scheduled_opti_conf.get(
0, dict())
init_conf = deepcopy(init_conf)
for element in _EXTERN_PARAM:
try:
init_conf.pop(element) # remove patience term
except KeyError:
pass
optimizer = opti(model.parameters(), **init_conf)
loss = loss.to(device)
# Init early stopping if needed
if early_stopping:
if not static_opti_conf is None:
early_control = EarlyStopping(static_opti_conf.get("patience", 7),
verbose=True)
else:
early_control = EarlyStopping(static_opti_conf.get(
0, {
"patience": 7
}).get("patience", 7),
verbose=True)
mean_train_loss = 0
mean_train_accuracy = 0
mean_valid_loss = 0
mean_valid_accuracy = 0
for curr_epoch in range(max_epoch):
print("epoch : ", curr_epoch)
if not scheduled_opti_conf is None:
_update_optimizer(optimizer,
scheduled_opti_conf.get(curr_epoch, None))
#Train part
optimizer.zero_grad()
model.train()
curr_batch = 0
for x, y in train_data:
curr_batch += 1
# Switch params
x = x.to(device)
y = y.to(device)
# Model forward
outputs = model(x)
optimizer.zero_grad()
# Loss backward
loss_val = loss(outputs, y)
mean_train_loss = (mean_train_loss *
((curr_batch - 1) * train_data.batch_size) +
train_data.batch_size * loss_val.item()) / (
curr_batch * train_data.batch_size)
# accuracy if needed
if not accuracy_method is None:
mean_train_accuracy = (
(curr_batch - 1) * train_data.batch_size *
mean_train_accuracy +
train_data.batch_size * accuracy_method(outputs, y)) / (
curr_batch * train_data.batch_size)
loss_val.backward()
# Do before methods
if "before_update" in list(model.__class__.__dict__.keys()):
model.before_update()
if not before_update is None:
before_update()
optimizer.step()
# Do after methods
if "after_update" in list(model.__class__.__dict__.keys()):
model.after_update()
if not after_update is None:
after_update()
# Eval part
if "after_epoch" in list(model.__class__.__dict__.keys()):
model.after_epoch(epoch=curr_epoch)
if not after_epoch is None:
after_epoch(epoch=curr_epoch, model=model)
#model.eval()
with torch.no_grad():
curr_batch = 0
for x, y in valid_data:
curr_batch += 1
# Switch params
x = x.to(device)
y = y.to(device)
# Model forward
outputs = model(x)
# Loss backward
loss_val = loss(outputs, y)
mean_valid_loss = (mean_valid_loss *
((curr_batch - 1) * valid_data.batch_size) +
valid_data.batch_size * loss_val.item()) / (
curr_batch * valid_data.batch_size)
# accuracy if needed
if not accuracy_method is None:
batch_accuracy = accuracy_method(outputs, y)
mean_valid_accuracy = (
(curr_batch - 1) * mean_valid_accuracy *
valid_data.batch_size + valid_data.batch_size *
batch_accuracy) / (curr_batch * valid_data.batch_size)
best_acc = max(best_acc, mean_valid_accuracy)
print("train acc : ", mean_train_accuracy)
print("train loss: ", mean_train_loss)
print("valid acc : ", mean_valid_accuracy)
history_acc_train.append(mean_train_accuracy)
history_acc_valid.append(mean_valid_accuracy)
#print("train loss: ", mean_valid_loss)
if early_stopping:
if not accuracy_method is None:
early_control(model, accuracy=mean_valid_accuracy)
else:
early_control(model, loss=mean_valid_loss)
if early_control.early_stop:
break
if early_stopping:
if early_control.early_stop:
mean_valid_accuracy = early_control.best_score #TODO Reload best model
print("final acc : ", best_acc)
return history_acc_train, history_acc_valid
