def Eval_I2P():
eval_loss = 0
eval_acc = 0
eval_data = I2P_data(os.path.join(dataroot, 'valid'), dataset_valid_I2P)
data_eval = torch.utils.data.DataLoader(eval_data,
batch_size=opt.batchSize,
pin_memory=True,
num_workers=4,
shuffle=True)
with torch.no_grad():
model.eval()
for i, (Inputf, Inputr, Inputt, Inputa, Label) in enumerate(data_eval):
Inputf, Inputr, Inputt, Inputa, Label = Inputf.to(
device), Inputr.to(device), Inputt.to(device), Inputa.to(
device), Label.to(device)
y = F.normalize(model(Inputf.float(), Inputr.float(),
Inputt.float(), Inputa.float()),
dim=1)
y = torch.cat((y[:, :2], y[:, 4:6]), axis=1)
Label = Label.reshape(Inputf.shape[0])
loss = criterion_I2P(y, Label)
eval_loss += loss.item() * Inputf.shape[0]
eval_acc += (y.argmax(1) == Label).sum().item()
epoch_eval_loss = eval_loss / len(eval_data)
epoch_eval_acc = eval_acc / len(eval_data)
return epoch_eval_loss, epoch_eval_acc
def train_model():
batch_loss = 0
batch_acc = 0
batch_loss_list = []
path = opt.Training_dataroot
train_data = I2P_data(path)
data_train = torch.utils.data.DataLoader(train_data,
batch_size=opt.batchSize,
pin_memory=True,
num_workers=4,
shuffle=True)
task_4_model.train()
for i, (Dic, Inputf, Inputr, Inputt, Inputa,
Label) in enumerate(data_train):
optimizer.zero_grad()
Inputf, Inputr, Inputt, Inputa, Label = Inputf.to(device), Inputr.to(
device), Inputt.to(device), Inputa.to(device), Label.to(device)
y = task_4_model(Inputf.float(), Inputr.float(), Inputt.float(),
Inputa.float())
Label = Label.reshape(Inputf.shape[0])
loss = criterion(y, Label)
batch_loss += loss.item() * Inputf.shape[0]
batch_loss_list.append(loss.item() * Inputf.shape[0] / len(train_data))
loss.backward()
#optimizer.module.step()
optimizer.step()
batch_acc += (y.argmax(1) == Label).sum().item()
epoch_loss = batch_loss / len(train_data)
epoch_acc = batch_acc / len(train_data)
return epoch_loss, epoch_acc, np.array(batch_loss_list)
def Eval():
eval_loss = 0
eval_acc = 0
path = opt.Validating_dataroot
eval_data = I2P_data(path)
data_eval = torch.utils.data.DataLoader(eval_data,
batch_size=opt.batchSize,
pin_memory=True,
num_workers=4,
shuffle=True)
failed_shape = []
with torch.no_grad():
task_4_model.eval()
for i, (Dic, Inputf, Inputr, Inputt, Inputa,
Label) in enumerate(data_eval):
Inputf, Inputr, Inputt, Inputa, Label = Inputf.to(
device), Inputr.to(device), Inputt.to(device), Inputa.to(
device), Label.to(device)
y = task_4_model(Inputf.float(), Inputr.float(), Inputt.float(),
Inputa.float())
Label = Label.reshape(Inputf.shape[0])
loss = criterion(y, Label)
Dic = np.array(Dic)
Dic = Dic[(y.argmax(1) != Label).cpu()]
for dic in Dic:
failed_shape.append(dic)
eval_loss += loss.item() * Inputf.shape[0]
eval_acc += (y.argmax(1) == Label).sum().item()
epoch_eval_loss = eval_loss / len(eval_data)
epoch_eval_acc = eval_acc / len(eval_data)
return failed_shape, epoch_eval_loss, epoch_eval_acc
def Train():
train_loss = 0
train_acc = 0
train_data = I2P_data(os.path.join(dataroot, 'train'), dataset_train_I2P)
data_train = torch.utils.data.DataLoader(train_data,
batch_size=opt.batchSize,
pin_memory=True,
num_workers=8,
shuffle=True)
model.train()
for i, (Inputf, Inputr, Inputt, Inputa, Label) in enumerate(data_train):
Inputf, Inputr, Inputt, Inputa, Label = Inputf.to(device), Inputr.to(
device), Inputt.to(device), Inputa.to(device), Label.to(device)
y = F.normalize(model(Inputf.float(), Inputr.float(), Inputt.float(),
Inputa.float()),
dim=1)
Label = Label.reshape(Inputf.shape[0])
loss = criterion_I2P(y, Label)
train_loss += loss.item() * Inputf.shape[0]
train_acc += (y.argmax(1) == Label).sum().item()
loss.backward()
optimizer.step()
epoch_train_loss = train_loss / len(train_data)
epoch_train_acc = train_acc / len(train_data)
return epoch_train_loss, epoch_train_acc