def train_model(cnn, criterion, optimizer, num_epochs=100):
loss_history = []
train_acc_history = []
val_acc_history = []
epoch_history = []
best_val_acc = 0.0
p = 0.5
learning_rate = [0.001]
for lr in learning_rate:
lr_msg = 'Learning Rate for this model: {}'.format(lr)
print(lr_msg)
epoch_history.append(lr_msg)
p_msg = 'Dropout p for this model: {}'.format(p)
print(p_msg)
epoch_history.append(p_msg)
cnn = CNN(p)
if use_gpu:
cnn.cuda()
for epoch in range(num_epochs):
optimizer = torch.optim.SGD(cnn.parameters(), lr=lr, momentum=0.9)
print('Starting epoch %d / %d' % (epoch + 1, num_epochs))
# print('Learning Rate for this epoch: {}'.format(learning_rate))
print('Learning Rate for this epoch: {}'.format(lr))
i = 0
for gender_labels, race_labels, img_names, images in train_loader:
gender_labels = torch.from_numpy(np.asarray(gender_labels))
race_labels = torch.from_numpy(np.asarray(race_labels))
images = Variable(images)
labels = Variable(gender_labels)
race_labels = Variable(race_labels)
if use_gpu:
images, labels = images.cuda(), labels.cuda()
race_labels = race_labels.cuda()
pred_labels, _, _ = cnn(images)
loss = criterion(pred_labels, labels)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
if (i + 1) % 5 == 0:
print(
'Epoch [%d/%d], Iter [%d/%d] CNN Loss: %.4f Adversary Loss: %.4f'
% (epoch + 1, num_epochs, i + 1, len(train_data) // 50,
loss.data, nn_loss.data))
i = i + 1
if epoch == 0 or epoch % 5 == 0 or epoch == num_epochs - 1:
train_acc = check_acc(cnn, train_loader)
train_acc_history.append(train_acc)
train_msg = 'Train accuracy for epoch {}: {} '.format(
epoch + 1, train_acc)
print(train_msg)
epoch_history.append(train_msg)
val_acc = check_acc(cnn, test_loader)
val_acc_history.append(val_acc)
val_msg = 'Validation accuracy for epoch {}: {} '.format(
epoch + 1, val_acc)
print(val_msg)
epoch_history.append(val_msg)
is_best = val_acc > best_val_acc
best_val_acc = max(val_acc, best_val_acc)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': cnn.state_dict(),
'best_val_acc': best_val_acc,
'optimizer': optimizer.state_dict()
}, is_best)
np.savetxt("training_log_cnn.out", epoch_history, fmt='%s')
transforms.CenterCrop(227),
transforms.ToTensor()
])
print('Loading images...')
batch_size = 50
root='UTKFace/val'
val_data = gender_race_dataset("val_labels_all.csv", root, test_transform)
val_loader = torch.utils.data.DataLoader(val_data,
batch_size=batch_size,shuffle=False)
if model == "cnn":
print("Using CNN")
cnn = CNN()
if use_gpu:
cnn.cuda()
optimizer = torch.optim.SGD(cnn.parameters(),lr=0.001,momentum=0.9)
SAVED_MODEL_PATH = 'cnn_model_best_SGD_adversary.pth.tar'
checkpoint = torch.load(SAVED_MODEL_PATH)
cnn.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
epoch = checkpoint['epoch']
best_val_acc = checkpoint['best_val_acc']
print("best model saved from epoch: ", epoch)
print("best val_acc = ", best_val_acc)
batch_size = 50
validation_split = .2
shuffle_dataset = True
random_seed = 42
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size,
shuffle=True)
val_loader = torch.utils.data.DataLoader(val_data,
batch_size=batch_size,
shuffle=False)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=batch_size,
shuffle=False)
cnn = CNN()
if use_gpu:
print('Using GPU for cnn')
cnn.cuda()
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(cnn.parameters(), lr=0.001, momentum=0.9)
def train_model(cnn, criterion, optimizer, num_epochs=100):
loss_history = []
train_acc_history = []
val_acc_history = []
epoch_history = []
best_val_acc = 0.0
p = 0.5
def train_model(cnn, adversary, criterion, nn_criterion, optimizer, nn_optimizer, num_epochs = 100):
loss_history = []
train_acc_history = []
val_acc_history = []
epoch_history = []
# learning_rate = 0.001
# learning_rate = np.logspace(-6,-2, num=15)
# learning_rate = np.logspace(-3,-2, num=8)
best_val_acc = 0.0
# alpha = 1.0
# p_vals = np.array([0.5])
p = 0.5
learning_rate = [0.0026826957952797246] #0.001 for Adam
alpha = [0.9]
# alpha = np.array([0.9, 1.0])
for layer in range(1,2):
for lr in learning_rate:
for a in alpha:
layer_msg = 'Layer for this model: {}'.format(layer)
print(layer_msg)
epoch_history.append(layer_msg)
lr_msg = 'Learning Rate for this model: {}'.format(lr)
print(lr_msg)
epoch_history.append(lr_msg)
p_msg = 'Dropout p for this model: {}'.format(p)
print(p_msg)
epoch_history.append(p_msg)
a_msg = 'Alpha for this model: {}'.format(a)
print(a_msg)
epoch_history.append(a_msg)
cnn = CNN(p)
cnn.cuda()
adversary = NN(512)
adversary.cuda()
for epoch in range(num_epochs):
optimizer = torch.optim.SGD(cnn.parameters(),lr=lr,momentum=0.9)
nn_optimizer = torch.optim.SGD(adversary.parameters(),lr=lr,momentum=0.9)
# optimizer = torch.optim.Adam(cnn.parameters(), lr=lr, betas=(0.9, 0.999))
# nn_optimizer = torch.optim.Adam(cnn.parameters(), lr=lr, betas=(0.9, 0.999))
print('Starting epoch %d / %d' % (epoch + 1, num_epochs))
# print('Learning Rate for this epoch: {}'.format(learning_rate))
# print('Learning Rate for this epoch: {}'.format(lr))
i = 0
for gender_labels, race_labels, img_names, images in train_loader:
gender_labels = torch.from_numpy(np.asarray(gender_labels))
race_labels = torch.from_numpy(np.asarray(race_labels))
images = Variable(images)
labels = Variable(gender_labels)
race_labels = Variable(race_labels)
if use_gpu:
images,labels = images.cuda(),labels.cuda()
race_labels = race_labels.cuda()
pred_labels, penultimate_weights, layer_2 = cnn(images)
if layer == 1:
nn_pred_labels = adversary(penultimate_weights)
if layer == 2:
nn_pred_labels = adversary(layer_2)
nn_loss = nn_criterion(nn_pred_labels, race_labels)
cnn_loss = criterion(pred_labels,labels)
#loss for gender prediction model
# loss = cnn_loss - alpha*nn_loss
loss = cnn_loss - a*nn_loss
optimizer.zero_grad()
loss.backward(retain_graph = True)
optimizer.step()
#loss for adversary model
nn_optimizer.zero_grad()
nn_loss.backward()
nn_optimizer.step()
if (i+1) % 5 == 0:
print ('Epoch [%d/%d], Iter [%d/%d] CNN Loss: %.4f Adversary Loss: %.4f'
%(epoch+1, num_epochs, i+1, len(train_data)//50, loss.data, nn_loss.data))
i = i + 1
# if epoch % 10 == 0:
# learning_rate = learning_rate * 0.9
# if epoch % 5 ==0 or epoch == num_epochs-1:
train_acc = check_acc(cnn,train_loader)
train_acc_history.append(train_acc)
train_msg = 'Train accuracy for epoch {}: {} '.format(epoch + 1,train_acc)
print(train_msg)
epoch_history.append(train_msg)
val_acc = check_acc(cnn,test_loader)
val_acc_history.append(val_acc)
val_msg = 'Validation accuracy for epoch {}: {} '.format(epoch + 1,val_acc)
print(val_msg)
epoch_history.append(val_msg)
#plot_performance_curves(train_acc_history,val_acc_history,epoch_history)
is_best = val_acc > best_val_acc
best_val_acc = max(val_acc,best_val_acc)
save_checkpoint(
{'epoch':epoch+1,
'state_dict':cnn.state_dict(),
'best_val_acc':best_val_acc,
'optimizer':optimizer.state_dict()},is_best)
np.savetxt("training_log_cnn_SGD.out", epoch_history, fmt='%s')