def train_baselines():
train_data, val_data = get_dataloader(96)
for model, batch in zip(models, batch_size):
name = str(model).split()[1]
print('*****Start Training {} with batch size {}******'.format(
name, batch))
print(
' epoch iter rate | smooth_loss | train_loss (acc) | valid_loss (acc) | total_train_loss\n'
)
logger = Logger(
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/log/full_data_{}_10xlr_2'
.format(name), name)
# load pre-trained model on train-37479
net = model(pretrained=True)
net = nn.DataParallel(net.cuda())
# load_net(net, name)
# optimizer = get_optimizer(net.module, lr=.005, pretrained=True, resnet=True if 'resnet' in name else False)
optimizer = get_optimizer(net.module,
lr=.01,
pretrained=True,
resnet=True if 'resnet' in name else False)
train_data.batch_size = batch
val_data.batch_size = batch
num_epoches = 60
print_every_iter = 20
epoch_test = 1
smooth_loss = 0.0
train_loss = np.nan
train_acc = np.nan
best_test_loss = np.inf
t = time.time()
for epoch in range(
num_epoches): # loop over the dataset multiple times
# train loss averaged every epoch
total_epoch_loss = 0.0
# lr_schedule(epoch, optimizer, base_lr=0.005, pretrained=True)
new_lr_schedule(epoch, optimizer)
rate = get_learning_rate(optimizer)[0] # check
sum_smooth_loss = 0.0
total_sum = 0
sum = 0
net.cuda().train()
num_its = len(train_data)
for it, (images, labels, indices) in enumerate(train_data, 0):
logits = net(Variable(images.cuda()))
probs = F.sigmoid(logits)
loss = multi_criterion(logits, labels.cuda())
optimizer.zero_grad()
loss.backward()
optimizer.step()
# additional metrics
sum_smooth_loss += loss.data[0]
total_epoch_loss += loss.data[0]
sum += 1
total_sum += 1
# print statistics
if it % print_every_iter == print_every_iter - 1:
smooth_loss = sum_smooth_loss / sum
sum_smooth_loss = 0.0
sum = 0
train_acc = multi_f_measure(probs.data, labels.cuda())
train_loss = loss.data[0]
print(
'\r{} {} {} | {} | {} {} | ... '.format(
epoch + it / num_its, it + 1, rate, smooth_loss,
train_loss, train_acc), )
total_epoch_loss = total_epoch_loss / total_sum
if epoch % epoch_test == epoch_test - 1 or epoch == num_epoches - 1:
net.cuda().eval()
test_loss, test_acc = evaluate(net, val_data)
print('\r')
print('{} {} {} | {} | {} {} | {} {} | {}'.format(
epoch + 1, it + 1, rate, smooth_loss, train_loss,
train_acc, test_loss, test_acc, total_epoch_loss))
# save if the current loss is better
if test_loss < best_test_loss:
print('save {} {}'.format(test_loss, best_test_loss))
torch.save(
net.state_dict(),
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/model/full_data_{}_10xlr_2.pth'
.format(name))
best_test_loss = test_loss
logger.add_record('train_loss', total_epoch_loss)
logger.add_record('evaluation_loss', test_loss)
logger.add_record('f2_score', test_acc)
logger.save()
logger.save_plot()
logger.save_time(start_time=t, end_time=time.time())
def train_baselines():
train_data, val_data = get_dataloader(96)
for model, batch in zip(models, batch_size):
name = str(model).split()[1]
print('*****Start Training {} with batch size {}******'.format(
name, batch))
print(
' epoch iter rate | smooth_loss | train_loss (acc) | valid_loss (acc) | total_train_loss\n'
)
logger = Logger('../log/{}'.format(name), name)
net = model(pretrained=True)
optimizer = get_optimizer(net,
lr=.01,
pretrained=True,
resnet=True if 'resnet' in name else False)
net = nn.DataParallel(net.cuda())
train_data.batch_size = batch
val_data.batch_size = batch
num_epoches = 50 #100
print_every_iter = 20
epoch_test = 1
# optimizer
# optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=0.0005)
# optimizer = optim.Adam(net.parameters(), lr=1e-4, weight_decay=5e-4)
smooth_loss = 0.0
train_loss = np.nan
train_acc = np.nan
# test_loss = np.nan
best_test_loss = np.inf
# test_acc = np.nan
t = time.time()
for epoch in range(
num_epoches): # loop over the dataset multiple times
# train loss averaged every epoch
total_epoch_loss = 0.0
lr_schedule(epoch, optimizer, pretrained=True)
rate = get_learning_rate(optimizer)[0] # check
sum_smooth_loss = 0.0
total_sum = 0
sum = 0
net.cuda().train()
num_its = len(train_data)
for it, (images, labels, indices) in enumerate(train_data, 0):
logits = net(Variable(images.cuda()))
probs = F.sigmoid(logits)
loss = multi_criterion(logits, labels.cuda())
optimizer.zero_grad()
loss.backward()
optimizer.step()
# additional metrics
sum_smooth_loss += loss.data[0]
total_epoch_loss += loss.data[0]
sum += 1
total_sum += 1
# print statistics
if it % print_every_iter == print_every_iter - 1:
smooth_loss = sum_smooth_loss / sum
sum_smooth_loss = 0.0
sum = 0
train_acc = multi_f_measure(probs.data, labels.cuda())
train_loss = loss.data[0]
print('\r{} {} {} | {} | {} {} | ... '.format(
epoch + it / num_its, it + 1, rate, smooth_loss,
train_loss, train_acc),
end='',
flush=True)
total_epoch_loss = total_epoch_loss / total_sum
if epoch % epoch_test == epoch_test - 1 or epoch == num_epoches - 1:
net.cuda().eval()
test_loss, test_acc = evaluate(net, val_data)
print('\r', end='', flush=True)
print('{} {} {} | {} | {} {} | {} {} | {}'.format(
epoch + 1, it + 1, rate, smooth_loss, train_loss,
train_acc, test_loss, test_acc, total_epoch_loss))
# save if the current loss is better
if test_loss < best_test_loss:
print('save {} {}'.format(test_loss, best_test_loss))
torch.save(net.state_dict(),
'../models/{}.pth'.format(name))
net.load_state_dict(
torch.load('../models/{}.pth'.format(name)))
print(evaluate(net, val_data))
best_test_loss = test_loss
logger.add_record('train_loss', total_epoch_loss)
logger.add_record('evaluation_loss', test_loss)
logger.add_record('f2_score', test_acc)
logger.save()
logger.save_plot()
logger.save_time(start_time=t, end_time=time.time())
