def train_and_save_scores(epoch):
net.train()
train_loss = 0
correct = 0
total = 0
abstain = 0
# for saving purpose
global training_loss
global train_abstained
global train_acc
global train_acc_without_abstained
global train_f1score
global train_f1score_without_abstained
global train_predicted
train_softmax_scores = []
total = 0
abstained = 0
loss_acc = np.array([])
batch_abstained = np.array([])
batch_true = np.array([])
batch_true_without_abstained = np.array([])
batch_f1score = np.array([])
batch_f1score_without_abstained = np.array([])
if args.dataset == 'mnist':
if int(epoch/args.epdl) > 5 and int(epoch/args.epdl) <= 20:
args.lr = 0.01
if int(epoch/args.epdl) > 20 and int(epoch/args.epdl) <=50:
args.lr = 0.001
#optimizer = optim.SGD(net.parameters(), lr=cf.learning_rate(args.lr, epoch), momentum=0.9,
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9,
nesterov=args.nesterov, weight_decay=5e-4)
print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, args.lr))
else: #cifar 10/100/stl-10/tin200/fashion
optimizer = optim.SGD(net.parameters(), lr=cf.learning_rate(args.lr, int(epoch/args.epdl)), momentum=0.9, weight_decay=5e-4,nesterov=args.nesterov)
# optimizer = optim.Adam(net.parameters(), lr=cf.learning_rate(args.lr, int(epoch/args.epdl)))
print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, cf.learning_rate(args.lr, int(epoch/args.epdl))))
print('\n => Training Epoch #%d, LR=%.4f' %(epoch, cf.learning_rate(args.lr, epoch)))
sys.stdout.flush()
#pdb.set_trace()
for batch_idx, (inputs, targets) in enumerate(trainloader):
#print(dir(inputs.cuda))
#quit()
optimizer.zero_grad()
inputs, targets = Variable(inputs.to(device)), Variable(targets.to(device))
if (args.net_type == 'tsc-lstm'): # update input to match lstm
inputs = inputs.view(-1, series_length, 1) # input dimensions for time series data set are 1
# print(inputs.shape)
outputs = net(inputs) # Forward Propagation
#pdb.set_trace()
if args.loss_fn is None:
loss = criterion(outputs, targets)
else:
loss = criterion(outputs, targets, epoch) # Loss
loss.backward() # Backward Propagation
optimizer.step() # Optimizer update
train_loss += loss.data.item()
_, predicted = torch.max(outputs.data, 1)
this_batch_size =targets.size(0)
total += this_batch_size
correct += predicted.eq(targets.data).cpu().sum().data.item()
abstained_now = predicted.eq(abstain_class_id).sum().data.item()
abstain += abstained_now
if total-abstain != 0:
#pdb.set_trace()
abst_acc = 100.*correct/(float(total-abstain))
else:
abst_acc = 1.
p_out = F.softmax(outputs,dim=1)
#pdb.set_trace()
total += p_out.size(0)
_,predicted = torch.max(p_out.data,1)
abstained += predicted.eq(abstain_class_id).sum().data.item()
train_softmax_scores.append(p_out.data)
loss_acc = np.append(loss_acc, loss.data.item())
f1score = f1_score(targets.cpu().numpy(), predicted.cpu().numpy(), average='weighted')
# TODO: check if it works (get accuracy without abstension)
batch_total = p_out.size(0)
batch_abstained = np.append(batch_abstained, predicted.eq(abstain_class_id).sum().data.item() / batch_total)
batch_true = np.append(batch_true, predicted.eq(targets.data).cpu().sum().data.item() / batch_total)
batch_true_without_abstained = np.append(batch_true_without_abstained, abst_acc)
batch_f1score = np.append(batch_f1score, f1score)
non_abstained_index = np.invert(predicted.eq(abstain_class_id).cpu().detach().numpy()).nonzero()[0]
f1score_without_abstained = f1_score(targets.cpu().numpy()[non_abstained_index], predicted.cpu().numpy()[non_abstained_index], average='weighted')
batch_f1score_without_abstained = np.append(batch_f1score_without_abstained, f1score_without_abstained)
#train_predicted = np.append(train_predicted, predicted.cpu())
#print('\r')
#print('| Epoch [%3d/%3d] Iter[%3d/%3d]\t\tAbstained %d Abstention rate %.4f Cumulative Abstention Rate: %.4f Loss: %.4f Acc@1: %.3f%% Acc@2: %.3f%%'
#%(epoch, num_epochs, batch_idx+1,
# (len(trainset)//batch_size)+1, abstain, float(abstained_now)/this_batch_size, float(abstain)/total, loss.data.item(), 100.*correct/float(total), abst_acc))
#sys.stdout.flush()
print('\r')
print('| Epoch [%3d/%3d] Iter[%3d/%3d]\t\tAbstained %d Abstention rate %.4f Cumulative Abstention Rate: %.4f Loss: %.4f Acc@1: %.3f%% Acc@2: %.3f%%'
%(epoch, num_epochs, batch_idx+1,
(len(trainset)//batch_size)+1, abstain, float(abstained_now)/this_batch_size, float(abstain)/total, loss.data.item(), 100.*correct/float(total), abst_acc))
sys.stdout.flush()
training_loss = np.append(training_loss, np.mean(loss_acc))
train_abstained = np.append(train_abstained, np.mean(batch_abstained))
train_acc = np.append(train_acc, np.mean(batch_true))
train_acc_without_abstained = np.append(train_acc_without_abstained, np.mean(batch_true_without_abstained))
train_scores = torch.cat(train_softmax_scores).cpu().numpy()
train_f1score = np.append(train_f1score, np.mean(batch_f1score))
train_f1score_without_abstained = np.append(train_f1score_without_abstained, np.mean(batch_f1score_without_abstained))
if args.save_train_scores:
print('Saving train softmax scores at Epoch %d' %(epoch))
fn = args.expt_name if args.expt_name else 'test'
np.save(args.output_path+fn+"/"+ fn +".train_scores.epoch_"+str(epoch), train_scores)
print("\n##### Epoch %d Train Abstention Rate at end of epoch %.4f"
%(epoch, float(abstained)/total))
def train(epoch):
net.train()
train_loss = 0
correct = 0
total = 0
abstain = 0
if args.dataset == 'mnist':
if int(epoch/args.epdl) > 5 and int(epoch/args.epdl) <= 20:
args.lr = 0.01
if int(epoch/args.epdl) > 20 and int(epoch/args.epdl) <=50:
args.lr = 0.001
#optimizer = optim.SGD(net.parameters(), lr=cf.learning_rate(args.lr, epoch), momentum=0.9,
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.9,
nesterov=args.nesterov, weight_decay=5e-4)
print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, args.lr))
else: #cifar 10/100/stl-10/tin200/fashion
optimizer = optim.SGD(net.parameters(), lr=cf.learning_rate(args.lr, int(epoch/args.epdl)),
momentum=0.9, weight_decay=5e-4,nesterov=args.nesterov)
print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, cf.learning_rate(args.lr, int(epoch/args.epdl))))
#print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, cf.learning_rate(args.lr, epoch)))
#pdb.set_trace()
for batch_idx, (inputs, targets) in enumerate(trainloader):
#print(dir(inputs.cuda))
#quit()
optimizer.zero_grad()
inputs, targets = Variable(inputs), Variable(targets)
if (args.net_type == 'tsc-lstm'): # update input to match lstm
inputs = inputs.view(-1, series_length, 1) # input dimensions for time series data set are 1
outputs = net(inputs) # Forward Propagation
#pdb.set_trace()
if args.loss_fn is None:
loss = criterion(outputs, targets)
else:
loss = criterion(outputs, targets, epoch) # Loss
loss.backward() # Backward Propagation
optimizer.step() # Optimizer update
train_loss += loss.data.item()
_, predicted = torch.max(outputs.data, 1)
this_batch_size =targets.size(0)
total += this_batch_size
correct += predicted.eq(targets.data).cpu().sum().data.item()
abstained_now = predicted.eq(abstain_class_id).sum().data.item()
abstain += abstained_now
if total-abstain != 0:
#pdb.set_trace()
abst_acc = 100.*correct/(float(total-abstain))
else:
abst_acc = 1.
sys.stdout.write('\r')
sys.stdout.write('| Epoch [%3d/%3d] Iter[%3d/%3d]\t\tAbstained %d Abstention rate %.4f Cumulative Abstention Rate: %.4f Loss: %.4f Acc@1: %.3f%% Acc@2: %.3f%%'
%(epoch, num_epochs, batch_idx+1,
(len(trainset)//batch_size)+1, abstain, float(abstained_now)/this_batch_size, float(abstain)/total, loss.data.item(), 100.*correct/float(total), abst_acc))
sys.stdout.flush()
def train(epoch):
net.train()
train_loss = 0
correct = 0
total = 0
abstain = 0
if args.dataset == 'mnist':
if int(epoch / args.epdl) > 5 and int(epoch / args.epdl) < 20:
args.lr = 0.01
if int(epoch / args.epdl) >= 20:
args.lr = 0.001
#optimizer = optim.SGD(net.parameters(), lr=cf.learning_rate(args.lr, epoch), momentum=0.9,
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=0.9,
nesterov=args.nesterov,
weight_decay=5e-4)
print('\n=> Training Epoch #%d, LR=%.4f' % (epoch, args.lr))
else: #cifar 10/100/stl-10/tin200/fashion
optimizer = optim.SGD(net.parameters(),
lr=cf.learning_rate(args.lr,
int(epoch / args.epdl)),
momentum=0.9,
weight_decay=5e-4,
nesterov=args.nesterov)
print('\n=> Training Epoch #%d, LR=%.4f' %
(epoch, cf.learning_rate(args.lr, int(epoch / args.epdl))))
#print('\n=> Training Epoch #%d, LR=%.4f' %(epoch, cf.learning_rate(args.lr, epoch)))
for batch_idx, (inputs, targets) in enumerate(trainloader):
#print(type(inputs))
#print(dir(inputs.cuda))
#quit()
if use_cuda:
#pdb.set_trace()
inputs, targets = inputs.cuda(cuda_device), targets.cuda(
cuda_device) # GPU settings
#inputs, targets = inputs.cuda(), targets.cuda() # GPU settings
optimizer.zero_grad()
inputs, targets = Variable(inputs), Variable(targets)
outputs = net(inputs) # Forward Propagation
if args.loss_fn is None:
loss = criterion(outputs, targets)
else:
loss = criterion(outputs, targets, epoch) # Loss
loss.backward() # Backward Propagation
optimizer.step() # Optimizer update
train_loss += loss.data.item()
_, predicted = torch.max(outputs.data, 1)
total += targets.size(0)
correct += predicted.eq(targets.data).cpu().sum().data.item()
abstain += predicted.eq(abstain_class_id).sum().data.item()
if total - abstain != 0:
#pdb.set_trace()
abst_acc = 100. * correct / (float(total - abstain))
else:
abst_acc = 1.
sys.stdout.write('\r')
sys.stdout.write(
'| Epoch [%3d/%3d] Iter[%3d/%3d]\t\tAbstained %d Loss: %.4f Acc@1: %.3f%% Acc@2: %.3f%%'
% (epoch, num_epochs, batch_idx + 1,
(len(trainset) // batch_size) + 1, abstain, loss.data.item(),
100. * correct / float(total), abst_acc))
sys.stdout.flush()
