def train(net, loader, criterion, optimizer, verbose=False):
net.train()
running_loss = 0
running_accuracy = 0
for i, (X, y) in enumerate(loader):
if args.cuda:
X, y = X.cuda(), y.cuda()
X, y = Variable(X), Variable(y)
output = net(X)
loss = criterion(output, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
#pdb.set_trace()
running_loss += loss.item()
acc = utils.get_multilabel_accuracy(output, y)
running_accuracy += acc
if i % 400 == 0 and verbose:
pct = float(i + 1) / len(loader)
curr_loss = running_loss / (i + 1)
curr_acc = running_accuracy / (i + 1)
print('Complete: {:.2f}, Loss: {:.2f}, Accuracy: {:.4f}'.format(
pct * 100, curr_loss, curr_acc))
return running_loss / len(loader), running_accuracy / len(loader)
def validate(net, loader, criterion):
net.eval()
running_loss = 0
running_accuracy = 0
targets = torch.FloatTensor(0, 17) # For fscore calculation
predictions = torch.FloatTensor(0, 17)
for i, (X, y) in enumerate(tqdm(loader, desc='Val')):
if args.cuda:
X, y = X.cuda(), y.cuda()
X, y = Variable(X, volatile=True), Variable(y)
output = net(X)
loss = criterion(output, y)
acc = utils.get_multilabel_accuracy(output, y)
targets = torch.cat((targets, y.cpu().data), 0)
predictions = torch.cat((predictions, output.cpu().data), 0)
running_loss += loss.item()
running_accuracy += acc
fscore = fbeta_score(targets.numpy(),
predictions.numpy() > 0.23,
beta=2,
average='samples')
return running_loss / len(loader), running_accuracy / len(loader), fscore