def train(train_loader, model, criterions, optimizer, epoch):
"""training the model"""
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
loss_protest = AverageMeter()
loss_v = AverageMeter()
protest_acc = AverageMeter()
violence_mse = AverageMeter()
visattr_acc = AverageMeter()
end = time.time()
loss_history = []
for i, sample in enumerate(train_loader):
# measure data loading batch_time
input, target = sample['image'], sample['label']
data_time.update(time.time() - end)
if args.cuda:
input = input.cuda()
for k, v in target.items():
target[k] = v.cuda()
target_var = {}
for k,v in target.items():
target_var[k] = Variable(v)
input_var = Variable(input)
output = model(input_var)
losses, scores, N_protest = calculate_loss(output, target_var, criterions)
optimizer.zero_grad()
loss = 0
for l in losses:
loss += l
# back prop
loss.backward()
optimizer.step()
if N_protest:
loss_protest.update(losses[0].item(), input.size(0))
loss_v.update(loss.item() - losses[0].item(), N_protest)
else:
# when there is no protest image in the batch
loss_protest.update(losses[0].item(), input.size(0))
loss_history.append(loss.item())
protest_acc.update(scores['protest_acc'], input.size(0))
violence_mse.update(scores['violence_mse'], N_protest)
visattr_acc.update(scores['visattr_acc'], N_protest)
batch_time.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}] '
'Time {batch_time.val:.2f} ({batch_time.avg:.2f}) '
'Data {data_time.val:.2f} ({data_time.avg:.2f}) '
'Loss {loss_val:.3f} ({loss_avg:.3f}) '
'Protest {protest_acc.val:.3f} ({protest_acc.avg:.3f}) '
'Violence {violence_mse.val:.5f} ({violence_mse.avg:.5f}) '
'Vis Attr {visattr_acc.val:.3f} ({visattr_acc.avg:.3f})'
.format(
epoch, i, len(train_loader), batch_time=batch_time,
data_time=data_time,
loss_val=loss_protest.val + loss_v.val,
loss_avg = loss_protest.avg + loss_v.avg,
protest_acc = protest_acc, violence_mse = violence_mse,
visattr_acc = visattr_acc))
return loss_history
def validate(val_loader, model, criterions, epoch):
"""Validating"""
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
loss_protest = AverageMeter()
loss_v = AverageMeter()
protest_acc = AverageMeter()
violence_mse = AverageMeter()
visattr_acc = AverageMeter()
end = time.time()
loss_history = []
for i, sample in enumerate(val_loader):
# measure data loading batch_time
input, target = sample['image'], sample['label']
if args.cuda:
input = input.cuda()
for k, v in target.items():
target[k] = v.cuda()
input_var = Variable(input)
target_var = {}
for k,v in target.items():
target_var[k] = Variable(v)
output = model(input_var)
losses, scores, N_protest = calculate_loss(output, target_var, criterions)
loss = 0
for l in losses:
loss += l
if N_protest:
loss_protest.update(losses[0].item(), input.size(0))
loss_v.update(loss.item() - losses[0].item(), N_protest)
else:
# when no protest images
loss_protest.update(losses[0].item(), input.size(0))
loss_history.append(loss.item())
protest_acc.update(scores['protest_acc'], input.size(0))
violence_mse.update(scores['violence_mse'], N_protest)
visattr_acc.update(scores['visattr_acc'], N_protest)
batch_time.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.2f} ({batch_time.avg:.2f}) '
'Loss {loss_val:.3f} ({loss_avg:.3f}) '
'Protest Acc {protest_acc.val:.3f} ({protest_acc.avg:.3f}) '
'Violence MSE {violence_mse.val:.5f} ({violence_mse.avg:.5f}) '
'Vis Attr Acc {visattr_acc.val:.3f} ({visattr_acc.avg:.3f})'
.format(
epoch, i, len(val_loader), batch_time=batch_time,
loss_val =loss_protest.val + loss_v.val,
loss_avg = loss_protest.avg + loss_v.avg,
protest_acc = protest_acc,
violence_mse = violence_mse, visattr_acc = visattr_acc))
print(' * Loss {loss_avg:.3f} Protest Acc {protest_acc.avg:.3f} '
'Violence MSE {violence_mse.avg:.5f} '
'Vis Attr Acc {visattr_acc.avg:.3f} '
.format(loss_avg = loss_protest.avg + loss_v.avg,
protest_acc = protest_acc,
violence_mse = violence_mse, visattr_acc = visattr_acc))
return loss_protest.avg + loss_v.avg, loss_history