def validate(model, val_loader, epoch, callback=None, **kwargs):
model.eval()
global writer
loss = AverageMeter()
metrics = AverageMeter()
data_time = time.time()
val_bar = tqdm(val_loader)
devices = 'cuda:0' if 'devices' not in kwargs.keys(
) else 'cuda:' + kwargs['devices'][0]
with torch.no_grad():
for idx, (data, target) in enumerate(val_bar):
data = data.to(devices)
target = target.to(devices)
data_time = time.time() - data_time
# forward
batch_time = time.time()
pred = model(data)
batch_time = time.time() - batch_time
batch_loss = model.loss(pred, target, mode=kwargs['mode'])
metrics.update(accuracy(pred, target), data.size(0))
loss.update(batch_loss.item(), data.size(0))
val_bar.set_description(
'Eval epoch {}/{}, data_time: {:.4f}, batch_time: {:.4f}, loss: {:.4f}, accuracy: {:.4f}'
.format(idx, epoch, data_time, batch_time, loss.avg,
metrics.avg.item()))
data_time = time.time()
return loss.avg, metrics.avg.item()
def train_epoch(model,
optimizer,
train_loader,
epoch,
callback=None,
**kwargs):
global writer
loss = AverageMeter()
metrics = AverageMeter()
data_time = time.time()
train_bar = tqdm(train_loader)
devices = 'cuda:0' if 'devices' not in kwargs.keys(
) else 'cuda:' + kwargs['devices'][0]
for idx, (data, target) in enumerate(train_bar):
data = data.to(devices)
target = target.to(devices)
data_time = time.time() - data_time
# forward
batch_time = time.time()
pred = model(data)
batch_time = time.time() - batch_time
batch_loss = model.loss(pred, target, mode=kwargs['mode'])
metrics.update(accuracy(pred, target), data.size(0))
loss.update(batch_loss.item(), data.size(0))
# backward
optimizer.zero_grad()
batch_loss.backward()
optimizer.step()
# print(metrics.avg,loss.avg,type(loss.avg))
train_bar.set_description(
'Epoch {}/{}, data_time: {:.4f}, batch_time: {:.4f}, loss: {:.4f}, accuracy: {:.4f}'
.format(idx, epoch, data_time, batch_time, loss.avg,
metrics.avg.item()))
data_time = time.time()
return loss.avg, metrics.avg.item()