def train(logbook, net, device, loss_fn, opt, train_l):
"""Run one epoch of the training experiment."""
logbook.meter.reset()
bar = FillingSquaresBar('Training \t', max=len(train_l))
controllers = indiv.Controller.getControllers(net)
for i_batch, data in enumerate(train_l):
# load data onto device
inputs, gt_labels = data
inputs = inputs.to(device)
gt_labels = gt_labels.to(device)
# forprop
pr_outs = net(inputs)
loss = loss_fn(pr_outs, gt_labels)
# update statistics
logbook.meter.update(pr_outs, gt_labels, loss.item(), track_metric=logbook.track_metric)
bar.suffix = 'Total: {total:} | ETA: {eta:} | Epoch: {epoch:4d} | ({batch:5d}/{num_batches:5d})'.format(
total=bar.elapsed_td,
eta=bar.eta_td,
epoch=logbook.i_epoch,
batch=i_batch + 1,
num_batches=len(train_l))
bar.suffix = bar.suffix + logbook.meter.bar()
bar.next()
# backprop
opt.zero_grad()
loss.backward()
opt.step()
for ctrl in controllers:
ctrl.step_postOptimStep()
bar.finish()
stats = {
'train_loss': logbook.meter.avg_loss,
'train_metric': logbook.meter.avg_metric
}
for k, v in stats.items():
if v:
logbook.writer.add_scalar(k, v, global_step=logbook.i_epoch)
logbook.writer.add_scalar('learning_rate', opt.param_groups[0]['lr'], global_step=logbook.i_epoch)
return stats
def test(logbook, net, device, loss_fn, test_l, valid=False, prefix=None):
"""Run a validation epoch."""
logbook.meter.reset()
bar_title = 'Validation \t' if valid else 'Test \t'
bar = FillingSquaresBar(bar_title, max=len(test_l))
with torch.no_grad():
for i_batch, data in enumerate(test_l):
# load data onto device
inputs, gt_labels = data
inputs = inputs.to(device)
gt_labels = gt_labels.to(device)
# forprop
tensor_stats, pr_outs = net.forward_with_tensor_stats(inputs)
loss = loss_fn(pr_outs, gt_labels)
# update statistics
logbook.meter.update(pr_outs, gt_labels, loss.item(), track_metric=True)
bar.suffix = 'Total: {total:} | ETA: {eta:} | Epoch: {epoch:4d} | ({batch:5d}/{num_batches:5d})'.format(
total=bar.elapsed_td,
eta=bar.eta_td,
epoch=logbook.i_epoch,
batch=i_batch + 1,
num_batches=len(test_l))
bar.suffix = bar.suffix + logbook.meter.bar()
bar.next()
bar.finish()
if prefix == None:
prefix = 'valid' if valid else 'test'
stats = {
prefix+'_loss': logbook.meter.avg_loss,
prefix+'_metric': logbook.meter.avg_metric
}
if valid:
for k, v in stats.items():
if v:
logbook.writer.add_scalar(k, v, global_step=logbook.i_epoch)
for name, tensor in tensor_stats:
logbook.writer.add_histogram(name, tensor, global_step=logbook.i_epoch)
return stats