def eval(cfg, model, loader, criterion, publisher="test"):
model.eval()
# metrics
acc_meter = MultiAssessmentMeter(num_classes=cfg.dataset.num_classes,
metrics=["class", "overall", "iou"])
batch_loss = LossMeter()
meters = (acc_meter, batch_loss)
with torch.no_grad():
for _, data in enumerate(loader):
_ = processing(model, criterion, data, meters, cfg.general.device)
# get epoch loss and accuracy
epoch_loss = batch_loss.compute()
epoch_acc = acc_meter.compute()
# save loss and acc to tensorboard
log_dict = {
"{}/loss".format(publisher): epoch_loss,
"{}/mAcc".format(publisher): epoch_acc["class"],
"{}/oAcc".format(publisher): epoch_acc["overall"],
"{}/IoU".format(publisher): epoch_acc["iou"]
}
return log_dict
def train(cfg,
model,
loader,
optimizer,
criterion,
scheduler,
publisher="train"):
model.train()
# metrics
acc_meter = MultiAssessmentMeter(num_classes=cfg.dataset.num_classes,
metrics=["class", "overall", "iou"])
batch_loss = LossMeter()
meters = (acc_meter, batch_loss)
for _, data in enumerate(loader):
optimizer.zero_grad()
loss = processing(model, criterion, data, meters, cfg.general.device)
loss.backward()
optimizer.step()
# nan
if torch.isnan(loss):
print("Training loss is nan.")
exit()
scheduler.step()
# get epoch loss and accuracy
epoch_loss = batch_loss.compute()
epoch_acc = acc_meter.compute()
# save loss and acc to tensorboard
lr = scheduler.get_last_lr()[0]
log_dict = {
"lr": lr,
"{}/loss".format(publisher): epoch_loss,
"{}/mAcc".format(publisher): epoch_acc["class"],
"{}/oAcc".format(publisher): epoch_acc["overall"],
"{}/IoU".format(publisher): epoch_acc["iou"]
}
return log_dict