def subVal(model, criterion, dataLoader, device):
# set to eval mode
model.eval()
intersectionMeter = common.AverageMeter()
unionMeter = common.AverageMeter()
targetMeter = common.AverageMeter()
lossMeter = common.AverageMeter()
for i, (x, y) in enumerate(dataLoader):
x = x.to(device)
y = y.to(device)
out = model(x)
mainLoss = criterion(out[0], y)
# update val loss
lossMeter.update(mainLoss.item(), x.shape[0])
# compute IoU/accuracy
result = out[0].max(1)[1]
intersection, union, target = common.intersectionAndUnionGPU(result, y, numClasses, 255)
intersection, union, target = intersection.cpu().numpy(), union.cpu().numpy(), target.cpu().numpy()
## update meter
intersectionMeter.update(intersection), unionMeter.update(union), targetMeter.update(target)
# show the log
IoU = intersectionMeter.sum/(unionMeter.sum + 1e-10)
accuracy = intersectionMeter.sum/(targetMeter.sum + 1e-10)
print(f'val loss:',lossMeter.avg)
for i in range(numClasses):
print('Class_'+str(i)+' IoU:',IoU[i],' acc:',accuracy[i])
def subTrain(model, optimizer, criterion, dataLoader, currentepoch, maxIter):
# set to train mode to enable dropout and bn
model.train()
intersectionMeter = common.AverageMeter()
unionMeter = common.AverageMeter()
targetMeter = common.AverageMeter()
lossMeter = common.AverageMeter()
for i, (x, y) in enumerate(dataLoader):
x = x.cuda(non_blocking=True)
y = y.cuda(non_blocking=True)
out = model(x)
mainLoss = criterion(out[0], y)
auxLoss = criterion(out[1], y)
# whole loss
loss = 0.4 * auxLoss + mainLoss
lossMeter.update(loss.item(), x.shape[0])
#print('loss is:', loss.item())
# step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# ajust lr
curIter = currentepoch * len(dataLoader) + i + 1
newLr = poly_learning_rate(baseLr, curIter, maxIter)
optimizer.param_groups[0]['lr'] = newLr
# compute IoU/accuracy
result = out[0].max(1)[1]
intersection, union, target = common.intersectionAndUnionGPU(
result, y, numClasses, 255)
intersection, union, target = intersection.cpu().numpy(), union.cpu(
).numpy(), target.cpu().numpy()
## update meter
intersectionMeter.update(intersection), unionMeter.update(
union), targetMeter.update(target)
# after every epoch, print the log
IoU = intersectionMeter.sum / (unionMeter.sum + 1e-10)
accuracy = intersectionMeter.sum / (targetMeter.sum + 1e-10)
print(f'[{currentepoch}/{globalEpoch}] loss:', lossMeter.avg)
'''
def sub_sn_train(model, optimizer, criterion, dataloader, currentepoch,
maxIter):
model.train()
intersectionmeter = common.AverageMeter()
unionmeter = common.AverageMeter()
targetmeter = common.AverageMeter()
lossmeter = common.AverageMeter()
for i, (x, y) in enumerate(dataloader):
x = x.to(DEVICE, non_blocking=True)
y = y.to(DEVICE, non_blocking=True)
out = model(x)
out[1] = nn.Upsample(scale_factor=8, mode='bilinear')(out[1])
loss = criterion(out[1], y)
lossmeter.update(loss.item(), x.shape[0])
optimizer.zero_grad()
loss.backward()
optimizer.step()
curiter = currentepoch * len(dataloader) + i + 1
newlr = poly_learning_rate(BASELR, curiter, maxIter)
optimizer.param_groups[0]['lr'] = newlr
#iou
result = out[1].max(1)[1]
intersection, union, target = common.intersectionAndUnionGPU(
result, torch.squeeze(y), NUM_CLASSES, 255)
intersection, union, target = intersection.cpu().numpy(), union.cpu(
).numpy(), target.cpu().numpy()
intersectionmeter.update(intersection), unionmeter.update(
union), targetmeter.update(target)
IoU = intersectionmeter.sum / (unionmeter.sum + 1e-10)
accuracy = intersectionmeter.sum / (targetmeter.sum + 1e-10)
print(f'[{currentepoch}/{GLOBALEPOCH}] loss:{lossmeter.avg}')
def sub_sn_val(model, criterion, dataloader):
model.eval()
intersectionmeter = common.AverageMeter()
unionmeter = common.AverageMeter()
targetmeter = common.AverageMeter()
lossmeter = common.AverageMeter()
for i, (x, y) in enumerate(dataloader):
x = x.to(DEVICE, non_blocking=True)
y = y.to(DEVICE, non_blocking=True)
out = model(x)
out[1] = nn.Upsample(scale_factor=8, mode='bilinear')(out[1])
mainloss = criterion(out[1], y)
lossmeter.update(mainloss.item(), x.shape[0])
result = out[1].max(1)[1]
intersection, union, target = common.intersectionAndUnionGPU(
result, torch.squeeze(y), NUM_CLASSES, 255)
intersection, union, target = intersection.cpu().numpy(), union.cpu(
).numpy(), target.cpu().numpy()
intersectionmeter.update(intersection), unionmeter.update(
union), targetmeter.update(target)
# save for debug
rt = result * 128
rt = rt.to(torch.uint8)
rt = rt.cpu().numpy()
cv2.imwrite('segout.jpg', rt[0])
#IoU
IoU = intersectionmeter.sum / (unionmeter.sum + 1e-10)
accuracy = intersectionmeter.sum / (targetmeter.sum + 1e-10)
print(f'val loss:{lossmeter.avg}')
for i in range(NUM_CLASSES):
print(f'class_{i} IoU:{IoU[i]}, acc: {accuracy[i]}')