def evaluate(model, criterion, eval_dataloader, is_aux=False):
model.eval()
eval_loss = 0
eval_acc = 0
eval_acc_cls = 0
eval_mean_iu = 0
eval_fwavacc = 0
for (val_img, val_label) in eval_dataloader:
val_img = val_img.cuda()
val_label = val_label.cuda()
# forward
if is_aux is True:
aux_out1, aux_out2, main_out = model(val_img)
else:
main_out = model(val_img)
val_loss = criterion(main_out, val_label)
eval_loss += val_loss.item()
label_pred = main_out.max(dim=1)[1].data.cpu().numpy()
label_true = val_label.data.cpu().numpy()
for lbt, lbp in zip(label_true, label_pred):
acc, acc_cls, mean_iu, fwavacc = label_accuracy_score(
lbt, lbp, config.num_class)
eval_acc += acc
eval_acc_cls += acc_cls
eval_mean_iu += mean_iu
eval_fwavacc += fwavacc
return eval_loss, eval_acc, eval_mean_iu
def train_epoch(self):
self.model.train()
for batch_idx, (data, target) in enumerate(self.train_loader):
score, loss, lbl_pred, lbl_true = self.forward(data, target)
self.optim.zero_grad()
loss.backward()
self.optim.step()
print("Seenmask Train Epoch {:<5} | Iteration {:<5} | Loss {:5.5f} | seenmask_score grad sum {:7.8f} | seenmask_upscore grad sum {:7.8f} | score sum {:10.5f}".format(
int(self.epoch), int(batch_idx), float(loss.data[0]), float(self.model.seenmask_score.weight.grad.sum().data[0]),
float(self.model.seenmask_upscore.weight.grad.sum().data[0]), float(score.sum().data[0])))
metrics = utils.label_accuracy_score(lbl_true.numpy(), lbl_pred, self.n_class)
with open(osp.join(self.log_dir, 'seenmask_train_log.csv'), 'a') as f:
elapsed_time = (datetime.datetime.now(pytz.timezone('US/Eastern')) - self.timestamp_start).total_seconds()
log = [self.epoch, self.iteration] + [loss.data[0]] + list(metrics) + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
# write to tensorboard
self.tb_writer.add_scalar('seenmask/train/loss', loss.data[0], self.iteration)
self.tb_writer.add_scalar('seenmask/train/pxl_acc', metrics[0], self.iteration)
self.tb_writer.add_scalar('seenmask/train/class_acc', metrics[1], self.iteration)
self.tb_writer.add_scalar('seenmask/train/mean_iu', metrics[2], self.iteration)
self.tb_writer.add_scalar('seenmask/train/fwavacc', metrics[3], self.iteration)
self.iteration += 1
def validation(epoch, model, data_loader, criterion, device):
print('Start validation fold{} #{}'.format(fold + 1, epoch))
model.eval()
with torch.no_grad():
total_loss = 0
cnt = 0
mIoU_list = []
for step, (images, masks, _) in enumerate(data_loader):
images = torch.stack(images) # (batch, channel, height, width)
masks = torch.stack(
masks).long() # (batch, channel, height, width)
images, masks = images.to(device), masks.to(device)
outputs = model(images)
loss = criterion(outputs, masks)
total_loss += loss
cnt += 1
outputs = torch.argmax(outputs.squeeze(),
dim=1).detach().cpu().numpy()
mIoUs = label_accuracy_score(masks.detach().cpu().numpy(),
outputs,
n_class=12)
mIoU_list += mIoUs
avrg_loss = total_loss / cnt
print(
'Validation {} #{} Average Loss: {:.4f}, mIoU: {:.4f}'.format(
fold + 1, epoch, avrg_loss, np.mean(mIoU_list)))
return np.mean(mIoU_list)
def _get_loss_gen(self):
batchsize = self.y_fake.data.shape[0]
L_mce = F.softmax_cross_entropy(self.pred_label_map,
self.ground_truth,
normalize=False)
L_bce = F.softmax_cross_entropy(
self.y_fake,
Variable(self.xp.ones(batchsize, dtype=self.xp.int32),
volatile=not self.gen.train))
loss = L_mce + self.L_bce_weight * L_bce
# log report
label_true = chainer.cuda.to_cpu(self.ground_truth.data)
label_pred = chainer.cuda.to_cpu(
self.pred_label_map.data).argmax(axis=1)
logs = []
for i in six.moves.range(batchsize):
acc, acc_cls, iu, fwavacc = utils.label_accuracy_score(
label_true[i], label_pred[i], self.n_class)
logs.append((acc, acc_cls, iu, fwavacc))
log = np.array(logs).mean(axis=0)
values = {
'loss': loss,
'accuracy': log[0],
'accuracy_cls': log[1],
'iu': log[2],
'fwavacc': log[3],
}
chainer.report(values, self.gen)
return loss
def val(net,val_loader,viz,criterion,iteration = None):
'''
validation our model is well?
'''
training = net.training
#eval mode for drop
device = torch.device("cuda:0")
net.eval()
val_true_win = 'val_true_image'
label_name = val_loader.dataset.class_names
val_loss =0.0
label_trues = []
label_preds = []
for batch_idx,(data,target) in tqdm.tqdm(enumerate(val_loader,1),total = len(val_loader),desc='Validation iteration {}'.format(iteration),ncols=80,leave = False):
data = data.to(device)
target = target.to(device)
scores = net(data)
with torch.set_grad_enabled(False):
loss =criterion(scores,target)
val_loss +=loss.item()
imgs =data.detach().cpu().numpy()
#get the idx value for which labels?
lbl_pred = scores.max(1)[1].cpu().numpy()[:,:,:]
lbl_true = target.detach().cpu().numpy()
label_trues.append(lbl_true)
label_preds.append(lbl_pred)
acc, acc_cls, mean_iu, fwavacc = utils.label_accuracy_score(label_trues,label_preds,21)
loss = val_loss/len(val_loader)
net.train()
return loss,acc,acc_cls,mean_iu,fwavacc
def val(model, dataloader):
model.eval()
eval_acc = 0
eval_acc_cls = 0
eval_mean_iu = 0
eval_fwavacc = 0
li_pred = []
li_gt = []
for i, (_, im, cloud, theta, shift, lb) in enumerate(dataloader):
im, cloud, theta, shift, lb = Variable(im), Variable(cloud), Variable(theta), Variable(shift), Variable(lb)
im, cloud, theta, shift, lb = im.float().cuda(), cloud.float().cuda(), theta.float().cuda(), shift.float().cuda(), lb.long().cuda()
_, pred = model(im, cloud, theta, shift)
# Mean IoU
label_true = lb.data.cpu().numpy().astype(np.int8)
label_pred = pred.data.cpu().numpy().squeeze(0)
label_pred = (label_pred > 0.5).astype(np.int8)
for (label, prob) in zip(label_true, label_pred):
acc, acc_cls, mean_iu, fwavacc = label_accuracy_score(label, prob, n_class=2)
eval_acc += acc
eval_acc_cls += acc_cls
eval_mean_iu += mean_iu
eval_fwavacc += fwavacc
# MaxF
label_pred = pred.data.cpu().numpy().squeeze()
label_true = lb.data.cpu().numpy().squeeze()
li_pred.append(label_pred)
li_gt.append(label_true)
print 'Validation ======ACC: %lf,Mean IoU: %lf======' % (eval_acc/dataloader.__len__(),
eval_mean_iu/dataloader.__len__())
max_f = eval_road(li_pred, li_gt)
model.train()
return max_f
def train_epoch(self):
self.model.train()
if self.epoch % 20 == 0 and self.epoch != 0:
for param_group in self.optim.param_groups:
param_group['lr'] = self.initlr * 0.1
self.initlr = self.initlr * 0.1
print('current learning rate is ')
print(self.initlr)
n_class = len(self.train_loader.dataset.class_names)
for batch_idx, (data, target) in tqdm.tqdm(
enumerate(self.train_loader), total=len(self.train_loader),
desc='Train epoch=%d' % self.epoch, ncols=80, leave=False):
iteration = batch_idx + self.epoch * len(self.train_loader)
#print('shape of validate data is ',data.shape)
if self.iteration != 0 and (iteration - 1) != self.iteration:
continue # for resuming
self.iteration = iteration
if self.iteration % self.interval_validate == 0:
self.validate()
if self.cuda:
data, target = data.cuda(), target.cuda()
data, target = Variable(data), Variable(target)
self.optim.zero_grad()
score = self.model(data)
loss = cross_entropy2d(score, target,
size_average=self.size_average)
loss.backward()
self.optim.step()
metrics = []
lbl_pred = score.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu().numpy()
for lt, lp in zip(lbl_true, lbl_pred):
acc, acc_cls, mean_iu, fwavacc = \
utils.label_accuracy_score(
[lt], [lp], n_class=n_class)
metrics.append((acc, acc_cls, mean_iu, fwavacc))
metrics = np.mean(metrics, axis=0)
with open(osp.join(self.out, 'log.csv'), 'a') as f:
elapsed_time = (
datetime.datetime.now(pytz.timezone('Asia/Shanghai')) -
self.timestamp_start).total_seconds()
log = [self.epoch, self.iteration] + [loss.data[0]] + \
metrics.tolist() + [''] * 5 + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
if self.iteration >= self.max_iter:
break
def eval_on_validation():
if not os.path.exists(config.run_dir):
os.mkdir(config.run_dir)
model = getattr(models, config.model)()
model = torch.nn.DataParallel(model) # multi-gpu
model.cuda()
print 'test on validation set.', config.model
print model
if config.load_model_path:
model.load_state_dict(torch.load(config.load_model_path))
# data
test_data = KITTIRoadFusion(config.root, split='val', num_features=19)
test_dataloader = DataLoader(test_data, batch_size=1, shuffle=False, num_workers=4)
# test
model.eval()
eval_acc = 0
eval_acc_cls = 0
eval_mean_iu = 0
eval_fwavacc = 0
li_pred = []
li_gt = []
total_time = 0
for i, (name, im, cloud, theta, shift, lb) in enumerate(test_dataloader):
im, cloud, theta, shift, lb = Variable(im), Variable(cloud), Variable(theta), Variable(shift), Variable(lb)
im, cloud, theta, shift, lb = im.float().cuda(), cloud.float().cuda(), theta.float().cuda(), shift.float().cuda(), lb.long().cuda()
start = time.clock()
_, pred = model(im, cloud, theta, shift) # inference
end = time.clock()
total_time += (end-start)
# pred = F.upsample_bilinear(pred, scale_factor=4)
# save image
label_pred = pred.data.cpu().numpy().squeeze()
label_pred = np.array(label_pred*255, dtype=np.uint8)
filename = os.path.join(config.run_dir, name[0])
print filename
# cv2.imwrite(filename, label_pred)
# Mean IoU
label_true = lb.data.cpu().numpy().astype(np.int8)
label_pred = pred.data.cpu().numpy().squeeze(0)
label_pred = (label_pred > 0.5).astype(np.int8)
for (label, prob) in zip(label_true, label_pred):
acc, acc_cls, mean_iu, fwavacc = label_accuracy_score(label, prob, n_class=2)
eval_acc += acc
eval_acc_cls += acc_cls
eval_mean_iu += mean_iu
eval_fwavacc += fwavacc
# MaxF
label_pred = pred.data.cpu().numpy().squeeze()
label_true = lb.data.cpu().numpy().squeeze()
li_pred.append(label_pred)
li_gt.append(label_true)
print 'Runtime ############# time(s) : %f ##########' % (total_time / test_dataloader.__len__())
print 'Validation ======ACC: %lf,Mean IoU: %lf======' % (eval_acc / test_dataloader.__len__(),
eval_mean_iu / test_dataloader.__len__())
eval_road(li_pred, li_gt)
def train(num_epochs, model, data_loader, val_loader, val_every, device, file_name):
learning_rate = 0.0001
from torch.optim.swa_utils import AveragedModel, SWALR
from torch.optim.lr_scheduler import CosineAnnealingLR
from segmentation_models_pytorch.losses import SoftCrossEntropyLoss, JaccardLoss
from adamp import AdamP
criterion = [SoftCrossEntropyLoss(smooth_factor=0.1), JaccardLoss('multiclass', classes=12)]
optimizer = AdamP(params=model.parameters(), lr=learning_rate, weight_decay=1e-6)
swa_scheduler = SWALR(optimizer, swa_lr=learning_rate)
swa_model = AveragedModel(model)
look = Lookahead(optimizer, la_alpha=0.5)
print('Start training..')
best_miou = 0
for epoch in range(num_epochs):
hist = np.zeros((12, 12))
model.train()
for step, (images, masks, _) in enumerate(data_loader):
loss = 0
images = torch.stack(images) # (batch, channel, height, width)
masks = torch.stack(masks).long() # (batch, channel, height, width)
# gpu 연산을 위해 device 할당
images, masks = images.to(device), masks.to(device)
# inference
outputs = model(images)
for i in criterion:
loss += i(outputs, masks)
# loss 계산 (cross entropy loss)
look.zero_grad()
loss.backward()
look.step()
outputs = torch.argmax(outputs.squeeze(), dim=1).detach().cpu().numpy()
hist = add_hist(hist, masks.detach().cpu().numpy(), outputs, n_class=12)
acc, acc_cls, mIoU, fwavacc = label_accuracy_score(hist)
# step 주기에 따른 loss, mIoU 출력
if (step + 1) % 25 == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, mIoU: {:.4f}'.format(
epoch + 1, num_epochs, step + 1, len(data_loader), loss.item(), mIoU))
# validation 주기에 따른 loss 출력 및 best model 저장
if (epoch + 1) % val_every == 0:
avrg_loss, val_miou = validation(epoch + 1, model, val_loader, criterion, device)
if val_miou > best_miou:
print('Best performance at epoch: {}'.format(epoch + 1))
print('Save model in', saved_dir)
best_miou = val_miou
save_model(model, file_name = file_name)
if epoch > 3:
swa_model.update_parameters(model)
swa_scheduler.step()
def func_eval(model, criterion, val_dataset, val_loader, post_crf=False):
print("Start validation.\n")
model.eval() # make model evaluation mode
with torch.no_grad():
n_class = 12
total_loss_sum = 0
mIoU_list = []
hist = np.zeros((n_class, n_class)) # confusion matrix
for step, (images, masks, img_info) in enumerate(val_loader):
images = torch.stack(images).to(
CFG.device) # (batch, channel, height, width)
masks = torch.stack(masks).long().to(
CFG.device) # (batch, channel, height, width)
# forward pass (get logits)
logits = model(images)
# loss 계산 (cross entropy loss)
loss = criterion(logits, masks)
total_loss_sum += loss.item() * images.shape[0]
# use softmax to get probability
probs = F.softmax(logits, dim=1)
probs = probs.data.cpu().numpy()
# Postprocessing
if post_crf:
pool = mp.Pool(mp.cpu_count())
images = images.data.cpu().numpy().astype(np.uint8).transpose(
0, 2, 3, 1)
probs = pool.map(dense_crf_wrapper, zip(images, probs))
pool.close()
# get class index which has biggest probability
preds = np.argmax(probs, axis=1)
masks = masks.detach().cpu().numpy()
hist = add_hist(hist, masks, preds, n_class=n_class)
if step == 0:
fig_mask = log_images(masks, preds, img_info)
del images, masks, logits, probs, preds
val_loss = total_loss_sum / len(val_dataset)
acc, acc_cls, mIoU, iu, fwavacc = label_accuracy_score(hist)
recycle = [
'Background', 'UNKNOWN', 'General trash', 'Paper', 'Paper pack',
'Metal', 'Glass', 'Plastic', 'Styrofoam', 'Plastic bag', 'Battery',
'Clothing'
]
mIoU_df = pd.DataFrame({'Recycle Type': recycle, 'IoU': iu})
return val_loss, acc, mIoU, mIoU_df, fig_mask
def validate(self):
training = self.model.training
self.model.eval()
n_class = len(self.val_loader.dataset.class_names)
val_loss = 0
visualizations = []
label_trues, label_preds = [], []
with torch.no_grad():
for batch_idx, (data, target) in tqdm.tqdm(
enumerate(self.val_loader),
total=len(self.val_loader),
desc='Valid iteration=%d' % self.iteration,
ncols=80,
leave=False):
if self.cuda:
data, target = data.to('cuda'), target.to('cuda')
score = self.model(data)
loss = cross_entropy2d(score,
target,
size_average=self.size_average)
if np.isnan(float(loss.item())):
raise ValueError('loss is nan while validating')
val_loss += float(loss.item()) / len(data)
imgs = data.data.cpu()
lbl_pred = score.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu()
for img, lt, lp in zip(imgs, lbl_true, lbl_pred):
img, lt = self.val_loader.dataset.untransform(img, lt)
label_trues.append(lt)
label_preds.append(lp)
if len(visualizations) < 9:
viz = fcn.utils.visualize_segmentation(lbl_pred=lp,
lbl_true=lt,
img=img,
n_class=n_class)
visualizations.append(viz)
metrics = utils.label_accuracy_score(label_trues, label_preds, n_class)
print(label_trues)
print(label_preds)
val_loss /= len(self.val_loader)
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
print('acc, acc_cls, mean_iu, fwavacc')
print('acc:', metrics[0])
print('acc_cls:', metrics[1])
print('mean_iu:', metrics[2])
print('fwavacc:', metrics[3])
return label_trues, label_preds, n_class
def train_epoch(self):
self.model.train()
n_class = len(self.train_loader.dataset.class_names)
for batch_idx, (data, target) in tqdm.tqdm(
enumerate(self.train_loader),
total=len(self.train_loader),
desc='Train epoch=%d' % self.epoch,
ncols=80,
leave=False):
iteration = batch_idx + self.epoch * len(self.train_loader)
if self.iteration != 0 and (iteration - 1) != self.iteration:
continue # for resuming
self.iteration = iteration
if self.iteration % self.interval_validate == 0:
self.validate()
assert self.model.training
if self.cuda:
data, target = data.to('cuda'), target.to('cuda')
self.optim.zero_grad()
score = self.model(data)
loss = cross_entropy2d(score,
target,
size_average=self.size_average)
loss /= len(data)
if np.isnan(float(loss.item())):
raise ValueError('loss is nan while training')
loss.backward()
self.optim.step()
metrics = []
lbl_pred = score.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu().numpy()
acc, acc_cls, mean_iu, fwavacc = \
utils.label_accuracy_score(
lbl_true, lbl_pred, n_class=n_class)
metrics.append((acc, acc_cls, mean_iu, fwavacc))
metrics = np.mean(metrics, axis=0)
with open(osp.join(self.out, 'log.csv'), 'a') as f:
elapsed_time = (
datetime.datetime.now(pytz.timezone('America/Bogota')) -
self.timestamp_start).total_seconds()
log = [self.epoch, self.iteration] + [loss.item()] + \
metrics.tolist() + [''] * 5 + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
if self.iteration >= self.max_iter:
break
def eval_on_validation_bev():
if not os.path.exists(config.run_dir):
os.mkdir(config.run_dir)
model = getattr(models, config.model)()
model = torch.nn.DataParallel(model) # multi-gpu
model.cuda()
print 'test on validation set.', config.model
print model
if config.load_model_path:
model.load_state_dict(torch.load(config.load_model_path))
# data
bev = BirdsEyeView()
test_data = KITTIRoadFusion(config.root, split='val', num_features=19, return_bev=True)
test_dataloader = DataLoader(test_data, batch_size=1, shuffle=False, num_workers=4)
# test
model.eval()
eval_acc = 0
eval_acc_cls = 0
eval_mean_iu = 0
eval_fwavacc = 0
li_pred = []
li_gt = []
for i, (name, im, cloud, theta, shift, _, lb) in enumerate(test_dataloader):
im, cloud, theta, shift, lb = Variable(im), Variable(cloud), Variable(theta), Variable(shift), Variable(lb)
im, cloud, theta, shift, lb = im.float().cuda(), cloud.float().cuda(), theta.float().cuda(), shift.float().cuda(), lb.long().cuda()
_, pred = model(im, cloud, theta, shift) # inference
pred = pred.data.cpu().numpy().squeeze()
theta = theta.data.cpu().numpy().squeeze()
shift = shift.data.cpu().numpy().squeeze()
label_pred = bev.transformLable2BEV((pred*255).astype(np.uint8), theta, shift)
label_true = lb.data.cpu().numpy().squeeze()
# save image
filename = os.path.join(config.run_dir, name[0])
print filename
cv2.imwrite(filename, label_pred)
label_pred = label_pred/255.
# Mean IoU
label_true_hard = np.expand_dims(label_true.astype(np.int8), axis=0)
label_pred_hard = np.expand_dims((label_pred > 0.5).astype(np.int8), axis=0)
for (label, prob) in zip(label_true_hard, label_pred_hard):
acc, acc_cls, mean_iu, fwavacc = label_accuracy_score(label, prob, n_class=2)
eval_acc += acc
eval_acc_cls += acc_cls
eval_mean_iu += mean_iu
eval_fwavacc += fwavacc
# MaxF
li_pred.append(label_pred)
li_gt.append(label_true)
print 'Validation ======ACC: %lf,Mean IoU: %lf======' % (eval_acc / test_dataloader.__len__(),
eval_mean_iu / test_dataloader.__len__())
eval_road(li_pred, li_gt)
def train(num_epochs, model, data_loader, val_loader, criterion, optimizer,
saved_dir, val_every, device, file_name, n_class):
print('Start training..')
best_mIoU = 0
for epoch in range(num_epochs):
hist = np.zeros((n_class, n_class))
model.train()
for step, (images, masks, _) in enumerate(data_loader):
# (batch, channel, height, width)
images = torch.stack(images)
# (batch, channel, height, width)
masks = torch.stack(masks).long()
# gpu 연산을 위해 device 할당
images, masks = images.to(device), masks.to(device)
# inference
outputs = model(images)
# loss 계산 (cross entropy loss)
loss = criterion(outputs, masks)
optimizer.zero_grad()
loss.backward()
optimizer.step()
outputs = torch.argmax(outputs.squeeze(),
dim=1).detach().cpu().numpy()
hist = add_hist(hist,
masks.detach().cpu().numpy(),
outputs,
n_class=n_class)
acc, acc_cls, mIoU, fwavacc = label_accuracy_score(hist)
wandb.log({"loss": loss, "mIoU": mIoU}) # wandb 로그출력
# step 주기에 따른 loss 출력
if (step + 1) % 25 == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, mIoU:{:.4f}'.
format(epoch + 1, num_epochs, step + 1,
len(train_loader), loss.item(), mIoU))
# validation 주기에 따른 loss 출력 및 best model 저장
# mIoU에 따라 모델 저장
if (epoch + 1) % val_every == 0:
avrg_loss, val_mIoU = validation(epoch + 1, model, val_loader,
criterion, device, n_class)
if val_mIoU > best_mIoU:
print('Best performance at epoch: {}'.format(epoch + 1))
print('Save model in', saved_dir)
best_mIoU = val_mIoU
save_model(model, saved_dir, file_name)
wandb.log({
"val_loss": avrg_loss,
"val_mIoU": val_mIoU,
"best_mIoU": best_mIoU
})
def eval_metric(self, score, target, n_class):
# -----------------------------------------------------------------------------
metrics = []
lbl_pred = score.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.cpu().numpy()
for lt, lp in zip(lbl_true, lbl_pred):
acc, acc_cls, mean_iu, fwavacc = \
utils.label_accuracy_score(
[lt], [lp], n_class=n_class)
metrics.append((acc, acc_cls, mean_iu, fwavacc))
metrics = np.mean(metrics, axis=0)
return metrics
def train(epoch_idx, net, train_loader, lr, logger, n_class):
net.cuda()
net.train()
base_params = list(map(id, net.base_net.parameters()))
top_params = filter(lambda p: id(p) not in base_params, net.parameters())
optimizer = torch.optim.SGD([{
'params': top_params
}, {
'params': net.base_net.parameters(),
'lr': lr * 0.1
}],
lr=lr,
momentum=0.9,
weight_decay=0.00004)
criterion = nn.CrossEntropyLoss(ignore_index=-1)
len_batch = len(train_loader)
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
score = net(data)
loss = criterion(score, target)
loss.backward()
optimizer.step()
_, predicted = score.max(1)
predicted, target = to_np(predicted), to_np(target)
acc, acc_cls, mean_iu = label_accuracy_score(target, predicted,
n_class)
info = {
'acc': acc,
'acc_cls': acc_cls,
'mean_iu': mean_iu,
'loss': loss.data[0]
}
for tag, value in info.items():
logger.scalar_summary(tag, value,
len_batch * epoch_idx + batch_idx + 1)
print(('train', batch_idx, epoch_idx))
if (epoch_idx + 1) % 10 == 0:
n = (epoch_idx + 1) / 10
state = net.state_dict()
torch.save(state, './deeplab_epoch_' + str(n) + '.pth')
def main():
model = Deeplab()
dataset = VOC2012ClassSeg('./dataset', split='train', transform=True)
val_loader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=1,
pin_memory=True)
# n_class = len(dataset.class_names)
# model_file = ''
# moda_data = torch.load(model_file)
# try:
# model.load_state_dict(model_data)
# except Exception:
# model.load_state_dict(model_data['model_state_dict'])
# if torch.cuda.is_available():
# model.cuda()
model.eval()
label_trues, label_preds = [], []
for batch_idx, (data, target) in enumerate(val_loader):
# if torch.cuda.is_available():
# data, target = data.cuda(), target.cuda()
data, target = Variable(data, volatile=True), Variable(target)
score = model(data)
_, predicted = score.max(1)
predicted = to_np(predicted)
target = to_np(target)
for lt, lp in zip(target, predicted):
label_trues.append(lt)
label_preds.append(lp)
if batch_idx == 5:
break
n_class = 21
print(len(label_preds))
metrics = label_accuracy_score(label_trues, label_preds, n_class=n_class)
metrics = np.array(metrics)
metrics *= 100
print(metrics)
def validate(self):
self.netd.eval()
self.netg.eval()
self.nets.eval()
progressbar = tqdm(self.valid_data_loader)
for ii, (imgs, _) in enumerate(progressbar):
normal, defect, target = imgs
if self.opt.use_gpu:
normal = normal.cuda()
defect = defect.cuda()
target = target.cuda()
repair = self.netg(defect)
if self.opt.with_segmentation:
seg_input = torch.cat([defect, repair], dim=1)
seg = self.nets(seg_input)
else:
seg = None
if self.opt.with_segmentation:
metrics = []
lbl_pred = seg.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu().numpy()
acc, acc_cls, mean_iu, fwavacc = \
label_accuracy_score(
lbl_true, lbl_pred, n_class=2)
metrics.append((acc, acc_cls, mean_iu, fwavacc))
metrics = np.mean(metrics, axis=0)
progressbar.set_description(
f'Acc: {metrics[0]:.5f}, Acc_cls: {metrics[1]:.5f}, MIU: {metrics[2]:.5f}, Fwavacc: {metrics[3]:.5f}'
)
if self.opt.debug:
if not os.path.exists(self.opt.val_save_path):
os.makedirs(self.opt.val_save_path)
imgs = torch.cat((defect, repair), 0)
tv.utils.save_image(imgs,
os.path.join(
self.opt.val_save_path,
'{}_defect_repair.jpg'.format(ii)),
normalize=True,
range=(-1, 1))
def validation(epoch, model, data_loader, criterion, device, n_class):
print('Start validation #{}'.format(epoch))
model.eval()
with torch.no_grad():
total_loss = 0
cnt = 0
mIoU_list = []
hist = np.zeros((n_class, n_class)) # 중첩을위한 변수
for step, (images, masks, _) in enumerate(data_loader):
# (batch, channel, height, width)
images = torch.stack(images)
# (batch, channel, height, width)
masks = torch.stack(masks).long()
images, masks = images.to(device), masks.to(device)
outputs = model(images)
loss = criterion(outputs, masks)
total_loss += loss
cnt += 1
outputs = torch.argmax(outputs.squeeze(),
dim=1).detach().cpu().numpy()
# 계산을 위한 중첩
hist = add_hist(hist,
masks.detach().cpu().numpy(),
outputs,
n_class=n_class)
# mIoU = label_accuracy_score(
# masks.detach().cpu().numpy(), outputs, n_class=12)[2]
# mIoU_list.append(mIoU)
# mIoU가 전체에대해 계산
acc, acc_cls, mIoU, fwavacc = label_accuracy_score(hist)
avrg_loss = total_loss / cnt
print('Validation #{} Average Loss: {:.4f}, mIoU: {:.4f}'.format(
epoch, avrg_loss, mIoU))
return avrg_loss, mIoU
def calc_loss(self):
batchsize = self.ground_truth.shape[0]
self.loss = F.softmax_cross_entropy(self.pred_label_map,
self.ground_truth,
normalize=False)
# log report
label_true = chainer.cuda.to_cpu(self.ground_truth.data)
label_pred = chainer.cuda.to_cpu(
self.pred_label_map.data).argmax(axis=1)
logs = []
for i in six.moves.range(batchsize):
acc, acc_cls, iu, fwavacc = utils.label_accuracy_score(
label_true[i], label_pred[i], self.n_class)
logs.append((acc, acc_cls, iu, fwavacc))
log = np.array(logs).mean(axis=0)
values = {
'loss': self.loss,
'accuracy': log[0],
'accuracy_cls': log[1],
'iu': log[2],
'fwavacc': log[3],
}
chainer.report(values, self.model)
def psudo_labeling(num_epochs, model, data_loader, val_loader,
unlabeled_loader, criterion, optimizer, device, n_class,
saved_dir, file_name, val_every):
# Instead of using current epoch we use a "step" variable to calculate alpha_weight
# This helps the model converge faster
step = 100
size = 256
transform = A.Compose([A.Resize(256, 256)])
preds_array = np.empty((0, size * size), dtype=np.long)
file_name_list = []
best_mIoU = 0
model.train()
for epoch in range(num_epochs):
hist = np.zeros((n_class, n_class))
for batch_idx, (imgs, image_infos) in enumerate(unlabeled_loader):
# Forward Pass to get the pseudo labels
# --------------------------------------------- test(unlabelse)를 모델에 통과
model.eval()
outs = model(torch.stack(imgs).to(device))
oms = torch.argmax(outs.squeeze(), dim=1).detach().cpu().numpy()
oms = torch.Tensor(oms)
oms = oms.long()
oms = oms.to(device)
# --------------------------------------------- 학습
model.train()
# Now calculate the unlabeled loss using the pseudo label
imgs = torch.stack(imgs)
imgs = imgs.to(device)
# preds_array = preds_array.to(device)
output = model(imgs)
unlabeled_loss = alpha_weight(step) * criterion(output, oms)
# Backpropogate
optimizer.zero_grad()
unlabeled_loss.backward()
optimizer.step()
output = torch.argmax(output.squeeze(),
dim=1).detach().cpu().numpy()
hist = add_hist(hist,
oms.detach().cpu().numpy(),
output,
n_class=n_class)
if (batch_idx + 1) % 25 == 0:
acc, acc_cls, mIoU, fwavacc = label_accuracy_score(hist)
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, mIoU:{:.4f}'.
format(epoch + 1, num_epochs, batch_idx + 1,
len(unlabeled_loader), unlabeled_loss.item(),
mIoU))
# For every 50 batches train one epoch on labeled data
# 50배치마다 라벨데이터를 1 epoch학습
if batch_idx % 50 == 0:
# Normal training procedure
for batch_idx, (images, masks, _) in enumerate(train_loader):
images = torch.stack(images)
# (batch, channel, height, width)
masks = torch.stack(masks).long()
# gpu 연산을 위해 device 할당
images, masks = images.to(device), masks.to(device)
output = model(images)
labeled_loss = criterion(output, masks)
optimizer.zero_grad()
labeled_loss.backward()
optimizer.step()
# Now we increment step by 1
step += 1
if (epoch + 1) % val_every == 0:
avrg_loss, val_mIoU = validation(epoch + 1, model, val_loader,
criterion, device, n_class)
if val_mIoU > best_mIoU:
print('Best performance at epoch: {}'.format(epoch + 1))
print('Save model in', saved_dir)
best_mIoU = val_mIoU
save_model(model, saved_dir, file_name)
wandb.log({
"val_loss": avrg_loss,
"val_mIoU": val_mIoU,
"best_mIoU": best_mIoU
})
model.train()
def validate(self):
n_class = self.train_loader.dataset.n_class
# os.system('play -nq -t alsa synth {} sine {}'.format(0.3, 440)) # sound an alarm
val_loss = 0
prec = 0
metrics = np.zeros((len(self.val_loader), 4), dtype=np.float64)
for batch_idx, (rgb_img, ddd_img,
target) in tqdm.tqdm(enumerate(self.val_loader),
total=len(self.val_loader),
desc=' val %d' % self.epoch,
ncols=80,
leave=False):
## validate
with torch.no_grad():
self.model.eval()
if self.cuda:
rgb_img = rgb_img.cuda()
ddd_img = ddd_img.cuda()
target = target.cuda()
output = self.model(rgb_img, ddd_img)
if self.val_loader.dataset.encode_label:
output = F.interpolate(output,
size=target.size()[2:],
mode='bilinear',
align_corners=False)
else:
output = F.interpolate(output,
size=target.size()[1:],
mode='bilinear',
align_corners=False)
loss = self.criterion(output, target)
loss_data = loss.data.item()
if np.isnan(loss_data):
raise ValueError('loss is nan while validating')
val_loss += loss_data / len(rgb_img)
## some stats
lbl_pred = output.data.max(1)[1].cpu().numpy().squeeze()
lbl_true = target.data.cpu().numpy().squeeze()
prec += compute_precision(lbl_pred, lbl_true)
m = label_accuracy_score(lbl_true, lbl_pred, n_class)
metrics[batch_idx, :] = np.array(m)
metrics = np.mean(metrics, axis=0)
val_prec = prec / len(self.val_loader)
with open(osp.join(self.output_path, 'log.csv'), 'a') as f:
metrics_str = ['%.10f' % (a) for a in list(metrics)]
elapsed_time = (
datetime.datetime.now(pytz.timezone('Asia/Jakarta')) -
self.timestamp_start).total_seconds()
val_loss /= len(self.val_loader)
log = [self.epoch, self.iteration] + [''] * 5 + \
['%.10f' %(val_loss)] + metrics_str + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
if is_best:
self.best_mean_iu = mean_iu
is_prec_best = val_prec > self.best_prec
if is_prec_best:
self.best_prec = val_prec
torch.save(
{
'epoch': self.epoch,
'iteration': self.iteration,
'arch': self.arch,
'optim_state_dict': self.optim.state_dict(),
'model_state_dict': self.model.state_dict(),
'best_mean_iu': self.best_mean_iu,
'best_prec': self.best_prec,
}, osp.join(self.output_path, 'checkpoint.pth.tar'))
if self.arch == 'rfnet':
torch.save(
{
'epoch': self.epoch,
'iteration': self.iteration,
'arch': self.arch,
'optim_state_dict': self.optim.state_dict(),
'optim_dec_state_dict': self.optim_dec.state_dict(),
'model_state_dict': self.model.state_dict(),
'best_mean_iu': self.best_mean_iu,
'best_prec': self.best_prec,
}, osp.join(self.output_path, 'checkpoint.pth.tar'))
if is_best:
shutil.copy(osp.join(self.output_path, 'checkpoint.pth.tar'),
osp.join(self.output_path, 'model_best.pth.tar'))
if is_prec_best:
shutil.copy(osp.join(self.output_path, 'checkpoint.pth.tar'),
osp.join(self.output_path, 'model_prec_best.pth.tar'))
self.writer.add_scalar('val/loss', val_loss, self.epoch)
self.writer.add_scalar('val/precision', val_prec, self.epoch)
self.writer.add_scalar('val/accuracy', metrics[0], self.epoch)
self.writer.add_scalar('val/acc_class', metrics[1], self.epoch)
self.writer.add_scalar('val/mean_iu', metrics[2], self.epoch)
self.writer.add_scalar('val/fwacc', metrics[3], self.epoch)
if self.scheduler != None:
self.scheduler.step(val_prec)
if self.training:
self.model.train()
def train_epoch(self):
self.model.train()
if self.freeze_bn:
self.model.apply(BNtoFixed)
n_class = self.train_loader.dataset.n_class
m = []
for batch_idx, (rgb_img, ddd_img,
target) in tqdm.tqdm(enumerate(self.train_loader),
total=len(self.train_loader),
desc=' epoch %d' % self.epoch,
ncols=80,
leave=False):
iteration = batch_idx + self.epoch * len(self.train_loader)
self.iteration = iteration
## prepare input and label
if self.cuda:
rgb_img = rgb_img.cuda()
ddd_img = ddd_img.cuda()
target = target.cuda()
## main training function
## compute output of feed forward
output = self.model(rgb_img, ddd_img)
if self.arch.startswith('bisenet'):
out_sup1 = F.interpolate(output[1],
size=target.size()[1:],
mode='bilinear')
out_sup2 = F.interpolate(output[2],
size=target.size()[1:],
mode='bilinear')
output = F.interpolate(output[0],
size=target.size()[1:],
mode='bilinear')
elif self.arch.startswith('icnet'):
out_sub24 = F.interpolate(output[1],
size=target.size()[1:],
mode='bilinear')
out_sub4 = F.interpolate(output[2],
size=target.size()[1:],
mode='bilinear')
output = F.interpolate(output[0],
size=target.size()[1:],
mode='bilinear')
elif self.train_loader.dataset.encode_label:
output = F.interpolate(output,
size=target.size()[2:],
mode='bilinear')
target = target.float()
## compute loss and backpropagate
loss = None
if self.arch.startswith('bisenet'):
loss_p = self.criterion(output, target)
loss_a1 = self.crit_aux1(out_sup1, target)
loss_a2 = self.crit_aux2(out_sup2, target)
loss = loss_p + self.alphas[0] * loss_a1 + self.alphas[
1] * loss_a2
elif self.arch.startswith('icnet'):
loss_sub124 = self.criterion(output, target)
loss_sub24 = self.crit_sub24(out_sub24, target)
loss_sub4 = self.crit_sub4(out_sub4, target)
loss = self.lambdas[0] * loss_sub4 + \
self.lambdas[1] * loss_sub24 + \
self.lambdas[2] * loss_sub124
else:
loss = self.criterion(output, target)
loss_data = loss.data.item()
self.optim.zero_grad()
if self.arch.startswith('rfnet'):
self.optim_dec.zero_grad()
if self.use_amp:
with amp.scale_loss(loss, self.optim) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if self.arch.startswith('bisenet') or 'effnet' in self.arch:
nn.utils.clip_grad_norm_(self.model.parameters(), 0.25)
self.optim.step()
if self.arch.startswith('rfnet'):
self.optim_dec.step()
## the stats
lbl_pred = output.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu().numpy()
metrics = label_accuracy_score(lbl_true, lbl_pred, n_class=n_class)
with open(osp.join(self.output_path, 'log.csv'), 'a') as f:
loss_data_str = '%.10f' % (loss_data)
metrics_str = ['%.10f' % (a) for a in list(metrics)]
elapsed_time = (
datetime.datetime.now(pytz.timezone('Asia/Jakarta')) -
self.timestamp_start).total_seconds()
log = [self.epoch, self.iteration] + [loss_data_str] + \
metrics_str + [''] * 5 + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
m.append(metrics)
if self.max_iter != None and self.iteration >= self.max_iter:
break
m = np.mean(np.array(m), axis=0)
self.writer.add_scalar('train/loss', loss_data, self.epoch)
self.writer.add_scalar('train/accuracy', m[0], self.epoch)
self.writer.add_scalar('train/acc_class', m[1], self.epoch)
self.writer.add_scalar('train/mean_iu', m[2], self.epoch)
self.writer.add_scalar('train/fwacc', m[3], self.epoch)
def test(test_loader, net, criterion, epoch, showall=False):
cnn0_loss, cnn0_accs, cnn0_mIoUs, cnn0_acc_clss, cnn0_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
cnn1_loss, cnn1_accs, cnn1_mIoUs, cnn1_acc_clss, cnn1_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
cnn2_loss, cnn2_accs, cnn2_mIoUs, cnn2_acc_clss, cnn2_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
# switch to evaluation mode
net.eval()
start_time = time.time()
for batch_idx, (datas, targets) in enumerate(test_loader):
if args.cuda:
datas = datas.cuda()
datas = Variable(datas, volatile=True)
# compute output
scores = net(datas)
multi_targets = combine_label(targets, COMB_DICTs)
multi_targets_tensor = torch.from_numpy(multi_targets).long()
if args.cuda:
multi_targets_tensor = multi_targets_tensor.cuda()
testlosses = []
for i, score in enumerate(scores):
targets_i = Variable(multi_targets_tensor[i, :, :, :])
testlosses.append(criterion(score, targets_i))
testloss = sum(testlosses)
# measure accuracy and record loss
preds = []
for score in scores:
p = score.data.max(1)[1]
preds.append(p)
for i, lbl_pred in enumerate(preds):
lbl_pred = lbl_pred.cpu().numpy()[:, :, :] # (n_batch, h, w)
lbl_true = multi_targets[i, :, :, :]
acc, acc_cls, mIoU, fscore = label_accuracy_score(
lbl_true, lbl_pred, n_class=NUM_CLASSES[i])
locals()['cnn%d_loss' % (i)].update(testlosses[i].data[0],
datas.size(0))
locals()['cnn%d_accs' % (i)].update(acc, datas.size(0))
locals()['cnn%d_acc_clss' % (i)].update(acc_cls, datas.size(0))
locals()['cnn%d_mIoUs' % (i)].update(mIoU, datas.size(0))
locals()['cnn%d_fscore' % (i)].update(fscore, datas.size(0))
if showall:
trues = decode_labels(targets, num_images=len(targets))
for i, t in enumerate(trues):
Image.fromarray(t).save(
'runs/{}_{}/results/{}_{}_gt.png'.format(
args.name, args.dataset, batch_idx, i))
Image.fromarray(
(unNormalize(datas.data).transpose(1, 2).transpose(
2, 3).cpu().numpy()[i] * 255).astype(np.uint8)).save(
'runs/{}_{}/results/{}_{}_img.png'.format(
args.name, args.dataset, batch_idx, i))
lbl_pred = preds[2]
pred = decode_labels(lbl_pred, num_images=len(lbl_pred))
for i, p in enumerate(pred):
Image.fromarray(p).save(
'runs/{}_{}/results/{}_{}_pred.png'.format(
args.name, args.dataset, batch_idx, i))
if showall and args.visdom:
plot_images(datas, [p.cpu().numpy() for p in preds],
multi_targets,
epoch,
split='test',
crop_size=map(int, args.input_size.split(',')))
duration = time.time() - start_time
print(
'\nTest set: Loss: {:.4f}, Acc: {:.2f}%, mIoU: {:.4f}, Acc_cls: {:.2f}%, f-score: {:.2f}% ({:.3f} sec)\n'
.format(cnn2_loss.avg, 100. * cnn2_accs.avg, cnn2_mIoUs.avg,
100 * cnn2_acc_clss.avg, 100 * cnn2_fscore.avg, duration))
if args.visdom:
for i in range(3):
plotter.plot('cnn%d_acc' % (i),
'test',
epoch,
locals()['cnn%d_accs' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_loss' % (i),
'test',
epoch,
locals()['cnn%d_loss' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_mIoU' % (i),
'test',
epoch,
locals()['cnn%d_mIoUs' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_acc_cls' % (i),
'test',
epoch,
locals()['cnn%d_acc_clss' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_fscore' % (i),
'test',
epoch,
locals()['cnn%d_fscore' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
# plot images in a grid
if epoch == 1 or epoch % 10 == 0:
plot_images(datas, [p.cpu().numpy() for p in preds],
multi_targets,
epoch,
split='test',
crop_size=map(int, args.input_size.split(',')))
return cnn2_accs.avg, cnn2_mIoUs.avg
def validate(self):
training = self.model.training
self.model.eval()
n_class = len(self.val_loader.dataset.class_names)
val_loss = 0
visualizations = []
label_trues, label_preds = [], []
with torch.no_grad():
for batch_idx, (data, target) in tqdm.tqdm(
enumerate(self.val_loader), total=len(self.val_loader),
desc='Valid iteration=%d' % self.iteration, ncols=80,
leave=False):
if self.cuda:
data, target = data.to('cuda'), target.to('cuda')
score = self.model(data)
loss = cross_entropy2d(score, target,
size_average=self.size_average)
if np.isnan(float(loss.item())):
raise ValueError('loss is nan while validating')
val_loss += float(loss.item()) / len(data)
imgs = data.data.cpu()
lbl_pred = score.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu()
for img, lt, lp in zip(imgs, lbl_true, lbl_pred):
img, lt = self.val_loader.dataset.untransform(img, lt)
label_trues.append(lt)
label_preds.append(lp)
if len(visualizations) < 9:
viz = fcn.utils.visualize_segmentation(
lbl_pred=lp, lbl_true=lt, img=img, n_class=n_class)
visualizations.append(viz)
metrics = utils.label_accuracy_score(
label_trues, label_preds, n_class)
out = osp.join(self.out, 'visualization_viz')
if not osp.exists(out):
os.makedirs(out)
out_file = osp.join(out, 'iter%012d.jpg' % self.iteration)
img_ = fcn.utils.get_tile_image(visualizations)
#scipy.misc.imsave(out_file, img_)
imageio.imwrite(out_file, img_)
plt.imshow(imageio.imread(out_file))
plt.savefig('imagesProduced/validate')
#plt.show()
val_loss /= len(self.val_loader)
with open(osp.join(self.out, 'log.csv'), 'a') as f:
elapsed_time = (
datetime.datetime.now(pytz.timezone('America/Bogota')) -
self.timestamp_start).total_seconds()
log = [self.epoch, self.iteration] + [''] * 5 + [val_loss] + list(metrics) + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
if is_best:
self.best_mean_iu = mean_iu
torch.save({
'epoch': self.epoch,
'iteration': self.iteration,
'arch': self.model.__class__.__name__,
'optim_state_dict': self.optim.state_dict(),
'model_state_dict': self.model.state_dict(),
'best_mean_iu': self.best_mean_iu,
}, osp.join(self.out, 'checkpoint.pth.tar'))
if is_best:
shutil.copy(osp.join(self.out, 'checkpoint.pth.tar'),
osp.join(self.out, 'model_best.pth.tar'))
if training:
self.model.train()
def validate(self):
# import matplotlib.pyplot as plt
training = self.model.training
self.model.eval()
n_class = len(self.val_loader.dataset.class_names)
val_loss = 0
visualizations = []
label_trues, label_preds = [], []
with torch.no_grad():
for batch_idx, (data, target) in tqdm.tqdm(
enumerate(self.val_loader),
total=len(self.val_loader),
desc="Valid iteration=%d" % self.iteration,
ncols=80,
leave=False,
):
data, target = data.to(self.cuda), target.to(self.cuda)
score = self.model(data)
loss = self.cross_entropy2d(score, target)
if np.isnan(float(loss.item())):
raise ValueError("loss is nan while validating")
val_loss += float(loss.item()) / len(data)
imgs = data.data.cpu()
lbl_pred = score.data.max(1)[1].cpu().numpy()[:, :, :]
lbl_true = target.data.cpu()
for img, lt, lp in zip(imgs, lbl_true, lbl_pred):
img, lt = self.val_loader.dataset.untransform(img, lt)
label_trues.append(lt)
label_preds.append(lp)
if len(visualizations) < 9:
viz = fcn.utils.visualize_segmentation(lbl_pred=lp,
lbl_true=lt,
img=img,
n_class=n_class)
visualizations.append(viz)
metrics = utils.label_accuracy_score(label_trues, label_preds, n_class)
out = osp.join(self.out, "visualization_viz")
if not osp.exists(out):
os.makedirs(out)
out_file = osp.join(out, "iter%012d.jpg" % self.iteration)
img_ = fcn.utils.get_tile_image(visualizations)
imageio.imwrite(out_file, img_)
# plt.imshow(imageio.imread(out_file))
# plt.show()
val_loss /= len(self.val_loader)
print(
"acc {:0.3f}, acc_cls {:0.3f}, mean_iu {:0.3f}, fwavacc {:0.3f}, val loss {}"
.format(metrics[0], metrics[1], metrics[2], metrics[3], val_loss))
with open(osp.join(self.out, "log.csv"), "a") as f:
elapsed_time = (datetime.datetime.now() -
self.timestamp_start).total_seconds()
log = ([self.epoch, self.iteration] + [""] * 5 + [val_loss] +
list(metrics) + [elapsed_time])
log = map(str, log)
f.write(",".join(log) + "\n")
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
if is_best:
self.best_mean_iu = mean_iu
torch.save(
{
"epoch": self.epoch,
"iteration": self.iteration,
"arch": self.model.__class__.__name__,
"optim_state_dict": self.optim.state_dict(),
"model_state_dict": self.model.state_dict(),
"best_mean_iu": self.best_mean_iu,
},
osp.join(self.out, "checkpoint.pth.tar"),
)
if is_best:
shutil.copy(
osp.join(self.out, "checkpoint.pth.tar"),
osp.join(self.out, "model_best.pth.tar"),
)
if training:
self.model.train()
def validate(self, both_fcn_and_seenmask=False):
self.model.eval()
val_loss = 0
lbl_trues, lbl_preds, visualizations = [], [], []
for batch_idx, (data, target) in enumerate(self.val_loader):
if both_fcn_and_seenmask:
score, loss, lbl_pred, lbl_true = self.forward_szn(
data, target)
else:
score, loss, lbl_pred, lbl_true = self.forward(data, target)
val_loss += float(loss.data[0])
print(
"Test Epoch {:<5} | Iteration {:<5} | Loss {:5.5f} | Score Sum {:10.5f}"
.format(int(self.epoch), int(batch_idx), float(loss.data[0]),
float(score.sum().data[0])))
img, lt, lp = data[0], lbl_true[0], lbl_pred[
0] # eliminate first dimension (n=1) for visualization
img, lt = self.val_loader.dataset.untransform(img, lt)
lbl_trues.append(lt)
lbl_preds.append(lp)
# generate visualization for first few images of val_loader
if len(visualizations) < 25:
viz = vis_utils.visualize_segmentation(
lbl_pred=lp,
lbl_true=lt,
img=img,
n_class=self.n_class,
label_names=self.label_names,
unseen=self.val_unseen)
visualizations.append(viz)
# save the visualizaton image
if both_fcn_and_seenmask:
out = osp.join(self.log_dir, 'szn_viz')
else:
out = osp.join(self.log_dir, 'fcn_viz')
if not osp.exists(out):
os.makedirs(out)
out_file = osp.join(out, 'epoch%d.jpg' % self.epoch)
viz_img = fcn.utils.get_tile_image(visualizations)
scipy.misc.imsave(out_file, viz_img)
# update the validation log for the current epoch
if self.unseen:
metrics = utils.label_accuracy_score(lbl_trues,
lbl_preds,
self.n_class,
unseen=self.val_unseen)
metrics, seen_metrics, unseen_metrics = metrics
self.tb_writer.add_scalar('fcn/val/seen/pxl_acc', seen_metrics[0],
self.epoch)
self.tb_writer.add_scalar('fcn/val/seen/class_acc',
seen_metrics[1], self.epoch)
self.tb_writer.add_scalar('fcn/val/seen/mean_iu', seen_metrics[2],
self.epoch)
self.tb_writer.add_scalar('fcn/val/seen/fwavacc', seen_metrics[3],
self.epoch)
self.tb_writer.add_scalar('fcn/val/unseen/pxl_acc',
unseen_metrics[0], self.epoch)
self.tb_writer.add_scalar('fcn/val/unseen/class_acc',
unseen_metrics[1], self.epoch)
self.tb_writer.add_scalar('fcn/val/unseen/mean_iu',
unseen_metrics[2], self.epoch)
self.tb_writer.add_scalar('fcn/val/unseen/fwavacc',
unseen_metrics[3], self.epoch)
print('seen pxl_acc: %.3f' % seen_metrics[0])
print('seen class_acc: %.3f' % seen_metrics[1])
print('seen mean_iu: %.3f' % seen_metrics[2])
print('seen fwavacc: %.3f' % seen_metrics[3])
print('unseen pxl_acc: %.3f' % unseen_metrics[0])
print('unseen class_acc: %.3f' % unseen_metrics[1])
print('unseen mean_iu: %.3f' % unseen_metrics[2])
print('unseen fwavacc: %.3f' % unseen_metrics[3])
else:
metrics = utils.label_accuracy_score(lbl_trues, lbl_preds,
self.n_class)
val_loss /= len(
self.val_loader) # val loss is averaged across all the images
with open(osp.join(self.log_dir, 'val_log.csv'), 'a') as f:
elapsed_time = datetime.datetime.now(
pytz.timezone('US/Eastern')) - self.timestamp_start
if self.unseen:
log = [self.epoch, self.iteration] + [val_loss] + list(
metrics) + list(seen_metrics) + list(unseen_metrics) + [
elapsed_time
]
else:
log = [self.epoch, self.iteration
] + [val_loss] + list(metrics) + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
# write metrics to tensorboard
self.tb_writer.add_scalar('fcn/val/loss', val_loss, self.epoch)
self.tb_writer.add_scalar('fcn/val/pxl_acc', metrics[0], self.epoch)
self.tb_writer.add_scalar('fcn/val/class_acc', metrics[1], self.epoch)
self.tb_writer.add_scalar('fcn/val/mean_iu', metrics[2], self.epoch)
self.tb_writer.add_scalar('fcn/val/fwavacc', metrics[3], self.epoch)
self.tb_writer.add_image('fcn/segmentations', viz_img, self.epoch)
print('overall pxl_acc: %.3f' % metrics[0])
print('overall class_acc: %.3f' % metrics[1])
print('overall mean_iu: %.3f' % metrics[2])
print('overall fwavacc: %.3f' % metrics[3])
# track and update the best mean intersection over union
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
if is_best:
self.best_mean_iu = mean_iu
# checkpoint the model's weights
torch.save(
{
'epoch': self.epoch,
'iteration': self.iteration,
'arch': self.model.__class__.__name__,
'optim_state_dict': self.optim.state_dict(),
'model_state_dict': self.model.state_dict(),
'best_mean_iu': self.best_mean_iu,
}, osp.join(self.log_dir, 'checkpoint'))
# save the weights for the best performing model so far
if is_best:
shutil.copy(osp.join(self.log_dir, 'checkpoint'),
osp.join(self.log_dir, 'best'))
def train(train_loader, net, criterion, optimizer, epoch):
cnn0_loss, cnn0_accs, cnn0_mIoUs, cnn0_acc_clss, cnn0_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
cnn1_loss, cnn1_accs, cnn1_mIoUs, cnn1_acc_clss, cnn1_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
cnn2_loss, cnn2_accs, cnn2_mIoUs, cnn2_acc_clss, cnn2_fscore = AverageMeter(
), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
# switch to train mode
net.train()
start_time = time.time()
all_start = start_time
print(' Train Epoch | Loss | Acc | mIoU |'
' Acc_cls | f-score | Time ')
for batch_idx, (datas, targets) in enumerate(train_loader):
if args.cuda:
datas = datas.cuda()
datas = Variable(datas)
# compute output
scores = net(datas)
multi_targets = combine_label(targets, COMB_DICTs)
multi_targets_tensor = torch.from_numpy(multi_targets).long()
if args.cuda:
multi_targets_tensor = multi_targets_tensor.cuda()
losses = []
for i, score in enumerate(scores):
targets_i = Variable(multi_targets_tensor[i, :, :, :])
losses.append(criterion(score, targets_i))
loss = sum(losses)
# compute gradient and do optimizer step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure accuracy and record loss
preds = []
for score in scores:
p = score.data.max(1)[1]
preds.append(p)
for i, lbl_pred in enumerate(preds):
lbl_pred = lbl_pred.cpu().numpy()[:, :, :] # (n_batch, h, w)
lbl_true = multi_targets[i, :, :, :]
acc, acc_cls, mIoU, fscore = label_accuracy_score(
lbl_true, lbl_pred, n_class=NUM_CLASSES[i])
locals()['cnn%d_loss' % (i)].update(losses[i].data[0],
datas.size(0))
locals()['cnn%d_accs' % (i)].update(acc, datas.size(0))
locals()['cnn%d_acc_clss' % (i)].update(acc_cls, datas.size(0))
locals()['cnn%d_mIoUs' % (i)].update(mIoU, datas.size(0))
locals()['cnn%d_fscore' % (i)].update(fscore, datas.size(0))
if batch_idx % args.log_interval == 0:
duration = time.time() - start_time
print(
'{:3d}[{:4d}/{:4d}] | {:.3f}({:.3f}) | {:2.2f}%({:2.2f}%) | {:.2f}({:.2f}) | {:2.2f}%({:2.2f}%) |'
' {:2.2f}%({:2.2f}%) | ({:.3f} sec)'.format(
epoch, batch_idx * len(datas), len(train_loader.dataset),
cnn2_loss.val, cnn2_loss.avg, 100. * cnn2_accs.val,
100. * cnn2_accs.avg, cnn2_mIoUs.val, cnn2_mIoUs.avg,
100. * cnn2_acc_clss.val, 100. * cnn2_acc_clss.avg,
100. * cnn2_fscore.val, 100. * cnn2_fscore.avg, duration))
start_time = time.time()
duration = time.time() - all_start
print(
'Train Summary: Epoch {}, Acc: {:.2f}%, mIoU: {:.2f}, Acc_cls: {:.2f}%, f-score: {:.2f}% ({:.3f} sec)'
.format(epoch, 100. * cnn2_accs.avg, cnn2_mIoUs.avg,
100 * cnn2_acc_clss.avg, 100 * cnn2_fscore.avg, duration))
# log avg values to visdom
if args.visdom:
for i in range(3):
plotter.plot('cnn%d_acc' % (i),
'train',
epoch,
locals()['cnn%d_accs' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_loss' % (i),
'train',
epoch,
locals()['cnn%d_loss' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_mIoU' % (i),
'train',
epoch,
locals()['cnn%d_mIoUs' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_acc_cls' % (i),
'train',
epoch,
locals()['cnn%d_acc_clss' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
plotter.plot('cnn%d_fscore' % (i),
'train',
epoch,
locals()['cnn%d_fscore' % (i)].avg,
exp_name=args.name + '_' + args.dataset)
# plot images in a grid
if epoch == 1 or epoch % 10 == 0:
plot_images(datas, [p.cpu().numpy() for p in preds],
multi_targets,
epoch,
split='train',
crop_size=map(int, args.input_size.split(',')))
def validate(self):
self.model.eval()
val_loss = 0
lbl_trues, lbl_preds, visualizations = [], [], []
for batch_idx, (data, target) in enumerate(self.val_loader):
score, loss, lbl_pred, lbl_true = self.forward(data, target)
val_loss += float(loss.data[0])
print("Seenmask Test Epoch {:<5} | Iteration {:<5} | Loss {:5.5f} | Score Sum {:10.5f}".format(int(self.epoch), int(batch_idx), float(loss.data[0]), float(score.sum().data[0])))
img, lt, lp = data[0], lbl_true[0], lbl_pred[0] # eliminate first dimension (n=1) for visualization
img, lt = self.val_loader.dataset.untransform(img, lt)
lbl_trues.append(lt)
lbl_preds.append(lp)
# generate visualization for first few images of val_loader
if len(visualizations) < 25:
viz = vis_utils.visualize_seenmask(lbl_pred=lp, lbl_true=lt, img=img, n_class=self.n_class, unseen=self.unseen)
visualizations.append(viz)
# save the visualizaton image
out = osp.join(self.log_dir, 'seenmask_viz')
if not osp.exists(out):
os.makedirs(out)
out_file = osp.join(out, 'epoch%d.jpg' % self.epoch)
viz_img = fcn.utils.get_tile_image(visualizations)
scipy.misc.imsave(out_file, viz_img)
metrics = utils.label_accuracy_score(lbl_trues, lbl_preds, self.n_class)
val_loss /= len(self.val_loader) # val loss is averaged across all the images
with open(osp.join(self.log_dir, 'seenmask_val_log.csv'), 'a') as f:
elapsed_time = datetime.datetime.now(pytz.timezone('US/Eastern')) - self.timestamp_start
log = [self.epoch, self.iteration] + [val_loss] + list(metrics) + [elapsed_time]
log = map(str, log)
f.write(','.join(log) + '\n')
# write metrics to tensorboard
self.tb_writer.add_scalar('seenmask/val/loss', val_loss, self.epoch)
self.tb_writer.add_scalar('seenmask/val/pxl_acc', metrics[0], self.epoch)
self.tb_writer.add_scalar('seenmask/val/class_acc', metrics[1], self.epoch)
self.tb_writer.add_scalar('seenmask/val/mean_iu', metrics[2], self.epoch)
self.tb_writer.add_scalar('seenmask/val/fwavacc', metrics[3], self.epoch)
self.tb_writer.add_image('fcn/segmentations', viz_img, self.epoch)
print('pxl_acc: %.3f'%metrics[0])
print('class_acc: %.3f'%metrics[1])
print('mean_iu: %.3f'%metrics[2])
print('fwavacc: %.3f'%metrics[3])
# track and update the best mean intersection over union
mean_iu = metrics[2]
is_best = mean_iu > self.best_mean_iu
if is_best:
self.best_mean_iu = mean_iu
self.checkpoint['model_state_dict'] = self.model.state_dict() # TODO: verify
torch.save(self.checkpoint, osp.join(self.log_dir, 'best'))
def pseudo_labeling(num_epochs, model, data_loader, val_loader,
unlabeled_loader, device, val_every, file_name):
# Instead of using current epoch we use a "step" variable to calculate alpha_weight
# This helps the model converge faster
from torch.optim.swa_utils import AveragedModel, SWALR
from segmentation_models_pytorch.losses import SoftCrossEntropyLoss, JaccardLoss
from adamp import AdamP
criterion = [
SoftCrossEntropyLoss(smooth_factor=0.1),
JaccardLoss('multiclass', classes=12)
]
optimizer = AdamP(params=model.parameters(), lr=0.0001, weight_decay=1e-6)
swa_scheduler = SWALR(optimizer, swa_lr=0.0001)
swa_model = AveragedModel(model)
optimizer = Lookahead(optimizer, la_alpha=0.5)
step = 100
size = 256
best_mIoU = 0
model.train()
print('Start Pseudo-Labeling..')
for epoch in range(num_epochs):
hist = np.zeros((12, 12))
for batch_idx, (imgs, image_infos) in enumerate(unlabeled_loader):
# Forward Pass to get the pseudo labels
# --------------------------------------------- test(unlabelse)를 모델에 통과
model.eval()
outs = model(torch.stack(imgs).to(device))
oms = torch.argmax(outs.squeeze(), dim=1).detach().cpu().numpy()
oms = torch.Tensor(oms)
oms = oms.long()
oms = oms.to(device)
# --------------------------------------------- 학습
model.train()
# Now calculate the unlabeled loss using the pseudo label
imgs = torch.stack(imgs)
imgs = imgs.to(device)
# preds_array = preds_array.to(device)
output = model(imgs)
loss = 0
for each in criterion:
loss += each(output, oms)
unlabeled_loss = alpha_weight(step) * loss
# Backpropogate
optimizer.zero_grad()
unlabeled_loss.backward()
optimizer.step()
output = torch.argmax(output.squeeze(),
dim=1).detach().cpu().numpy()
hist = add_hist(hist,
oms.detach().cpu().numpy(),
output,
n_class=12)
if (batch_idx + 1) % 25 == 0:
acc, acc_cls, mIoU, fwavacc = label_accuracy_score(hist)
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}, mIoU:{:.4f}'.
format(epoch + 1, num_epochs, batch_idx + 1,
len(unlabeled_loader), unlabeled_loss.item(),
mIoU))
# For every 50 batches train one epoch on labeled data
# 50배치마다 라벨데이터를 1 epoch학습
if batch_idx % 50 == 0:
# Normal training procedure
for batch_idx, (images, masks, _) in enumerate(data_loader):
labeled_loss = 0
images = torch.stack(images)
# (batch, channel, height, width)
masks = torch.stack(masks).long()
# gpu 연산을 위해 device 할당
images, masks = images.to(device), masks.to(device)
output = model(images)
for each in criterion:
labeled_loss += each(output, masks)
optimizer.zero_grad()
labeled_loss.backward()
optimizer.step()
# Now we increment step by 1
step += 1
if (epoch + 1) % val_every == 0:
avrg_loss, val_mIoU = validation(epoch + 1, model, val_loader,
criterion, device)
if val_mIoU > best_mIoU:
print('Best performance at epoch: {}'.format(epoch + 1))
print('Save model in', saved_dir)
best_mIoU = val_mIoU
save_model(model, file_name=file_name)
model.train()
if epoch > 3:
swa_model.update_parameters(model)
swa_scheduler.step()
