def evaluate(args):
test_data, test_label = load_data(
'experiment/data/modelnet40_ply_hdf5_2048/', train=False)
testDataset = ModelNetDataLoader(test_data, test_label)
testDataLoader = torch.utils.data.DataLoader(testDataset,
batch_size=args.batch_size,
shuffle=False)
log.debug('Building Model', args.model_name)
if args.model_name == 'pointnet':
num_class = 40
model = PointNetCls(num_class, args.feature_transform)
else:
model = PointNet2ClsMsg()
torch.backends.cudnn.benchmark = True
model = torch.nn.DataParallel(model).cuda()
log.debug('Using gpu:', args.gpu)
if args.pretrain is None:
log.err('No pretrain model')
return
log.debug('Loading pretrain model...')
state_dict = torch.load(args.pretrain)
model.load_state_dict(state_dict)
acc = test_clf(model.eval(), testDataLoader)
log.msg(Test_Accuracy='%.5f' % (acc))
def test_kitti_semseg(model, loader, model_name, num_classes, class_names):
ious = np.zeros((num_classes, ), dtype=np.float32)
count = np.zeros((num_classes, ), dtype=np.uint32)
count[0] = 1
accuracy = []
for points, target in tqdm(loader,
total=len(loader),
smoothing=0.9,
dynamic_ncols=True):
batch_size, num_point, _ = points.size()
points = points.float().transpose(2, 1).cuda()
target = target.long().cuda()
with torch.no_grad():
if model_name == 'pointnet':
pred, _ = model(points)
else:
pred = model(points)
pred_choice = pred.argmax(-1)
target = target.squeeze(-1)
for class_id in range(num_classes):
I = torch.sum((pred_choice == class_id)
& (target == class_id)).cpu().item()
U = torch.sum((pred_choice == class_id)
| (target == class_id)).cpu().item()
iou = 1 if U == 0 else I / U
ious[class_id] += iou
count[class_id] += 1
correct = (pred_choice == target).sum().cpu().item()
accuracy.append(correct / (batch_size * num_point))
categorical_iou = ious / count
df = pd.DataFrame(categorical_iou, columns=['mIOU'], index=class_names)
df = df.sort_values(by='mIOU', ascending=False)
log.info('categorical mIOU')
log.msg(df)
acc = np.mean(accuracy)
miou = np.mean(categorical_iou[1:])
return acc, miou
def train(args):
experiment_dir = mkdir('experiment/')
checkpoints_dir = mkdir('experiment/%s/' % (args.model_name))
kitti_utils = Semantic_KITTI_Utils(KITTI_ROOT, subset=args.subset)
class_names = kitti_utils.class_names
num_classes = kitti_utils.num_classes
if args.subset == 'inview':
train_npts = 8000
test_npts = 24000
if args.subset == 'all':
train_npts = 50000
test_npts = 100000
log.info(subset=args.subset, train_npts=train_npts, test_npts=test_npts)
dataset = SemKITTI_Loader(KITTI_ROOT,
train_npts,
train=True,
subset=args.subset)
dataloader = DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
test_dataset = SemKITTI_Loader(KITTI_ROOT,
test_npts,
train=False,
subset=args.subset)
testdataloader = DataLoader(test_dataset,
batch_size=int(args.batch_size / 2),
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.model_name == 'pointnet':
model = PointNetSeg(num_classes, input_dims=4, feature_transform=True)
else:
model = PointNet2SemSeg(num_classes, feature_dims=1)
if args.optimizer == 'SGD':
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
elif args.optimizer == 'Adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=args.learning_rate,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=1e-4)
torch.backends.cudnn.benchmark = True
model = torch.nn.DataParallel(model)
model.cuda()
log.info('Using gpu:', args.gpu)
if args.pretrain is not None:
log.info('Use pretrain model...')
model.load_state_dict(torch.load(args.pretrain))
init_epoch = int(args.pretrain[:-4].split('-')[-1])
log.info('Restart training', epoch=init_epoch)
else:
log.msg('Training from scratch')
init_epoch = 0
best_acc = 0
best_miou = 0
for epoch in range(init_epoch, args.epoch):
model.train()
lr = calc_decay(args.learning_rate, epoch)
log.info(subset=args.subset,
model=args.model_name,
gpu=args.gpu,
epoch=epoch,
lr=lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
for points, target in tqdm(dataloader,
total=len(dataloader),
smoothing=0.9,
dynamic_ncols=True):
points = points.float().transpose(2, 1).cuda()
target = target.long().cuda()
if args.model_name == 'pointnet':
logits, trans_feat = model(points)
else:
logits = model(points)
#logits = logits.contiguous().view(-1, num_classes)
#target = target.view(-1, 1)[:, 0]
#loss = F.nll_loss(logits, target)
logits = logits.transpose(2, 1)
loss = nn.CrossEntropyLoss()(logits, target)
if args.model_name == 'pointnet':
loss += feature_transform_reguliarzer(trans_feat) * 0.001
optimizer.zero_grad()
loss.backward()
optimizer.step()
torch.cuda.empty_cache()
acc, miou = test_kitti_semseg(model.eval(), testdataloader,
args.model_name, num_classes,
class_names)
save_model = False
if acc > best_acc:
best_acc = acc
if miou > best_miou:
best_miou = miou
save_model = True
if save_model:
fn_pth = '%s-%s-%.5f-%04d.pth' % (args.model_name, args.subset,
best_miou, epoch)
log.info('Save model...', fn=fn_pth)
torch.save(model.state_dict(),
os.path.join(checkpoints_dir, fn_pth))
else:
log.info('No need to save model')
log.warn('Curr', accuracy=acc, mIOU=miou)
log.warn('Best', accuracy=best_acc, mIOU=best_miou)
def train(args):
experiment_dir = mkdir('experiment/')
checkpoints_dir = mkdir('experiment/%s/'%(args.model_name))
kitti_utils = Semantic_KITTI_Utils(ROOT, subset = args.subset)
class_names = kitti_utils.class_names
num_classes = kitti_utils.num_classes
if args.subset == 'inview':
train_npts = 2000
test_npts = 2500
if args.subset == 'all':
train_npts = 50000
test_npts = 100000
log.info(subset=args.subset, train_npts=train_npts, test_npts=test_npts)
dataset = SemKITTI_Loader(ROOT, train_npts, train = True, subset = args.subset)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
test_dataset = SemKITTI_Loader(ROOT, test_npts, train = False, subset = args.subset)
testdataloader = DataLoader(test_dataset, batch_size=int(args.batch_size/2), shuffle=False,
num_workers=args.workers, pin_memory=True)
if args.model_name == 'pointnet':
model = PointNetSeg(num_classes, input_dims = 4, feature_transform = True)
else:
model = PointNet2SemSeg(num_classes, feature_dims = 1)
if args.optimizer == 'SGD':
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
elif args.optimizer == 'Adam':
optimizer = torch.optim.Adam(
model.parameters(),
lr=args.learning_rate,
betas=(0.9, 0.999),
eps=1e-08,
weight_decay=1e-4)
torch.backends.cudnn.benchmark = True
model = torch.nn.DataParallel(model)
#use more than 1 gpu
model.cuda()
log.info('Using gpu:',args.gpu)
if args.pretrain is not None:
log.info('Use pretrain model...')
model.load_state_dict(torch.load(args.pretrain))
init_epoch = int(args.pretrain[:-4].split('-')[-1])
log.info('Restart training', epoch=init_epoch)
else:
log.msg('Training from scratch')
init_epoch = 0
best_acc = 0
best_miou = 0
#->5.12 add
# to show details of epoch training
loss_list = []
miou_list = []
acc_list = []
epoch_time = []
lr_list = []
#-<5.12 add
for epoch in range(init_epoch,args.epoch):
model.train()
lr = calc_decay(args.learning_rate, epoch)
log.info(model=args.model_name, gpu=args.gpu, epoch=epoch, lr=lr)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
for points, target in tqdm(dataloader, total=len(dataloader), smoothing=0.9, dynamic_ncols=True):
points = points.float().transpose(2, 1).cuda()
target = target.long().cuda()
if args.model_name == 'pointnet':
logits, trans_feat = model(points)
else:
logits = model(points)
#logits = logits.contiguous().view(-1, num_classes)
#target = target.view(-1, 1)[:, 0]
#loss = F.nll_loss(logits, target)
logits = logits.transpose(2, 1)
loss = nn.CrossEntropyLoss()(logits, target)
if args.model_name == 'pointnet':
loss += feature_transform_reguliarzer(trans_feat) * 0.001
# loss_list.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
torch.cuda.empty_cache()
acc, miou = test_kitti_semseg(model.eval(), testdataloader,
args.model_name,num_classes,class_names)
# miou_list.append(np.asscalar(miou))
# acc_list.append(np.asscalar(acc))
save_model = False
if acc > best_acc:
best_acc = acc
if miou > best_miou:
best_miou = miou
save_model = True
#->5.12 add
loss_list.append(loss.item())
miou_list.append(np.asscalar(miou))
acc_list.append(np.asscalar(acc))
epoch_time.append(epoch)
lr_list.append(lr)
#->5.12 add
if save_model:
fn_pth = '%s-%.5f-%04d.pth' % (args.model_name, best_miou, epoch)
log.info('Save model...',fn = fn_pth)
torch.save(model.state_dict(), os.path.join(checkpoints_dir, fn_pth))
else:
log.info('No need to save model')
# 3.31 add |>
# show(args)
# 3.31 add |<
log.warn('Curr',accuracy=acc, mIOU=miou)
log.warn('Best',accuracy=best_acc, mIOU=best_miou)
# 5.15 add
label_size = {"size":40}
fig = plt.figure()
plt.plot(epoch_time,loss_list,label = "loss")
plt.plot(epoch_time,miou_list,label = "mIOU")
plt.plot(epoch_time,acc_list,label = "accuracy")
# plt.plot(epoch_time,lr_list,label = "learning rate")
plt.xlabel("epoch time", fontsize=40)
plt.ylabel("value", fontsize=40)
plt.title("training trendency", fontsize=60)
plt.tick_params(labelsize=40)
plt.legend(prop = label_size)
plt.show()