def evaluate(model_test):
model_test.eval()
total_correct = 0
data_eval = shapenet_dataset(datalist_path=args.data_eval)
eval_loader = torch.utils.data.DataLoader(data_eval,
num_workers=4,
batch_size=4,
shuffle=True,
collate_fn=pts_collate_seg)
print("dataset size:", len(eval_loader.dataset))
for batch_idx, (pts, label, seg) in enumerate(eval_loader):
## pts [N,P,3] label [N,] seg [N,P]
if is_GPU:
pts = Variable(pts.cuda())
label = Variable(label.cuda())
seg_label = Variable(seg.cuda())
else:
pts = Variable(pts)
label = Variable(label)
seg_label = Variable(seg)
## pred [N,50,P] trans [N,64,64]
pred, trans = net(pts)
_, pred_index = torch.max(pred, dim=1) ##[N,P]
num_correct = (pred_index.eq(seg_label)).data.cpu().sum()
total_correct += num_correct.item()
print('the average correct rate:{}'.format(
total_correct * 1.0 / (len(eval_loader.dataset) * 2048)))
model_test.train()
with open(logname, 'a') as f:
f.write('\nthe evaluate average accuracy:{}'.format(
total_correct * 1.0 / (len(eval_loader.dataset) * 2048)))
def evaluate(model_test):
model_test.eval()
total_correct = 0
data_eval = shapenet_dataset(root=args.data_eval,
classification=True,
train=False)
eval_loader = torch.utils.data.DataLoader(data_eval,
batch_size=args.batch_size,
shuffle=True,
num_workers=4,
collate_fn=pts_collate)
print("dataset size:", len(eval_loader.dataset))
for batch_idx, (split_dims, pts, label) in enumerate(eval_loader):
t1 = time.time()
print('batch {}'.format(batch_idx))
if is_GPU:
pts = Variable(pts.cuda())
label = Variable(label.cuda())
else:
pts = Variable(pts)
label = Variable(label)
pred = net(pts, split_dims)
_, pred_index = torch.max(pred, dim=1)
num_correct = (pred_index.eq(label)).data.cpu().sum().item()
total_correct += num_correct
print('the average correct rate:{}'.format(total_correct * 1.0 /
(len(eval_loader.dataset))))
model_test.train()
with open(logname, 'a') as f:
f.write('\nthe evaluate average accuracy:{}\n'.format(
total_correct * 1.0 / (len(eval_loader.dataset))))
def evaluate(model_test):
model_test.eval()
total_correct = 0
total_seen_class = [0 for _ in range(NUM_CLASSES)]
total_mIOU_class = [0 for _ in range(NUM_CLASSES)]
data_eval = shapenet_dataset(datalist_path=args.data_eval)
eval_loader = torch.utils.data.DataLoader(data_eval,num_workers=4,
batch_size=4, shuffle=True, collate_fn=pts_collate_seg)
print("dataset size:", len(eval_loader.dataset))
for batch_idx, (pts, label, seg) in enumerate(eval_loader):
## pts [N,P,3] label [N,] seg [N,P]
if is_GPU:
pts = Variable(pts.cuda())
label = Variable(label.cuda())
seg_label = Variable(seg.cuda())
else:
pts = Variable(pts)
label = Variable(label)
seg_label = Variable(seg)
## pred [N,50,P] trans [N,64,64]
pred, trans = net(pts)
_, pred_index = torch.max(pred, dim=1) ##[N,P]
num_correct = (pred_index.eq(seg_label)).data.cpu().sum()
print('in batch{} acc={}'.format(batch_idx, num_correct.item() * 1.0 / (4 * 2048)))
total_correct += num_correct.item()
################
## compute mIOU
iou_batch = []
for i in range(pred.size()[0]): ## B
iou_pc = []
for part in part_label[label[i]]: ## for each shape
gt = (seg[i] == part) ## gt of this part_idx
predict = (pred_index[i] == part).cpu()
intersection = (gt + predict) == 2
union = (gt + predict) >= 1
# print(intersection)
# print(union)
# assert False
if union.sum() == 0:
iou_part = 1.0
else:
iou_part = intersection.int().sum().item() / (union.int().sum().item() + 0.0001)
iou_pc.append(iou_part)
##np.asarray(iou_pc).mean() the mIOU of this shape
#iou_batch.append(np.asarray(iou_pc).mean())
total_mIOU_class[label[i]] += np.asarray(iou_pc).mean()
total_seen_class[label[i]] += 1
## mIOU of each class
mIOU_class = np.array(total_mIOU_class) / np.array(total_seen_class,dtype=np.float)
print ('##############################################################')
print('the average correct rate:{}'.format(total_correct * 1.0 / (len(eval_loader.dataset)*2048)))
print('the mean IOU overall :{}'.format((mIOU_class * weight_cls).sum()))
print ('##############################################################')
print ('mIOU of classes')
for i,cls in enumerate(data_eval.classname):
print ('{}: {}'.format(cls,mIOU_class[i]))
metavar='LR',
help='initial learning rate')
parser.add_argument('--resume',
default='checkpoint.pth',
type=str,
metavar='PATH',
help='path to latest checkpoint ')
args = parser.parse_args()
logname = args.log
if is_GPU:
torch.cuda.set_device(args.gpu)
my_dataset = shapenet_dataset(root=args.data, classification=True)
data_loader = torch.utils.data.DataLoader(my_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4,
collate_fn=pts_collate)
net = KDNet()
if is_GPU:
net = net.cuda()
optimizer = torch.optim.Adam(net.parameters(), lr=args.lr, betas=(0.5, 0.999))
critenrion = nn.NLLLoss()
def save_checkpoint(epoch, model, num_iter):
torch.save(
x = self.conv5(x) ##[N,1024,P]
global_feature = torch.squeeze(self.max_pool(x)) ##[N,1024]
global_feature = global_feature.view(-1, 1024, 1).repeat(1, 1, self.num_pts)
pts_seg_feature = torch.cat([global_feature, pts_feature_align], dim=1) ##[N,1088,P]
scores = self.classifer(pts_seg_feature) ##[N,50,P]
pred = F.log_softmax(scores, dim=1)
return pred,trans2
if __name__ == '__main__':
from data_utils import shapenet_dataset, pts_collate_seg
my_dataset = shapenet_dataset('/home/gaoyuzhe/Downloads/3d_data/hdf5_data/test_hdf5_file_list.txt')
data_loader = torch.utils.data.DataLoader(my_dataset,
batch_size=2, shuffle=True, collate_fn=pts_collate_seg)
net = PointNet_seg()
for batch_idx, (pts, label, seg) in enumerate(data_loader):
if False:
pts = Variable(pts.cuda())
label = Variable(label.cuda())
else:
pts = Variable(pts)
label = Variable(label)
pred,trans = net(pts)
print (pred.size())
print (trans.size())
exit()
def evaluate(model_test):
NUM_CLASSES = 16
part_label = [[0, 1, 2, 3], [4, 5], [6, 7], [8, 9, 10,
11], [12, 13, 14, 15],
[16, 17, 18], [19, 20, 21], [22, 23], [24, 25, 26, 27],
[28, 29], [30, 31, 32, 33, 34, 35], [36, 37], [38, 39, 40],
[41, 42, 43], [44, 45, 46], [47, 48, 49]]
num_dataset_cls = np.array(
[341, 14, 11, 158, 704, 14, 159, 80, 286, 83, 51, 38, 44, 12, 31, 848])
weight_cls = num_dataset_cls.astype(
np.float32) * 1.0 / num_dataset_cls.sum().astype(np.float32)
model_test.eval()
total_correct = 0
total_seen_class = [0 for _ in range(NUM_CLASSES)]
total_mIOU_class = [0 for _ in range(NUM_CLASSES)]
data_eval = shapenet_dataset(datalist_path=args.data_eval)
eval_loader = torch.utils.data.DataLoader(data_eval,
num_workers=4,
batch_size=4,
shuffle=True,
collate_fn=pts_collate_seg)
print("dataset size:", len(eval_loader.dataset))
for batch_idx, (pts, label, seg) in enumerate(eval_loader):
## pts [N,P,3] label [N,] seg [N,P]
if is_GPU:
pts = Variable(pts.cuda())
label = Variable(label.cuda())
seg_label = Variable(seg.cuda())
else:
pts = Variable(pts)
label = Variable(label)
seg_label = Variable(seg)
## pred [N,50,P] trans [N,64,64]
pred = net(pts)
_, pred_index = torch.max(pred, dim=1) ##[N,P]
num_correct = (pred_index.eq(seg_label)).data.cpu().sum()
print('in batch{} acc={}'.format(batch_idx,
num_correct.item() * 1.0 /
(4 * 2048)))
total_correct += num_correct.item()
################
## compute mIOU
iou_batch = []
for i in range(pred.size()[0]): ## B
iou_pc = []
for part in part_label[label[i]]: ## for each shape
gt = (seg[i] == part) ## gt of this part_idx
predict = (pred_index[i] == part).cpu()
intersection = (gt + predict) == 2
union = (gt + predict) >= 1
# print(intersection)
# print(union)
# assert False
if union.sum() == 0:
iou_part = 1.0
else:
iou_part = intersection.int().sum().item() / (
union.int().sum().item() + 0.0001)
iou_pc.append(iou_part)
##np.asarray(iou_pc).mean() the mIOU of this shape
#iou_batch.append(np.asarray(iou_pc).mean())
total_mIOU_class[label[i]] += np.asarray(iou_pc).mean()
total_seen_class[label[i]] += 1
## mIOU of each class
mIOU_class = np.array(total_mIOU_class) / np.array(total_seen_class,
dtype=np.float)
with open(logname, 'a') as f:
f.write(
'##############################################################\n')
f.write('the average correct rate:{}\n'.format(
total_correct * 1.0 / (len(eval_loader.dataset) * 2048)))
f.write('the mean IOU overall :{}\n'.format(
(mIOU_class * weight_cls).sum()))
f.write(
'##############################################################\n')
resume = os.path.join(LOG_DIR, "checkpoint.pth")
else:
resume = args.resume
logname = os.path.join(LOG_DIR, 'log.txt')
optfile = os.path.join(LOG_DIR, 'opt.txt')
with open(optfile, 'wt') as opt_f:
opt_f.write('------------ Options -------------\n')
for k, v in sorted(vars(args).items()):
opt_f.write('%s: %s\n' % (str(k), str(v)))
opt_f.write('-------------- End ----------------\n')
if is_GPU:
torch.cuda.set_device(args.gpu)
my_dataset = shapenet_dataset(args.data)
data_loader = torch.utils.data.DataLoader(my_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4,
collate_fn=pts_collate_seg)
net = pointnet2_seg(input_dim=3)
if is_GPU:
net = net.cuda()
optimizer = torch.optim.Adam(net.parameters(), lr=args.lr, betas=(0.9, 0.999))
critenrion = nn.NLLLoss()
def save_checkpoint(epoch, model, num_iter):
torch.save(