How to use Synapse dataset?
Opened this issue · 2 comments
Hi!Your team have made a amazing work!But I got some troubles about it.
I saw you used Synapse dataset in your paper,but i don't find the usage about Synapse dataset in your github.
And then I try to reproduce the code and make some modification on my own way for Synapse dataset,but i got a different result and some bug:
1.When i used --max_iterations 10000,i got
Traceback (most recent call last):
File "train_fully_supervised_2D_VIM.py", line 257, in
train(args, snapshot_path)
File "train_fully_supervised_2D_VIM.py", line 186, in train
metric_i = test_single_volume(
File "/home/zh_701/new_sda_2T/LHJ/Mamba-UNet/code/val_2D.py", line 39, in test_single_volume
metric_list.append(calculate_metric_percase(
File "/home/zh_701/new_sda_2T/LHJ/Mamba-UNet/code/val_2D.py", line 14, in calculate_metric_percase
hd95 = metric.binary.hd95(pred, gt)
File "/home/zh_701/anaconda3/envs/Munet/lib/python3.8/site-packages/medpy/metric/binary.py", line 413, in hd95
hd1 = __surface_distances(result, reference, voxelspacing, connectivity)
File "/home/zh_701/anaconda3/envs/Munet/lib/python3.8/site-packages/medpy/metric/binary.py", line 1269, in __surface_distances
raise RuntimeError(
RuntimeError: The second supplied array does not contain any binary object.
2.When i used --max_iterations 1000,i got:
mean_dice : 0.407416 mean_hd95 : 99.048800
I coudn't reproduce the result,I don't how do fix it .
I will be appreciate it if your could help me .
Hi, for Synapse datasets (this is somehow different from ACDC and prostate), please import the dataset:
from dataloaders.dataset import BaseDataSets_Synapse
and then replace the orginal dataset.
and the bugs indicates the evaluation metrics error. This is not a significant problems, and I guess you could just potentially not use the HD95 metrics, that would be fine. because ths loss is based on dice, not HD95.
hi, have you tried to train a model with 10,000 iterations? or 30,000 iterations?
Thanks.
Hi, for Synapse datasets (this is somehow different from ACDC and prostate), please import the dataset:
from dataloaders.dataset import BaseDataSets_Synapse
and then replace the orginal dataset.
and the bugs indicates the evaluation metrics error. This is not a significant problems, and I guess you could just potentially not use the HD95 metrics, that would be fine. because ths loss is based on dice, not HD95.
hi, have you tried to train a model with 10,000 iterations? or 30,000 iterations?
Thanks.
It did works!But I got a new trouble when i run test_2D_fully.py
I ran as :
python test_2D_fully.py --root_path ../data/Synapse --exp Synapse/VIM --model mambaunet --num_classes 9 --labeled_num 8
And then got a bug:
test_2D_fully.py:12: DeprecationWarning: Please use zoom from the scipy.ndimage namespace, the scipy.ndimage.interpolation namespace is deprecated.
from scipy.ndimage.interpolation import zoom
=> merge config from ../code/configs/swin_tiny_patch4_window7_224_lite.yaml
None
../model/Synapse/VIM_8_labeled/mambaunet/mambaunet_best_model.pth
Traceback (most recent call last):
File "test_2D_fully.py", line 125, in
metric = Inference(FLAGS)
File "test_2D_fully.py", line 105, in Inference
net.load_state_dict(torch.load(save_mode_path))
AttributeError: 'NoneType' object has no attribute 'load_state_dict'
Notice net is a None type.I trace back in networks.net_factory ,and found:
def net_factory(net_type="unet", in_chns=1, class_num=4):
if net_type == "unet":
net = UNet(in_chns=in_chns, class_num=class_num).cuda()
elif net_type == "enet":
net = ENet(in_channels=in_chns, num_classes=class_num).cuda()
elif net_type == "unet_ds":
net = UNet_DS(in_chns=in_chns, class_num=class_num).cuda()
elif net_type == "unet_cct":
net = UNet_CCT(in_chns=in_chns, class_num=class_num).cuda()
elif net_type == "unet_urpc":
net = UNet_URPC(in_chns=in_chns, class_num=class_num).cuda()
elif net_type == "efficient_unet":
net = Effi_UNet('efficientnet-b3', encoder_weights='imagenet',
in_channels=in_chns, classes=class_num).cuda()
elif net_type == "ViT_Seg" :#or "mambaunet"
net = ViT_seg(config, img_size=args.patch_size,
num_classes=args.num_classes).cuda()
elif net_type == "pnet":
net = PNet2D(in_chns, class_num, 64, [1, 2, 4, 8, 16]).cuda()
elif net_type == "nnUNet":
net = initialize_network(num_classes=class_num).cuda()
else:
net = None
return net
It mean that model" mambaunet" actually has no net
I dont konw how to solve it .
Thank a lot for your reply!And could you please help me for the new problem?Thanks again!