作者您好！
我猜测在issue种很多人无法复现您的结果是因为您在您的代码种用os.listdir获得目录然后再划分训练集和测试集，一旦更换主机，os.listdir得到的文件的顺序是不一样的，所以我们最终划分的数据集和您的数据集是不一样的。我分别尝试在两台主机上打印os.listdir获得的目录顺序，发现其不一致。

主机一：

主机二：

我建议您对os.listdir后的all_dirs进行排序后再划分数据集！

I guess that many people in the issue cannot reproduce your results because you use os.listdir in your code to obtain the dirs and then divide the training set and test set. Once the host is changed, the order of the files obtained by os.listdir is not the same. The same, so the data set we finally divided is different from your data set. I tried to print the directory order obtained by os.listdir on the two hosts and found that it was inconsistent. @ge-xing

Hi ge-xing:
When I run test.py after training the model with train.py, I get the following error

How can I solve this problem

The loss jitter is large when training your own dataset, how to solve this problem?

Hi, thanks for sharing the code base. I try to reproduce the results on BraTS 2020 dataset, but the results I got are much worse than the paper. Here are the details:

For model training:
wt is 0.8498, tc is 0.4873, et is 0.4150, mean_dice is 0.5840

The tensorboard files are:

The final model files are:

My settings are default settings:
env = "DDP"
max_epoch = 300
batch_size = 2
num_gpus = 4
GPU type: A100

Then I use the best model (best_model_0.5975.pt) to do evaluation on the test set, and I got:

My python environment is:
Python 3.8.10
monai 1.1.0
numpy 1.22.2
SimpleITK 2.2.1
torch 1.13.0a0+936e930

The most strange thing is the segmentation performance of TC and ET is quite bad. Do you have any idea why the performance is so weird, and could you give me some advice on model training? BTW, could you please share the conda env file and your model weights for BraTS 2020 dataset? If you can create and share a docker image I think that could be perfect! Thanks.

self.embed_model = BasicUNetEncoder(3, 1, 2, [64, 64, 128, 256, 512, 64])
self.model = BasicUNetDe(3, 14, 13, [64, 64, 128, 256, 512, 64]

why the embed_model not be BasicUNetEncoder(3, 1, 13, [64, 64, 128, 256, 512, 64])
2.
why use the resample_img(), and if I run on my own dataset, in which case I should use it.

我在使用AbdomenCT-1K這包data遇到上述的問題，不知道你們有沒有解法，我針對MONAI版本去做調整目前都沒用。

Hello, your work is so great!!
i ve already got the results on brats2020,now iam work on my own datasets,but my dataset only have 1 modality （t1），i want to know where can i change the input channel from 4 to 1 ,and what else do i need to change,thx!!!!

Thanks for your great work but I have met a little problem:

Why is the implementation in this line different from the formula in the paper?

Should I change sample_outputs[i]["all_samples"][index].cpu() to uncer_out in

Hi. Thank you for your contribution.
Could you please share the environment file of your experiment? I met several re-implementation errors, which I think may due to different environment and packages.
Looking forward to your reply.

Thanks for releasing the paper and the source code!

I experimented with GPU=2 since I only have 2 GPUs.
I experimented with batch_size=4 and found that the mean_dice during training was abnormally low compared to other people's results.

Results of the experiment with batch_size=4
training mean_dice = 0.5956
wt = 0.9142
tc = 0.6453
et = 0.6193

Are there any parameters or codes that should be changed by reducing the number of GPUs?

Hello, I tested the model after training, but I encountered a error stating that "maximum recursion depth exceeded while calling a Python object".


Could you help me with the problem? Thanks a lot!

您好，能提供一下您训练好的模型吗？我的邮箱是[email protected],万分感谢

I had an issue installing "mpi4py", is it possible to avoid using this library or is there an alternative?

Thank!

Hi,

I notice that the optimizer is defined only for the denoising unet model here

When I executed the train.py, I encountered the following error, how can I solve it?

Traceback (most recent call last):
File "train.py", line 192, in
trainer.train(train_dataset=train_ds, val_dataset=val_ds)
File "/home/hdc/zjh/Diff-UNet-main/LiTS/light_training/trainer.py", line 262, in train
self.train_epoch(
File "/home/hdc/zjh/Diff-UNet-main/LiTS/light_training/trainer.py", line 361, in train_epoch
for idx, batch in enumerate(loader):
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 517, in next
data = self._next_data()
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1182, in _next_data
idx, data = self._get_data()
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1148, in _get_data
success, data = self._try_get_data()
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 986, in _try_get_data
data = self._data_queue.get(timeout=timeout)
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/multiprocessing/queues.py", line 116, in get
return _ForkingPickler.loads(res)
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/site-packages/torch/multiprocessing/reductions.py", line 282, in rebuild_storage_fd
fd = df.detach()
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/multiprocessing/resource_sharer.py", line 58, in detach
return reduction.recv_handle(conn)
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/multiprocessing/reduction.py", line 189, in recv_handle
return recvfds(s, 1)[0]
File "/home/hdc/anaconda3/envs/pytorch/lib/python3.8/multiprocessing/reduction.py", line 164, in recvfds
raise RuntimeError('received %d items of ancdata' %

Hello, I used BTCV code to run on the MSD Task-08Hepatic Vessel dataset and LITS dataset, but the Dice coefficients I obtained were both very low. What is the reason for this? If it's convenient, can you give me a BTCV model that you have trained? Thank you very much! My email is [email protected] Thank you very much!

作者您好,我没有太好的GPU进行训练， 能提供一下您已经训练好的模型吗？邮箱：[email protected]，万分感谢！

I trained on the LiTS data set, and the Dice Score on the training set can reach 0.9035, but only 0.6898 on the test set. When I visualized the results, I found an interesting thing. The predicted shapes are very similar, but there seems to be some Offset, where might the problem be?

BTCV test.py script's code below at line 151. The function get_loader_btcv's arguments set batch_size, but this function parameter does not have batch_size. Here is the function parameter get_loader_btcv(data_dir, cache=True). I assume data_dir=./RawData/Training because get_loader_btcv read image and label files and RawData/Testing has only img directory. Am I setting the function arguments correct? or Am I using a wrong dataset? Please let me know.

train_ds, val_ds, test_ds = get_loader_btcv(batch_size=batch_size, fold=0, cache=False)

First of all, thanks for publicly sharing your paper and source-code.

I have used train.py to train a model on BraTS2020 dataset, but the results I got mismatches the results reported in the paper, do you have any idea why is that?

I got the following Dice scores (after training for 300 epochs):
wt is 0.8935, tc is 0.7762, et is 0.7637, mean_dice is 0.8111

From your figure 1 in the paper, we know your method directly predict x_0 in training, but why you inference the result step by step in testing? I do not understand.

def training_step(self, batch):
    image, label = self.get_input(batch)
    x_start = label

    x_start = (x_start) * 2 - 1
    x_t, t, noise = self.model(x=x_start, pred_type="q_sample")
    pred_xstart, pred_y = self.model(x=x_t, step=t, image=image, pred_type="denoise")

    loss_dice = self.dice_loss(pred_xstart, label)
    loss_bce = self.bce(pred_xstart, label)

    pred_xstart = torch.sigmoid(pred_xstart)
    loss_mse = self.mse(pred_xstart, label)

    loss = loss_dice + loss_bce + loss_mse

In the training phase, the noise generated from q_sample, the next step of prediction is the loss calculated with label. In my understanding, xt should calculate the loss with xt-1, and then generate x0 for the final target detection. But here why not calculate the loss with the next step, but directly calculate with the standard label?

大佬你好，该模型能否在AMOS22上运行，有没有考虑将Mamba与扩散模型结合

Hi, thank you for your contributions.

I'm confused by the window_infer, which seems to be a MONAI function and is unable to handel the parameter pred_type="ddim_sample .

Hello, I used BTCV code to run on the MSD Task-08Hepatic Vessel dataset and LITS dataset, but the Dice coefficients I obtained were both very low. What is the reason for this? If it's convenient, can you give me a BTCV model that you have trained? Thank you very much! My email is [email protected] Thank you very much!

According to your code, the diffusion model is not a faster version with large steps. How long the training time you cost?

May I ask a question, can the model do multi-class segmentation？

Hello, how is the BTCV training set divided, and can you directly use the original dataset without data preprocessing?
I use the list of BTCV training datasets below and get an error.

Hello there, author! Can you send me the segmentation model on brats2020 dataset? My email address is [email protected]. Thank you very much!

def training_step(self, batch):
        image, label = self.get_input(batch)
        x_start = label

        x_start = (x_start) * 2 - 1

why x_start = （x_start）*2 -1 ？

Can I use x_start directly, instead of two times and minus1？

Could you please share your requirements.txt file, thank you!

When I apply this model to the verse2020 dataset, I get stuck at the ninth epoch every time (it will directly terminate the prompt RuntimeError: DataLoader worker (pid 9063) is killed by signal: killed) When I change the higher performance GPU and CPU, adjust the learning rate and batch, etc., I still get stuck at the ninth epoch, showing that it takes ten hours

Can you share the code of MSD Liver
I want to reproduce the experimental results

I always get an error below when I set device=device = "cuda:0"

File "test.py", line 170, in
v_mean, v_out = trainer.validation_single_gpu(val_dataset=test_ds)

Diff-UNet/BTCV/light_training/trainer.py", line 168, in validation_single_gpu
val_out = self.validation_step(batch)
File "test.py", line 112, in validation_step
output = self.window_infer(image, self.model, pred_type="ddim_sample")
File "/opt/monai/monai/inferers/inferer.py", line 521, in call
return sliding_window_inference(
File "/opt/monai/monai/inferers/utils.py", line 256, in sliding_window_inference
seg_prob_out = predictor(win_data, *args, **kwargs) # batched patch
File "/usr/local/lib/python3.8/dist-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "test.py", line 72, in forward
sample_return += sample.cpu()
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

Program file "BTCV/text.py " a line of code at the line number 100 below. This line of code should be batch["label"] instead of batch["raw_label"]. Because batch["raw_label"] does not match with input's volume sizes.

label = batch["raw_label"] --> label = batch["label"]

Thank you a lot for your great work. I'd like to visualize my Brats testing segmentation results, which parts of code should I modify ?