liuquande / ms-net Goto Github PK

Hi Quande,
Thanks for patient answering my previous questions. I think I am good with the dataset now, however, I am dealing with issues derived by incompatible TensorFlow version (I am using TF 2.2). May I know the library versions of this project? (Especially TensorFlow)

By the way, I am trying to make this project compatible with TF2 and get some trouble with the Batchnorm function. You are using a function which is deprecated by TF2.0+. Could you offer a possible solution using "tf.contrib.layers.batch_norm" to replace the "tf.contrib.layers.batch_norm"? or maybe you could tell me what the meaning of the “scope” parameter. I am new to TF, and the scope you are using with BN is confusing to me.

I used the clipped_nii data on the website https://liuquande.github.io/SAML/ to reproduce the result. (Now there is 'Processed_data_nii' on the website, are they the same? )
But I got bad results and the loss trend is strange as the picture shows. I'm debugging and have some questions.

1. Have the clipped_nii been propressed or it need to be propressed with some steps in the data_loader.py ?
2. I removed the data augmentation from the data_loader.py because it cause errors and you mention it in the issue5 that it's not necessary. And I set allow_soft_placement=True as I use only one gpu. I'm not sure if these matter. And is there anything else I should pay attention to?
   

I tried to reproduce your result with the data you mentioned in website https://liuquande.github.io/SAML/. It seems the data mentioned in the two papers are same partly.
I tried to train with Raw_nii and clipped_nii, but with either I got the errors such as 'valueError: could not broadcast input array from shape (384,0,3) into shape (384,384,3)' in data augmentation part. I tried to remove the data augmentation, but got results 'dice_avg_student 0.0574, asd_avg_student 93.4718, dice_avg_student 0.0391, asd_avg_student 94.6999, dice_avg_student 0.0256, asd_avg_student 106.5785'.
I splited the data to 60% for training, 20% for validation, and 20% for test.
I'm not sure if the data is not suitable for this experiment, or there are some other preprocessing steps I should do.
Wishing for your instruction.

Hi：
Thanks for sharing your code.I read your code,and I am puzzled about the knowledge transformer loss your artical proposed .It is seemd that you have not use this loss,the loss is just dice loss and CE loss.

Wish your reply.

Hi Quande,
Thanks for your help and I have finally set up the code and environment and successfully trained a model. Here I got some questions about your experiment settings since I cannot replicate your result.
The following figure is about the overall dice loss curves (train & val) during my training and it seems a little bit poor compared to your overall dice curves in the paper.

1. I am not sure if there is any setting needed to be manually changed such as the "cost_kwargs" and "opt_kwargs".
2. Are there any preprocessing steps besides a) resampling to 384x384 (I use the zoom2shape function in data_loader.py) b) the data augmentation(flip & shift) in your codes?
3. I center-cropped the data from Site-C with a size of 256x256. What size did you crop with?
   The blue line is around 0.2.
   

Besides, for the results in this table, do they represent the student dice score for each site?
Thanks.

Hello, I read the paper and your work is pretty good. Did you implement your code with Pytorch? I am new to Tensorflow and it's a little bit hard for me to read your code.
Thanks.

Where can I get the dataset?

您好，刘老师。
有几个问题想请教您一下，我在复现您的实验结果过程中遇到了以下问题：
①、ValueError: Tensor conversion requested dtype int32 for Tensor with dtype float32: 'Tensor("teacher_1/conv2d_transpose/output_shape:0", shape=(4,), dtype=float32, device=/device:GPU:0)'
②、TypeError: Input 'input_sizes' of 'Conv2DBackpropInput' Op has type float32 that does not match expected type of int32.
③、在构建网络过程中使用了 GPU:0 1 2，是不是需要拥有多块GPU的设备呢

Thanks a lot.