poor result on isbi challenge about image_segmentation

[LeeJunHyun]

Oh, I implemented only for batch size 1.
At the first time, I had wanted to support more than batch size, but I didn't need that.
Sorry for confusion of code.
I'm not 100% sure, but I think I applied sigmoid function because of thresholding.
(To make binary segmentation results with threshold 0.5)
You can fix that part as you want :)

[LeeJunHyun]

Hi, @Lijiatu321 .
I am not familiar with ISBI challenge, but is it classification task?
(because you only mentioned accuracy metric)
U_Net architecture is for segmentation, not for classification.
Could you please explain about the task and the dataset?

[Lijiatu321]

Hi,@LeeJunHyun Thank you for you reply.I'm sorry I didn't make it clear before. isbi challenge is a segmantation task.In this challenge, a full stack of EM slices will be used to train machine learning algorithms for the purpose of automatic segmentation of neural structures .The original paper(U-net) use the isbi challenge data to train the network.Here is the linkhttp://brainiac2.mit.edu/isbi_challenge/. There are only 30 pictures in the training set and 30 picture in the testing set. The picture and the mask are all gray. Here is the example of the picture and the ground truth.

[LeeJunHyun]

Thank you for kind explanation.
How about result images? Did you check the results?

[Lijiatu321]

Hi, @LeeJunHyun . I have finally find the problem. There are two things.
One thing is that something wrong with the evaluation. Let me take get_accuracy() as an example, when you call the function you can calculate the accuracy of one batch size, but in the solver.train() when a epoch finished you do the acc = acc/length. And the length is the size of data(number of the train picture), in my opinion i think the length should be the length of the train loader. When we use the batch size 1 that is all right but if we change the batch size the accuracy will be decline. I am not sure if i am right . So, would you like to check you code and give me the answer?
The other thing is that the result of the model that i trained is not bad (accuracy : 90 percent), but i have some questions about the output of the network. When training you use the output of the network and the
gt as the input of the evaluate function but during validate you add the sigmod function. i don't understand why.

[Lijiatu321]

Thank you very much for your patient and time , wish you have a good day!!