I have trained this code multiple times.but the highest value of the dice coefficient i have got on the test dataset is 0.54.Could you please give me some advices?

[haibo-yu-yy]

I have trained this code multiple times.But the highest value of the dice coefficient i have got on the test dataset is 0.54.Could you please give me some advices?

[yelusaleng]

@haibo-yu-yy u should say some detailed information about your training

[haibo-yu-yy]

Thank you for your reply.Here are the details of my training: Dataset: CASIA 2.0, total 1825 splicing images.
train dataset : test dataset = 8:2, no validation dataset, i test once per epoch on test dataset. use data augmentation on the train dataset ,then there are 8760 images for training. batch_size=14 lr = 0.01 in epoch 0~49, lr = 0.001 in epoch 50~99, lr = 0.0001 in epoch 100~149. Besides, i also trained my model with no data augmentation train dataset, but i got a worse value. Look forward to your reply!

[yelusaleng]

I used the size 256384 (HW) of the image, what is yours?

[haibo-yu-yy]

256 * 256 I rewrote a file named dataset.py to deal with images.

[yelusaleng]

u can try the size I used.

[haibo-yu-yy]

ok, i will have a try

[SwallowIsXYZ]

I have trained this code multiple times.But the highest value of the dice coefficient i have got on the test dataset is 0.54.Could you please give me some advices?

could you give me your trained code , I can not run the code the author give above ,there is always a wrong , I have done all the thing I could done , it include resize the image size to the 384 X 256 . but I can not solve it , It made me so upset . I saw that you have trained the code ,can you give me your changed code or can you help me solve the problem? my email 1483367209@qq.com I am very appreciate it if you can help me

true_masks = np.array([i[1] for i in b]).astype(np.float32) / 255. ValueError: could not broadcast input array from shape (256,256,3) into shape (256,256)

[haibo-yu-yy]

u can check the shape of the images and masks. input images should have 3 channels true_masks have only 1 channel

[SwallowIsXYZ]

------------------ 原始邮件 ------------------ 发件人: "yelusaleng/RRU-Net" <notifications@github.com>; 发送时间: 2020年12月14日(星期一) 下午5:25 收件人: "yelusaleng/RRU-Net"<RRU-Net@noreply.github.com>; 抄送: "燕子"<1483367209@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [yelusaleng/RRU-Net] I have trained this code multiple times.but the highest value of the dice coefficient i have got on the test dataset is 0.54.Could you please give me some advices? (#5)

u can check the shape of the images and masks. input images should have 3 channels true_masks have only 1 channel

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Thank you for your quick reply ,I will try again follow your advice

[haibo-yu-yy]

@yelusaleng
Could you provide a pre-trained model?

[yelusaleng]

@haibo-yu-yy i have uploaded the model

[haibo-yu-yy]

thank you!