This line
output = F.sigmoid(self.conv10(conv9))
cause
UserWarning: nn.functional.sigmoid is deprecated. Use torch.sigmoid instead. warnings.warn("nn.functional.sigmoid is deprecated. Use torch.sigmoid instead.")

It should be:
output = torch.sigmoid(self.conv10(conv9))

Hello, I'm using the same dataset for my final project. For 3 weeks I am trying to finetune my network to obtain something good but no result. I am using FNC8 with vgg encoder.
What I want to ask, if you have some time, I can give you my notebook wich run in colab, and if i can inspire from your work using your Unet arhitecture?
Sorry, it isnt an issue, but i write this, maybe I can talk to you somehow.

Can you show us a quick way how to customize the network to use larger input size?
I am willing to test it with 1024x1024 size.

Hi, have try to run your code with Brats Dataset? what is need to make it run with your code

Hi, I tried running your code but it hangs around the line below with `"unet TypeError: function takes exactly 1 argument (3 given)".
I have commented the section. Please kindly advise,

Training process

if TRAIN:
unet_model.train()
path = os.path.join('saved_models',MODEL_NAME) if SAVE_MODEL else None
unet_train_history = unet_classifier.train(EPOCHS,trainloader,mini_batch=100,save_best=path)
print(f'Training Finished after {EPOCHS} epoches')

Testing process on test data.

unet_model.eval()
unet_score = unet_classifier.test(testloader) #unet TypeError: function takes exactly 1 argument (3 given)
print(f'\n\nDice Score {unet_score}')

Dice Score 0.7446110107881675

`

May you consider to update torchsummary to torchinfo, thanks!

• torchsummary - https://github.com/sksq96/pytorch-summary
• torchinfo - https://github.com/TylerYep/torchinfo

Hello, Im trying to train on your netowork. With my training function, augmentation I only obtain 45% IOU accuracy with dice loss after 100 epochs, but learned very hard, small updates in accuracy. I try to use Crossentropy, but loss is constant, so I verify gradients and they 0, or almost 0. Do you have any ideas?

Hi,
the original code is really a nice implementation of u-net architecture for binary class segmentation.
May you consider to adapt this code for multiple class segmentation?
Thanks!