[Peter Herrmann]

Dear Youssef,

I ran your U-Net example.
But unfortunately nothing is happens on my computer.
Loss remains unchanged and the prediction image looks messy.

Is there anything else I need to consider?
Is the GPU on by default?

I use a Win10 workstation with Intel Xeon Gold processors, 384 GB RAM and a 48 GB Nvidia GPU

Best regards

Peter

Topic

[Youssef MENJOUR]

Dear Peter,

We are currently making a lot of changes, so we apologize for any inconvenience.

First of all we will verify if your CUDA and CuDNN and patch (we also patch windows for a cuda bug) are well installed.

To do this, we’ll have to start from scratch.

First you will have to download the latest version of HAIBAL plateform installer  you can found also the latest version on this page. Install it and if we are lucky just patch and our problem is solved :).

 

Unfortunately, this shortcut didn’t work, so we’ll have to make sure that the installation of your machine is properly done.

————————————— Step 1 ——————————————————————————————————

 

Uninstall the version of CUDA present on your machine.   I invite you to reboot your PC once this operation is done.

 

————————————— Step 2 ——————————————————————————————————

Now we are going to make sure that the directory where the CuDNN DLLs are located is also properly deleted.
So we have to delete this directory.

The folder is named :”NVIDIA GPU Computing Toolkit” (C:\Program Files)

————————————— Step 3 ——————————————————————————————————

Now you will launch the HAIBAL installer that you already installed

and logically it will detect only your device.

For me he has detected a Geforce RTX 3090.

————————————— Step 4 ——————————————————————————————————

So Launch the CUDA installation.

Once this installation is finished, restart the platform installer (a new version with an automatic refresh will come in the future – so for now restart the tools)

 

————————————— Step 5 ——————————————————————————————————

and patch !!!

When the operation is finish please restart your computer.

Let’s first do that and tell me if it works now.

 

• This reply was modified 1 year ago by Youssef MENJOUR.

[Peter Herrmann]

Dear Youssef,

I’ve done all the steps outlined.

Now it’s working.
Thanks!

Peter

[Youssef MENJOUR]

Dear Peter,

An example page was create for U-Net: Convolutional Networks for Biomedical Image Segmentation.

https://graiphic.io/examples/u-net-biomedical/

Be carefull,  this example is not optimized because the best practice to monitor your test is to” forward” the test image after a full batch – outside of the batch loop, we purposely did it like this to display the whole code on one page without subVI. Normaly you can divide by 2 the speed of train process 🙂 .

A fit functionality will be propose before march aimed to train efficiently any model with limited monitoring.

 

 

 

• This reply was modified 1 year ago by Youssef MENJOUR.

[Peter Herrmann]

Dear Youssef,

nice page.

The only thing that isn’t quite right is the HAIBAL architecture compared to the model design.
In your architecture, the number of convolution filters is constant (16). I call this the simple variant, which I also started with.

However, the model design starts with 32 conv filters and increases to 1024 filters at the bottom of the UNet.

[Youssef MENJOUR]

ha! I didn’t realize it !
the most important thing is that you can go ahead with your project 🙂

Do not hesitate Peter if you have any question.

[Peter Herrmann]

Thanks a lot Youssef!

[Author]

Replies