model pretrained in BiaPy - version 3.6.8 GUI

[OlGiSa]

Good morning,

I wonder if you can help us with the proceeding of a pre-trained biaPy model built in "semantic segmentation" .

First, I generated the model with train and test TIFF images. That was succesful and generate good binary iamges (semantic segmentation).

Now, I would like to use that model ( I have both files YAML and PTH) in a different folder with new TIFF images, just to do semantic segmentation test. I dont want to do any training here.

In the Model Source I select " I have a model previously trained in BiaPy".

But the problem is that the proceeding asks for the training images. There is no way to skip this part. It allways asks for the training folder.

In the end, if I accept to train again, it does properly the the semantic segmentation in the new images; but I would not like to retrain.

Do you know if there is any issue with the GUI version? or maybe I am not doing properly the steps... I am following the wizard proceeding.

Thank you in advanced
kind regards
Olga

[OlGiSa]

Apologizes, I have realized I should have post this in the GUI section.

[danifranco]

Hello!

You need to select only "Test a model" like in the screenshot below:

You are probably selecting the first "Train a model" or the last option "Train and test a model".

PD: No worries about posting here on in the GUI project, always glad to help :D

[OlGiSa]

Dear Daniel

It doesnot work. It only works if I chose the train and test model. I 'd like to resum:

1.- I have a Train folder with X and Y folders. X for TIFF images and Y for corresponding masks groundtruths.
Also have Test folder with X folder. X with new TIFFs images.

If I generate a model with Train and Test option, the system generates the model and the corresponding masks TIFFs for the Test TIFFs. They are good. So I suppose that the train images and their corresponding masks need to be good for generating the model. Good so far.

2.- If I use the Test option, using this already done mask ( I say I have a biaPy model and select it ), and select the Test a model option. And select the TEST folder. It warns that the configuration file checked and some error were found

Model checkpoint not valid: path/myfile.pth Please correct the errors above before proceeding.

But it doesnot specify error. And this is the pth model I had before.

3.- I have tried a different thing: I've generated the model with the "train a model option", again built from scratch.

Exactly using the same pair of images.

It gives me this error message:

[14:44:02.435294] # LOAD TRAINING DATA #
[14:44:02.435303] ##########################
[14:44:02.435361] ### LOAD ###
[14:44:02.438408] Gathering raw images for training data . . .
Traceback (most recent call last):
File "/installations/BiaPy/main.py", line 4, in
main()
File "/installations/BiaPy/biapy/init.py", line 141, in main
_biapy.run_job()
File "/installations/BiaPy/biapy/_biapy.py", line 1135, in run_job
self.train()
File "/installations/BiaPy/biapy/_biapy.py", line 245, in train
self.workflow.train()
File "/installations/BiaPy/biapy/engine/base_workflow.py", line 847, in train
self.load_train_data()
File "/installations/BiaPy/biapy/engine/base_workflow.py", line 478, in load_train_data
) = load_and_prepare_train_data(
File "/installations/BiaPy/biapy/data/data_manipulation.py", line 388, in load_and_prepare_train_data
X_train = samples_from_image_list(
File "/installations/BiaPy/biapy/data/data_manipulation.py", line 1733, in samples_from_image_list
raise ValueError(
ValueError: Channel of the patch size given 1 does not correspond with the loaded image 3. Please, check the channels of the images!

Is this giving you any idea?
thanks
Olga

[danifranco]

Hi,

Regarding your 2nd point
I’ve identified and fixed a bug in the code. Please download the updated version of the GUI from the landing page and restart the process.

Regarding your 3rd point
You should be able to proceed using your previously trained model. However, if you'd like to troubleshoot the error: it appears there is a mismatch in image channels. The configuration file is set for 1-channel (grayscale) images, but the folder currently contains at least one RGB (3-channel) image.

This usually happens if the folder contents changed after the YAML file was first generated. To fix this, please regenerate the YAML file using the wizard to ensure all paths and image specifications are correctly synchronized.

[OlGiSa]

Dear Dani,

Apologizes for the delay in this answer .(We cannot install by our own anything; here it is mandatory to wait until someone from our IT team is free and come to do this :(

Anyway, the new executable has fixed the point 2. Thanks so much! So, now we can work with the researchers images.

As for the 3rd point, it still doesnot work. It refers that the RGB images are not alowed if I select the only Train model, but it admits the same RGB images if I select the Train and Test model.

It doesn't matter at the moment for us. We can work perfectly with the 2nd point already fixed. We are very happy and my microscopy team "te hacemos la ola".

Thank you so much
looking forward hearing from you

Vero and Olga

[danifranco]

I'm glad you can work with it!! For the 3rd case, maybe you are doing something wrong as it works as it should. If you want we can arrange a quick meeting so you can show me how you do it and so I can explain in case there are doubts. My email: daniel.franco@dipc.org ;)