Training for CSBDeep plugin

I’m confused on how to train the CSBDeep plugin, which looks very promising and I would love to use. How do we generate the trained model files to implement on our image data? It looks like I cannot generate the models in the plugin itself. Do I have to use TensorFlow to generate the trained model files? Are there plans to include training steps in the CSBDeep plugin itself? Thank you very much for your consideration.


I am also waiting for training step on CSBDeep.

I have understood this plugin can read a trained model which is generated by ourself on TensorFlow.
*I checked this plugin, but I have not checked the running with my model.(I can not use TensorFlow,)


Thank you for your response. I think this plugin would be heavily used if it were made friendlier for those of us that don’t know how to code. I hope there are plans to develop model training if future versions of CSBDeep. I presume .model files generated in Trainable Weka Segmentation cannot be used in CSBDeep?

Hi, thanks for posting, I’m happy you show interest in the plugin!

You are right, training is not possible in Fiji at the moment. It has to be done e.g. in python, the CARE networks are one example of matching image to image networks (installation, example notebooks). The notebooks use Keras and at the end of these notebooks the trained model is exported as a TensorFlow SavedModelBundle. This is what you can apply to your data using the CSBDeep Fiji plugin (provided as ZIP file or URL).

We are discussing training strategies. There is no roadmap yet.

About the Trainable Weka Segmentation: I don’t know how to convert such a model to TensorFlow. I will try to get to know more about it. It would be indeed nice to connect the two plugins.


Thank you for your prompt response. If I understand correctly, the image models are made in Keras, then exported in TensorFlow, and then processed/restored in CSBDeep?

Another trainable plugin is Trainable Deep Segmentation, which is in Plugins\Segmentation\Development\Trainable Deep Segmentation. If there were a way to link this plugin with CSBDeep, that would be great! I look forward to the future developments in CSBDeep! I have some I intravital microscopy data I would to have restored! Thank you.

1 Like

The models of the Trainable Weka Segmentation (TWS) plugin are regular WEKA models not TensorFlow ones. May I ask why you need TWS models imported in CSBDeep?

I was merely speculating as to whether a deep learning plugin in Fiji could be implemented in training for the CARE plugin. My image processing skills aren’t that advanced and I don’t understand how to use Python or TensorFlow to generate model files for CARE to implement. It frustrates me when plugins that could help our core facility users tremendously aren’t accessible to our beginning and intermediate users of Fiji/ImageJ.

1 Like