Trained model sharing and questions on test module

[JunMa11]

Dear MIC-DKFZ,

Thanks for the great respo. I'm a pytorch starter, and I'm using this U-Net example as a beginning. Enviroment: Ubuntu 16.0, Titan XP, pytorch 1.0, batchgenerators 0.19

• Following the guideline, the 2D training process works well. The trained model can be found here.

• The 3D training process also works but we need to modify the loggers in train3D.py as 2D case. Trained model of 3D U-Net

loggers={ "visdom": ("visdom", {"auto_start": c.start_visdom})}

• batchgenerators 0.19 released recently. I use the new verson and it works well. pytorch 1.0 also works well. So the requirements.txt may be updated.

The test module in experiments does not implement. It is a litter difficult for pytorch beginners to finish it. Would it be possible for you to complete the test module in U-Net 2D and 3D at your convenience? So this respo will be a complete U-Net example.

I think it is worth to enrich this great respo and make it to be a good begining/tutorial for pytorch starters in medical image segmentation, becuase I do not find such starter-friendly respo in github. Most of the U-Net respos are for CV rather than for medical image segmentation. Although I can not contribute the valuable code now, I can do the code testing work and share the trained model.

Looking forward to your reply! Best, Jun

[elpequeno]

Hi @JunMa11,

thank you for sharing your experience with our repo. I think I can provide an example "test()" method in the future. I'm on vacation for the next 2 weeks, but maybe after that. Thank you for testing the newest versions of the imported frameworks. I'll update the requirements.txt accordingly.

I'll contact you again in about 2 weeks. Cheers, André

[JunMa11]

Hi @elpequeno , Thanks for your quick reply and approval. I wish you a very good journey.

[elpequeno]

Thank you :-D You are welcome. If you want to practice a little since you say you are a beginner: The test() method will look similar to the validation. You need to load your test data, iterate over them and calculate the metrics you are interested in. For a start, I would say DICE, IOU, precision, and recall. Of cause, there are many more, but that sounds like a good start for me.

[JunMa11]

Thanks for your guidance. I will try it.

[elpequeno]

I added a test method to the 2D UNetExperiment. Results are printed in the terminal and also stored in a JSON file. Check it out and tell me if this helps. I'll close this task.

[JunMa11]

Get it. Thank you very much.