wrong detection for more images

[ghost]

Hello Dear, before all thing, I want to say your codes are so usefull and many thanks. I implemented everthing as you said. However, when I implement last code (6-WindowsPolypsDetection.ipynb: ) for detection of images from cvc-clinicdb sometimes it gets fine, and sometimes box is in wrong place (not detect polips). I checked there was no overfitting in my network. What is wrong with this ? Many thanks.

[muntisa]

Thanks for the note! The localization of the polyps is just a simple and brute one: window sliding. This means that it depends on the step and it is using a lot of predictions for each step. Better methods could be implemented. About the polyp classification: I have no other tests for this classification, so I should check other metrics, optimize more the model, etc. Fastai version of the same task has better accuracy: https://github.com/muntisa/Fastai-Colon-Polyps.

[ghost]

Thanks for quick reply. I'll check fastai-colon-polyps. Your previous work (colon-detection) was very usefull, if you check and test for other images please let me know. last thing, in your fine tunnig code, there is no code cell to save the model. I tried to save but always get errors such as model is not defined (also delete the lines for delete top model etc). If you edit this one it will be perfect. thanks again.

[phamchien98]

Can you share checkpoint "saved_models/model_best_Conv-Conv-Conv-FC_full.h5" for me for debugging? I try to run code in colab but can not train model by myself.