Hi,
I think my problem related more to basic concepts of deep learning than to your specific project, but I'm not really sure.
My problem is that after several experiments I realized that unfortunately, the model didn't succeed to get a higher value from 4.9 (for the dice soft). I tried the model on 15 samples, but I doubt that this is the reason for the insufficient result. After all, the amount of the samples usually affects the test set but not the train set. I'm hesitating that I got wrong when I chose the value of the iterations and the ephocs or something like that... (Although the weights updated every batch/iteration, so 200 ephocs and 100 iterations spouse to be equal to 100 ephocs and 200 iterations).
Note, that when I used the nifty slicer interface on my data, I got much more satisfactory results.
Another interesting point. I used your trained model which you shared with me in the past, and it seems that the loss value starts from a worse position than a new training. I also used the "full image" analysis at this time. I assume that it can be also affected by a wrong choice of patch shape values or something like that...
I attached here a link to google colab. I can also share my data if it's necessary.
Hi, I think my problem related more to basic concepts of deep learning than to your specific project, but I'm not really sure.
My problem is that after several experiments I realized that unfortunately, the model didn't succeed to get a higher value from 4.9 (for the dice soft). I tried the model on 15 samples, but I doubt that this is the reason for the insufficient result. After all, the amount of the samples usually affects the test set but not the train set. I'm hesitating that I got wrong when I chose the value of the iterations and the ephocs or something like that... (Although the weights updated every batch/iteration, so 200 ephocs and 100 iterations spouse to be equal to 100 ephocs and 200 iterations).
Note, that when I used the nifty slicer interface on my data, I got much more satisfactory results.
Another interesting point. I used your trained model which you shared with me in the past, and it seems that the loss value starts from a worse position than a new training. I also used the "full image" analysis at this time. I assume that it can be also affected by a wrong choice of patch shape values or something like that...
I attached here a link to google colab. I can also share my data if it's necessary.
https://colab.research.google.com/drive/1TjthgsjZwQZaOqy6DJqE1JSrpatEV2Xk?usp=sharing