About dice loss about tf_unet HOT 6 CLOSED

Not sure if my thought were correct back then, so any feedback is welcome.
The dice coefficient reaches 1 as best value (full aggreement). Hence, we want to reduce the difference between the current score and the best achiveable value:
lossFunction = 1 - currentScore

from tf_unet.

I agree with your point, I am convinced till dice coefficient calculation , but when batch contains two volumes one is label and one is image then what I should do ? Take mean or sum then subtract from 1.
currently I am getting following plot so i am little bit confused

from tf_unet.

This looks indeed a bit odd. (btw, you could use the Tensorboard so that you don't have to plot the metrics by yourself).

when batch contains two volumes one is label and one is image

not quite sure what you mean. Could you clarify a bit? What do you mean exaclty with batch and volume?

from tf_unet.

Ohh sorry for confusion I am using caffe and in caffe you pass data into batch ,volume is just what is size of your layer for ex (124 x 64 x 64) . But I am using dice loss layer, the code I am referring they are using

loss =np.sum(2*intersection/(union))

where as in your case its

loss = 1 - tf.reduce_mean(2 * intersection/ (union))

here is conflict, because the loss function is same but different approach to get loss value but with sum I am not getting desired results. Hope you got me now

from tf_unet.

I just pushed a new branch where I changed the computation of the dice coefficient.
I'm also using the predictions instead of the logits as suggested in #15 . I tested it with the toy example of tf_unet.

from tf_unet.

I merged the branch into master. Hope this resolves the issue

from tf_unet.