ultrasound-nerve-segmentation
ultrasound-nerve-segmentation copied to clipboard
jocicmarko
accuracy exceed 1.8000
@jocicmarko Thank you for sharing your implementation.
I'm using the model for training on 10000+ rgb images and the dice-coef exceeds 1.8000 (loss is obviously negative of that number). Shouldn't 1 be considered as the maximum value can be achieved for the accuracy? Appreciate your thoughts on that.
Same issue, my dice coefficient seems to exceed 1 from time to time. Seems to be related to the batch size
Before I investigate this further, can you provide me some details?
- Do you still use sigmoid as an activation function?
- Did you make any changes to outputs? I.e. more output classes instead of binary
- Did you make sure you scaled your ground truth values to [0,1]
Yes, sigmoid. Output is binary and the ground truth values are 0 and 1. THe only thing that I am doing differently is feeding the data. I am using a data generator that produces arrays of shape (batch_size,height, width,channels) and I am using TF backend:
model.fit_generator(DataGenerator(img_list,,batch_size) ,steps_per_epoch = 570,epochs = 50)
Did anyone solve this problem? When I apply this model for different dataset, same issue happened. In my case, I don't use data generator for implementing.
Most likely something is wrong with your input that you provide. I don't remember exactly what the issue was but I fixed it by fixing the input.
@luckustus I found easy careless mistakes which made mask input [0-255]. Now it works well. Thanks.
@akirakitamura19 is that a normalization layer that you are talking about?
@AloshkaD Yes, it is. Just make sure that values of train and train_make are [0 - 1].