aryansethi / no-blur Goto Github PK

Hello, again. I've noticed your progress through the past few days, and I thought I could help. I've noticed your choice of loss function and last layer activation are binary_crossentropy and sigmoid respectively. however this combination is used for binary classification tasks, which is not what you're doing. your output is not a probability distribution nor are you solving a classification problem. a more appropriate loss function would be mean squared error or mse for short since your outputs are linear values
(0 - 255 pixel data), now that I've mentioned linearity, it's time to say that the last layer's activation should be linear (aka no activation at the end of the network). Now to evaluate your model, you need to look at the loss, not the accuracy. accuracy is only meaningful in classification problems. I would encourage you to read some other open source projects that do the same task and learn from them, it will be very beneficial. best of luck to you.