Open ainazHjm opened 5 years ago
I tried the model without the rotation part and without using any weights for positive samples (i). The loss function is decreasing nicely but I guess the output will be zeros every where because the ratio of the positive samples are ~0.001. I'm going to train another model with weights and see how well that would do. I also need to train another model with oversampling of positive images and more data (stride < 200). Here is the loss function for (i):