Marissa Connor
I’m trying to understand how to best control which images from a data set are used for each step of training. I want to train two networks side by side on the same images. I want one of the networks to take several gradient steps using the same image and while the other network only takes a single step. Currently, if I want to take five steps in one network (with a batch size of 1), that uses five different images and creates a mismatch between the images in the two networks . Is there a way to specify that I want the same image to be used several times or to ensure that the order of the images in the two networks is the same? I see that there is a cursor in the data_layer.cpp that moves to the next image using cursor_ ->Next() and that’s where I was looking at making some changes but I wanted to check if others had any ideas for easy ways to control this.