I have trained a model on some labeled images gathered from a public source. The set has echocardiogram scannings of the heart with classes of no, mild or severe disease. It was trained with a CNN network.
When I run predict(model, ..) or model%>%predict on an entire array of images, I get a decent distribution among the classes, but if I do on the same pictures one by one nearly all of them switches to "no", even in the rows where it was "severe" before.
Certainly I need to know what to do here?
