Given 177 pictures, the tasks is to analyze the metal element of it.
For each picture, we need to analyze the copper content(%), gold and silver content(ppm).
We try to use simple CNN structure to complete this task.
The original dataset is too small, so I sampling in dataset, and do some proc to img (rotate or flip).
Original value of label will cause gradient burst, so we need to regularize it.
We regularize the label by dividing label with scale [100, 1000, 1000]
Revise the effect of label's regularization:
(1) rc-net_mirror-scale.py
In this version we mutiply the regularization scale to the the front model (fit and train on regularized label)'s result.
Though the result seems normal and some predictions close to true label, the acc is quite low.
(2) rc-net_2steps.py
In this version we add 2 dense layers after the front model and fit it to true data.
2 dense layers: 2 dense layers perform well than 1 dense layer, and 3 dense layers will result in overfitting to some specific data.
The result is greater than rc-net_mirror-scale, acc is higher than it and the loss reduces.