Densenet applied on multi-class unbalanced datasets I have 30000 image data with 5 class grading label (0,1,2,3,4). 
I use dynamic resampling technique. At beginning I resample each class to equal amount. Then I decrease the minor class(1,2,3,4). In the end, the ratio approach to 1:2:2:2:2 at 200 epoch.
I applied densenet169 with cross entropy loss on raw image:

I found most prediction value is 0.
Then I applied a color strengthen technique to preprocess the image. Then send to Densenet169:

I found class 3 got activated at some epoch in the end.
Then I merged these two datasets and another preprocessed datasets on channel. I got 512*512*9 input size. Here is the result from Densenet:

The results seem really unstable.
Any suggestions on how to future improve the results?
 A: Neural nets are very noisy; and empirical. If you think supervised image nets are bad, you can have a look at RL, in this blog post, https://www.alexirpan.com/2018/02/14/rl-hard.html shudder in horror, and then look back in relief at how relatively stable your supervised curves are.
Conv nets are not convex, and no-one knows why they converge so well at all. Actual convex things, like perhaps sum-product networks might be convex (I dont remember?) tend to empirically not work very well. So we're stuck with empirically working, theoretically how does it work?, noisy things.
Typically, one way to improve results on conv nets is to first train the network on some massive dataset, ie ImageNet. This will allow the network to learn some reasonable priors over images first. Then fine-tune on your own images. It'll still be noisy, but might, potentially, be mildly less noisy, possibly, depending on how well your own data matches imagenet data...
A: What are the units of axes? If it's the loss value, plot accuracy on a separate graph, as its values should not be > 1.
If validation accuracy is decreasing, it might a sign of the model overfitting to training data. Increasing regularization might help.
