I'm trying to build a CNN model, I've about 44,000 images with 6 class, image size being 99X99. First i tried ResNet10 using SGD optimizer, i set it for 30 epochs with learning rate of 0.001 and the best model received was at 22 epoch, but accuracy for last 2 class was not that good, Then tried U-Net and added Autrous Spacial Pyramid Pooling as middle layer, and used ADAM optimizer and set it for 30 epochs with learning rate of 0.001, and got best model at 5 epoch and the accuracy is good, so what i wanna know is, is 5 epoch too soon for a model to converge, should i rely on this model?
Asked
Active
Viewed 1,285 times
1 Answers
0
Well, ADAM optimizer does tend to converge faster making training fast but it have it's own pitfalls https://shaoanlu.wordpress.com/2017/05/29/sgd-all-which-one-is-the-best-optimizer-dogs-vs-cats-toy-experiment/ look at this blog, here they describe SGD works better than ADAM