Lesson2 --dogbreed accuracy not improving at all

Part 1 (2018)
1

Hi All,

I am following lesson 2 and working on dogbreed dataset. I hve used all the mentioned techniques like:
1)LR =.005, it gives me 83.57%

learn.fit(lr, 2)

poch trn_loss val_loss accuracy
0 0.647937 0.553398 0.826321
1 0.62565 0.532092 0.835127

Then with augmentation = False and learn.fit(lr, 3, cycle_len=1, cycle_mult=2)
epoch trn_loss val_loss accuracy
0 0.640235 0.512987 0.840998
1 0.639211 0.509754 0.84002
2 0.599696 0.510067 0.84638
3 0.622977 0.508784 0.842955
4 0.522688 0.491936 0.84638
5 0.544614 0.498515 0.84002
6 0.601776 0.494757 0.844912

after unfreezing and using

lr=np.array([1e-4,1e-3,1e-2])
learn.fit(lr, 3, cycle_len=2, cycle_mult=2)

epoch trn_loss val_loss accuracy
0 0.542071 0.495063 0.840509
1 0.560871 0.495163 0.847358
2 0.530614 0.496824 0.840509

training loss is improved but still validation loss is high. It is still underfit… What should i try now??
Just disclaimer i am still using sz = 224 and bs = 16

2
  1. At the very first stage, you show that you fit for 2 epochs. Your validation loss was still going down. Why did you stop fitting?
  2. In my experience, the learning rate plot suggested that a much higher learning rate than 0.005 would be useful. I ended up using 0.1, for example. Here is what mine looked like:

image

3

i thought in case of under fit scenario, we need to add SGR and if not improving than add further cycle_mult