Training

Example:

learner = cnn_learner(databunch, pretrained_model, metrics=accuracy)
learn.fit_one_cycle(1,1e-2)

Optimum Learning Rate

Refer to this link

What is Learning Rate?

• Hyper parameter that decides how much gradient should be back propogated

  • i.e. How much we move towards minimum

  • Small learning rate → Converge slowly to minimum

  • Large learning rate → Diverges

learn.lr_find()
learn.recorder.plot(suggestion=True) # Draws where minimum numerical gradient is (which should be your learning rate)

• Choose value that is in the middle of the sharpest downward slope

The 1cycle Policy

Refer to this link

• Similarly to the regular learner.fit, we need to find the optimum learning rate using lr_finder.