When we:

  1. fit with a pre-trained model
  2. save
  3. unfreeze
  4. fit
  5. load what we saved in step 2,

are we overwriting the weights/activations on the final layers with the saved version? I mean, do we end up with:

  • weights on the final layers according to the fit from step 1; and
  • weights on the earlier layers according to the fit from step 4?

The weights we get at the end of step 5 is not the same as that we got at the end of step 1, correct?


(Wayde Gilliam) #2

If you’re loading what you saved in step 2 then anything beyond step 2 is irrelevant and doesn’t apply to your model because you’re reverting back to an earlier state of it. The weights in the final layers are the one’s you just finished training on, and the weights in the earlier layers are still untouched/the pretrained values.