Hi, I tried playing around with the notebook from lesson 1 as Jeremy said in course so I created mine.
I had some struggle but I want some feedback on what was right and what was wrong,… ? and Thank you
Github repo link
Hi @zowlex, this seems like a good start to me !
However, just after the interp.plot_top_losses
cell, you unfreeze your model then fit it for 1 epoch. If you look carefully to the original notebook, this step was there to show that it really hurts the model (Jeremy reloads previous weights just after this cell I if remember correctly). This is why your error_rate
goes to 0.07. So you really shouldn’t do this step without setting your learning rate accordingly
Ways I see to improve your model:
-
Run the learning rate finder at the beginning, instead of going with the default value. It will also give you more experience to choose correctly your learning rate.
-
Your images seem to be initially quite big. So maybe you can add a training step after what you have already done, where you use bigger sized training set (currently you use a size of 224x224, which can remove a lot of details). This can be done by creating another Databunch, with a size of, say 448, and then do
learn.data = your_new_data
Thank you so much for your time, I’ll make sure to work on the points you mentioned.