View Data
These images are labeled satellite image chips with atmospheric conditions and various classes of land cover/land use. Resulting algorithms will help the global community better understand where, how, and why deforestation happens all over the world - and ultimately how to respond.
data.show_batch(rows=3, figsize=(12,9))
Choose Learning Rate
learn.lr_find()
learn.recorder.plot()
There is a kind of rule of thumb that to choose a learning rate of the point that the loss goes down most sharply will make the result better. So I choose lr = 0.01
Training
learn.fit_one_cycle(5, slice(lr))
Results
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_multi_top_losses(9, figsize=(12,9))
These images are the top loss images, we can see that they have been mis-predicted.
I also learned some concepts of Neural Networks, such as gradient descent, auto gradient, loss function, parameter and activation, to understand how the training works. And also, I started to learn to use pytorch.