The test images (downloaded at the beginning of the lesson) are sometimes re-downloaded and used in the training set. I have suggested a simple fix (discard the training image if it is the same size as the [resized] test image). There are better ways to do it (and probably more elegant ways to write the fix).
A better implementation may be to pull out a few of the 600 downloaded images, and use these as test images - but I thought this didn't fit in with the flow of the lesson as well.
This is the only lesson I have done - and love it - thanks for all the hard work.
A suggested exercise for students could include - how does the resnet18 model predict without the fine tuning? (Answer is very poorly!)
The test images (downloaded at the beginning of the lesson) are sometimes re-downloaded and used in the training set. I have suggested a simple fix (discard the training image if it is the same size as the [resized] test image). There are better ways to do it (and probably more elegant ways to write the fix).
A better implementation may be to pull out a few of the 600 downloaded images, and use these as test images - but I thought this didn't fit in with the flow of the lesson as well.
This is the only lesson I have done - and love it - thanks for all the hard work.
A suggested exercise for students could include - how does the resnet18 model predict without the fine tuning? (Answer is very poorly!)