I will be holding office hours today at 5:30pm in our classroom at USF.
Hi @yinterian,
Had a doubt regarding last weeks lesson.
I called the learn.predict() function without calling the learn.fit() at all, and it still correctly classified most of the outputs. Wondering how that was possible?
Is it because this line that created the learn object already worked on the training + validation set?
learn = ConvLearner.pretrained(resnet34, data, precompute=True)
Image showing the correct outputs.
I don’t think so. Not sure why you got that.
That’s odd. Can others replicate this? If so, it would be interesting to look in to…
In crestle I deleted the tmp and all previous model files, commented the learn.fit(0.01, 1) in notebook and this is the prediction
Sure, it’ll get half correct from random chance - but it shouldn’t be getting more than half correct. You could check the confusion matrix, for instance.
The image you show doesn’t actually demonstrate that. You’ve only shown what happens when you specifically plot those that are correct - but it doesn’t show what % of the dataset are correct.
1 Like
Yes… I wrongly assumed that by looking at the correctly classified images.
So, ran the confusion matrix on a model that does not call learn.fit and got bad results as expected.
1 Like
Can i conclude that we should always plot confusion matrix after prediction to check if model is working fine or not.
Is Confusion matrix the only way to check this or there are more ways?
Just the accuracy printed out after every epoch is fine for this.