Accuracy(log_preds, y), where's the y from?

Part 1 (2019)
1

Hi, I was testing the accuracy of my algo using:
accuracy(log_preds, y)

but I couldn’t figure how it can calculate the accuracy if I don’t have the ground truths of the test set. Where are they y’s from?

I’m using TTA.

2

For the test set, you cannot calculate accuracy if you don’t have any labels.

3

When looking at the source code, you can see that when you call learn.TTA(), it returns two things:

return np.stack(preds1+preds2), targs

It returns your TTA predictions but also your targets (i.e. y). So you can simply do preds, y = learn.TTA() and then pass those to your accuracy function.

4

That’s the odd thing, nowhere in my code am I saying: “here are the test set labels.”

However TTA is still able to test and I’m able to get an accuracy %.

5

When we add a test set in fastai, fastai creates an emptylabellist for it.

if label is None: 
   labels = EmptyLabelList([0] * len(items))

Now by default as you see it is labeling with index=0, so due to this you can get accuracy values.

So if you run it for CIFAR10, then all the labels in the test set would be aeroplane.

7

How can I set my own y’s?

8

Create a new databunch and predict on it.