Why does the sum of the confusion matrix numbers not equal the number of items in the validation set?

fastai
1

Hello all,

FastAI version: fastai==1.0.60

I noticed that when I plot the confusion matrix (on my validation set) and sum the numbers that it does not correspond to the number of items in my validation set.

Example:

The confusion matrix (array):


array([[50, 13],
       [ 0, 65]])

Adding the numbers above = 50 + 13 + 65 = 128

The number of items in the validation set:

learn.data.valid_ds

LabelList (140 items)
x: ImageList ...

I would expect these two values to be the same or am I missing something?

Thanks!

1 Like
2

Some fastai functions drop the last partial batch.

This is just a guess about what is happening - you would need to trace code to be sure.

1 Like
3

This seems plausible, as 128 is a typical batch size (or a multiple of a typical batch size such as 32 or 64).

2 Likes
4

@Pomo @machinethink
Thanks for your inputs! Apologies for the late reply. So yes, it does seem like the batch gets dropped. So all I did was to change my Batch Size to 70. So now my confusion matrix matches with my validation set size.

Thanks again!

1 Like