AI Benchmark Accuracy/Error Questions

[yuchai84]

Hi,

I have been trying to understand the Accuracy / Error meaning from AI benchmark. Here is my questions:

1. For "err, L1", my understand is that it is Per-Label L1 loss. In the case of multiple images input, how shall I interpret the result? Is is as "sum of each image's Per-Label Loss / Image number"? Or is it as Per-Lable L1 loss among all the images' lables?

2. For "error, px", in UNET for example, what is the meaning for that?

 

[Andrey Ignatov]

Hi @yuchai84,

In the case of multiple images input, how shall I interpret the result? Is is as "v"? Or is it as Per-Lable L1 loss among all the images' lables?

sum of each image's Per-Label Loss / Image number

For "error, px", in UNET for example, what is the meaning for that?

Average per-pixel L1 error.