Why are the error rate accumulated?

[OleguerCanal]

❓ Questions & Help

Details

Each call of the error rate accumulates the distance and length. Why is that?Is it to have a running average kind of thing? Why don't you just return the point-wise wer? @upskyy

[upskyy]

@OleguerCanal We did so because it's good to understand the tendency when it's accumulated as a whole. Is there a reason why you want to do point-wise wer?

[OleguerCanal]

This makes sense but if using wandb or tensorboard it can already be smoothed like this right?

I'm saying because I was training an architecture with a ctc head and an attn head and wanted to compare the wers of each one. Since I used the same instance of wer_estimator, the values got mixed without me knowing

[sooftware]

Hi @OleguerCanal! Thank you for your good opinion!
I recycled the code that I made before, so I made it like this.
@upskyy As @OleguerCanal said, why don't we add a way to show the ER for each batch? Let's add this as an option.
--error_rate_logging: accumulate, batch like this?

[OleguerCanal]

I like this solution @sooftware ! :+1:

[sooftware]

I think batch would be good as the default value.