What exactly is "pos", "neg", and "total" on the verbose of CEBRA during training #216
-
|
A typical message CEBRA outputs during training with tqdm looks like this: What are those numbers on the left? Are they the losses for positive samples, negative samples, and their sum? Would a good model have them as small as possible? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
Hi @icarosadero , the "positive" and "negative" are more for technical/debugging purposes. The number that should converge during training is the total loss. Here is how the three numbers relate:
When interpreting the total loss, always keep in mind that the worst possible loss is log(batchsize), i.e. the magnitude scales considerably with the number of samples. You might find it interesting to look at #190 for more details on how to interpret the losses! |
Beta Was this translation helpful? Give feedback.
-
|
Ok, thanks for clarifying |
Beta Was this translation helpful? Give feedback.
Hi @icarosadero , the "positive" and "negative" are more for technical/debugging purposes.
The number that should converge during training is the total loss.
Here is how the three numbers relate:
When interpreting the total loss, always keep in mind that the worst possible loss is log(batchsize), i.e. the magnitude scales considerably with the number of samples. You might find it interesting to look at #190 for more details on how to interpret the losses!