Add penalization term in loss funtion

SergioQuijanoRey / TFG

Trabajo Fin de Grado

1 stars 0 forks source link

Add penalization term in loss funtion #45

Closed SergioQuijanoRey closed 1 year ago

SergioQuijanoRey commented 1 year ago

In train logs, we see that the norm of the network outputs (the embeddings) goes to zero really fast. This can be why the network is not producing nice results.

Example of how the norm goes fast to zero

SergioQuijanoRey commented 1 year ago

Last change is not enough. With PENALTY_FACTOR = 1 we got the following curves:

Running loss and embedding norms

SergioQuijanoRey commented 1 year ago

Closed in #46. The feature is added, but it does not produce desired results