@AlecRad is this similar to the model you used most recently for generating faces?
Samples after one epoch on 2.7M LSUN bedroom samples suggests high visual quality is not just overfitting.pic.twitter.com/XekfbKMLTa
-
-
-
@kcimc Yep, pretty much the same architecture. I'm trying to do that paper thing and details should be up on arxiv within the week. - Voir les réponses
Nouvelle conversation -
-
-
@AlecRad How does that follow? Isn't one epoch 2.7 million gradient updates? -
@roydanroy 128mb, ~20K updates. My understanding is overfitting is tied to # of iterations over data with SGD. http://arxiv.org/abs/1509.01240 - Voir les réponses
Nouvelle conversation -
-
-
@AlecRad what's stopping the network from hooking early on to a few modes and freezing .... in that case, one epoch argument isn't valid. -
@amiconfusediam Would that be better described as underfitting the true distribution not overfitting? It's quite likely that's happening. - Voir les réponses
Nouvelle conversation -
-
-
@AlecRad try finding nearest neighbors in your trainset to get a feeling for whether bedroom parts are copied wholesale fr training examplesMerci. Twitter en tiendra compte pour améliorer votre fil. SupprimerSupprimer
-
Le chargement semble prendre du temps.
Twitter est peut-être en surcapacité ou rencontre momentanément un incident. Réessayez ou rendez-vous sur la page Twitter Status pour plus d'informations.