I have a question regarding auto encoders. Let's say we want to train an auto encoder to be able to learn a latent space representation of some pictures (dogs pictures for example..). I was wondering how do we train those models in practice, do we:
train a model for every image sample, so that the model will perfectly fit to every picture
Or,
train an unique model on all the samples that we have, so that the model will learn the distribution of the data and be able to encode it correctly in the latent space
What would be the loss function for this case ? MSE ?
Hi Younes,
1) You would train a single model on all of your data, so the model learns to generalize across natural images. Training a single image has little use.
2) Yes, in general, we would train with MSE.
Best,
Semih
Question about Auto Encoders
Hi !
I have a question regarding auto encoders. Let's say we want to train an auto encoder to be able to learn a latent space representation of some pictures (dogs pictures for example..). I was wondering how do we train those models in practice, do we:
Or,
What would be the loss function for this case ? MSE ?
Thank you for your time and for your dedication
1
Hi Younes,
1) You would train a single model on all of your data, so the model learns to generalize across natural images. Training a single image has little use.
2) Yes, in general, we would train with MSE.
Best,
Semih
1
Add comment