You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I would like to weight the contribution of each item to the overall loss based on some metadata. For instance, I have a few different types of images (i.e. sunny, cloudy, snowy), one of which is particularly challenging (snowy). I'd like the training to focus more on those instances by increasing their corresponding weight.
Any other context?
I know I could resort to a sampler to sample more often those types of images, but it's not what I want. I'd like to sample all items with equal frequency, just modify their contribution to the loss depending on their type.
The text was updated successfully, but these errors were encountered:
I've submitted a draft PR with my suggestion on how to potentially deal with this. Let me know your thought, more than happy to add the tests if you believe this approach is on the right track.
What's the feature?
I would like to weight the contribution of each item to the overall loss based on some metadata. For instance, I have a few different types of images (i.e. sunny, cloudy, snowy), one of which is particularly challenging (snowy). I'd like the training to focus more on those instances by increasing their corresponding weight.
Any other context?
I know I could resort to a sampler to sample more often those types of images, but it's not what I want. I'd like to sample all items with equal frequency, just modify their contribution to the loss depending on their type.
The text was updated successfully, but these errors were encountered: