Revamp "sample" function to be a __call__ method #32
Replies: 3 comments 3 replies
-
|
This seems great👍 |
Beta Was this translation helpful? Give feedback.
-
|
Glad you like the idea! But this is a major refactor that I would want to handle myself. I'd appreciate any ideas you have on how to make this work though! For example, what sorts of housekeeping should be in the call method vs the sample method? When a user writes their own custom models, what should they overload, the sample method, or some other sub function? I was hoping to discuss the idea here before going and making a bunch of changes. |
Beta Was this translation helpful? Give feedback.
-
|
can u tell me where is the scripts where u will be making changes? |
Beta Was this translation helpful? Give feedback.
-
|
It would be a change to the core |
Beta Was this translation helpful? Give feedback.
-
|
Added a new branch |
Beta Was this translation helpful? Give feedback.
-
|
Ok, the new interface has been merged, all the tests have passed, looks good! |
Beta Was this translation helpful? Give feedback.
-
This would be a major refactor of the interface. The idea is that instead of calling:
model.sample(image)one would simply call:
model(image)This is somewhat cleaner, but also makes it clearer what the core purpose of the model is for, sampling images. Internally there could still be a sample method but it would be more low level/functional. For those trying to use AutoProf for a more specialized workflow it may be helpful to have a clear way to call lower level functions as well. The inspiration comes from pytorch nn.Module which has users implement a
forwardfunction however the modules themselves are called by__call__so that the nn.Module can handle hooks and other housekeeping beforeforwardis called. This seems like a useful separation of tasks.At this point it's just an idea, something to workshop.
Beta Was this translation helpful? Give feedback.
All reactions