Skip to content

improve the initializer interface for layers #5632

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
dzhwinter opened this issue Nov 14, 2017 · 7 comments
Closed

improve the initializer interface for layers #5632

dzhwinter opened this issue Nov 14, 2017 · 7 comments
Assignees

Comments

@dzhwinter
Copy link
Contributor

dzhwinter commented Nov 14, 2017

The parameters in layers need a better flexible and user-friendly initializer interface. Currently, most of our layers have the fixed init method, such as conv2d use the gaussian initializer conv2d.
We need to rewrite that part.

@jacquesqiao
Copy link
Member

jacquesqiao commented Nov 16, 2017

this initializer for conv2d is designed in purpose, and it's a good initialization method. But it needs to allow users to define their own initializer~

@qingqing01
Copy link
Contributor

For the convolution operators, maybe we need to define a MSRAInit and set it by default. The MSRAInit in caffe2 : https://github.com/caffe2/caffe2/blob/master/caffe2/operators/filler_op.h#L462

@abhinavarora
Copy link
Contributor

I agree with @qingqing01. I have opened an issue for this #5752.

@dzhwinter
Copy link
Contributor Author

I think we need to make the initialized method configurable. Even though the MSRAInit is perfect, when a user wants to do some init trick, our design will block their use cases.

@dzhwinter
Copy link
Contributor Author

@abhinavarora
Copy link
Contributor

@dzhwinter, I am working on a PR to add that flexibility. Will send you all review soon.

@abhinavarora
Copy link
Contributor

abhinavarora commented Nov 18, 2017

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants