-
Notifications
You must be signed in to change notification settings - Fork 1.6k
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鈥檒l occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add Support for Recurrent Neural Networks (RNNs) in KFACOptimizer Class #3683
Comments
This sounds like an interesting addition that we would tentatively consider adding. Could you please join our office hours to discuss with us (MWF at 9am PST)? https://forum.deepchem.io/t/announcing-the-deepchem-office-hours/293 |
Hi, Thanks for the reply. I'm interested in joining the office hours to discuss this feature request. Could you please specify which day of the week the office hours occur? |
Hi @neuronphysics the office hours happen at 9am PST on Mondays, Wednesdays and Fridays |
Hi, I joined the Google Meet session on Friday for the office hour, but I wasn't admitted. |
Hi @neuronphysics there was no office hours on Friday due to thanksgiving, you can join on Monday (27th Nov 2023) |
Hi I am trying to attend your office hour. Should I use the old google meet? |
Hi @neuronphysics , the office hours is delayed by 15 min today, requesting you to wait, you will be admitted as soon as the office hours start |
@rbharath sir , @shreyasvinaya sir can I work on this issue? |
馃殌 Feature
I would like to request the addition of support for Recurrent Neural Networks (RNNs) in the KFACOptimizer class. Currently, the KFACOptimizer class works for linear and 2D convolution layers, but it does not support RNN layers. RNNs are widely used in various applications, and adding support for them in the KFACOptimizer class would be highly beneficial.
Proposal
I propose adding an RNN module to the KFACOptimizer class to handle RNN layers. This would involve modifying the KFACOptimizer class to compute the necessary statistics and whitened tensors for RNN layers, similar to how it currently handles linear and convolution layers.
Motivation
RNNs are a fundamental component of many deep learning models, especially in tasks involving sequential data, natural language processing, and time-series analysis. Efficient optimization techniques, such as KFAC, can significantly improve the training of RNN-based models by providing better approximations of the Fisher information matrix.
Adding support for RNNs in the KFACOptimizer class would enable researchers and practitioners to apply the KFAC optimization technique to a wider range of models, resulting in improved training efficiency and convergence. This enhancement would benefit the deep learning community by making state-of-the-art optimization techniques more accessible and effective for RNN-based applications.
Additional Context
I have provided an initial implementation attempt in the feature request description, which includes modifications to the KFACOptimizer class to handle RNN layers. I kindly request the repository maintainers to evaluate this implementation, provide feedback, and consider including it in the official codebase if deemed appropriate.
Additionally, I would like to refer the maintainers to the following paper for more details on the topic:
I believe that adding RNN support to the KFACOptimizer class would be a valuable addition to the repository and benefit the deep learning community as a whole. Thank you for considering this feature request.
The text was updated successfully, but these errors were encountered: