## Description ## This implements a special version of AdaGrad with a learning rate per row instead of for every parameter. This is useful for learning large embedding matrices, as it significantly reduces the memory requirements for the state array as well as reduces over-fitting due to having less learning rates. Consequently only 2D weight arrays are supported.
Compared to the AdaGrad optimizer already included in MXNet, this Optimizer also supports the group lasso proximal operator for inducing group sparsity, where each row of the weight is considered a group. As there is a single learning rate per group, we can use a [closed form solution for the proximal operator](https://en.wikipedia.org/wiki/Proximal_gradient_methods_for_learning#Group_lasso) which is not possible if different learning rates are used for parameters within a single group. This Optimizer is useful for the GluonNLP project. ## Checklist ## ### Essentials ### Please feel free to remove inapplicable items for your PR. - [X] Changes are complete (i.e. I finished coding on this PR) - [X] All changes have test coverage: - Unit tests are added for small changes to verify correctness (e.g. adding a new operator) - Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore) - Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL) - [X] Code is well-documented: - For user-facing API changes, API doc string has been updated. - For new C++ functions in header files, their functionalities and arguments are documented. - For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable - Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html - [X] To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change ### Changes ### - [X] Add contrib.optimizer.ProximalGroupAdaGrad ## Comments ## - If you have any suggestions for the API naming or other suggestions pleas let me know. - Forcing an eager update of the lazily accumulated group lasso regularization for a set of rows requires access to some private API. I am not sure how this could be solved without access to private API: fake_grad = mx.nd.sparse.row_sparse_array( (mx.nd.zeros((len(indices_to_force_update), weight.shape[1])), indices_to_force_update)) weight.grad()[:] = fake_grad weight.data()._fresh_grad = True trainer._optimizer._index_update_count[0] -= 1 trainer._optimizer.num_update -= 1 trainer.step(batch_size=1) @szha @szhengac @eric-haibin-lin [ Full content available at: https://github.com/apache/incubator-mxnet/pull/12365 ] This message was relayed via gitbox.apache.org for [email protected]
