[GitHub] eric-haibin-lin commented on a change in pull request #12365: Proximal Adagrad optimizer with row-wise learning rate

[GitBox <gi...@apache.org>]

eric-haibin-lin commented on a change in pull request #12365: Proximal Adagrad optimizer with row-wise learning rate
URL: https://github.com/apache/incubator-mxnet/pull/12365#discussion_r212859994
 
 

 ##########
 File path: python/mxnet/contrib/optimizer.py
 ##########
 @@ -0,0 +1,140 @@
+# coding: utf-8
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+# pylint: disable=too-many-lines
+"""Contrib optimizers."""
+from ..ndarray import (NDArray, clip, full, mean, norm,
+                       proximal_group_adagrad_update, sqrt, square, zeros)
+from ..optimizer import Optimizer
+
+# convenience wrapper for Optimizer.Register
+register = Optimizer.register  # pylint: disable=invalid-name
+
+
+@register
+class ProximalGroupAdaGrad(Optimizer):
+    """Proximal Adagrad optimizer with row-wise learning rates.
+
+    This class implements the AdaGrad optimizer described in *Adaptive
+    Subgradient Methods for Online Learning and Stochastic Optimization*, and
+    available at http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf but
+    uses only a single learning rate for every row of the parameter array.
+
+    This optimizer updates each weight by::
+
+        grad = clip(grad * rescale_grad, clip_gradient)
+        history += mean(square(grad), axis=1, keepdims=True)
+        div = grad / sqrt(history + float_stable_eps)
+        weight += (div + weight * wd) * -lr
+
+    If `l2_regularization_strength > 0` a proximal operator is used to optimize
+    with group lasso objective. Weights are updated lazily if the gradient is
+    sparse. In particular, before using a set of weights for a forward pass,
+    you may want to ensure that the lazily accumulated group lasso
+    regularization is applied. This can be achieved by creating a sparse
+    gradient array that contains explicit 0 data for the indices to be updated:
+
+        fake_grad = mx.nd.sparse.row_sparse_array(
+            (mx.nd.zeros((len(indices), dim)), indices))
+        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)
 
 Review comment:
   This seems not very easy to use..

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services