xidulu commented on a change in pull request #18403:
URL: https://github.com/apache/incubator-mxnet/pull/18403#discussion_r447495416



##########
File path: python/mxnet/gluon/probability/block/stochastic_block.py
##########
@@ -0,0 +1,127 @@
+# 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.
+
+# coding: utf-8
+# pylint: disable=abstract-method
+"""Stochastic block class."""
+__all__ = ['StochasticBlock', 'StochasticSequential']
+
+from functools import wraps
+from ...block import HybridBlock
+from ...utils import _indent
+
+
+class StochasticBlock(HybridBlock):
+    """`StochasticBlock` extends `HybridBlock` to support accumulating loss
+    in the forward phase, which is extremely useful in building Bayesian 
Neural Network,
+    where the loss function is composed of a classification loss and a KL loss.
+
+    """
+
+    def __init__(self, **kwargs):
+        super(StochasticBlock, self).__init__(**kwargs)
+        self._losses = []
+        self._losscache = []
+
+    def add_loss(self, loss):
+        self._losscache.append(loss)
+
+    @staticmethod
+    def collectLoss(func):
+        """To accumulate loss during the forward phase, one could first 
decorate
+        hybrid_forward with `StochasticBlock.collectLoss,
+        and then collect the loss tensor `x` by calling self.add_loss(x).
+        For example, in the following forward function,
+        we generate samples from a Gaussian parameterized by `loc` and `scale` 
and
+        accumulate the KL-divergence between it and its prior into the block's 
loss storage.:
+        @StochasticBlock.collectLoss
+        def hybrid_forward(self, F, loc, scale):
+            qz = mgp.Normal(loc, scale)
+            # prior
+            pz = mgp.Normal(F.np.zeros_like(loc), F.np.ones_like(scale))
+            self.add_loss(mgp.kl_divergence(qz, pz))
+            return qz.sample()
+        """
+        @wraps(func)
+        def inner(self, *args, **kwargs):
+            # Loss from hybrid_forward
+            func_out = func(self, *args, **kwargs)
+            collected_loss = self._losscache
+            self._losscache = []
+            return (func_out, collected_loss)
+
+        return inner
+
+    def __call__(self, *args, **kwargs):
+               # pylint: disable=arguments-differ
+        out = super().__call__(*args, **kwargs)
+        self._losses.extend(out[1])
+        return out[0]

Review comment:
       @leezu 
   
   I add two checks here: 
https://github.com/apache/incubator-mxnet/pull/18403/files#diff-85458cf5116b137da8148bf5b38bcfaeR74
   
https://github.com/apache/incubator-mxnet/pull/18403/files#diff-85458cf5116b137da8148bf5b38bcfaeR78
   
   To make it clearer, I list several possible situations:
   
   1. Users call add_loss inside functions decorated by CollectLoss, add_loss 
appends losses into _losscache, _losscache would then get cleared in 
CollectLoss, len(_losscache) becomes 0 when __call__ is invoked.
   
   2. Users call add_loss without using CollectLoss, add_loss appends losses 
into _losscache, _losscache  still contains value when entering  __call__, in 
this case, a exception will be raised.
   
   3. Users use CollectLoss without calling add_loss, self._losses = out[1] = []
   
   4. Users use StochasticBlock without calling CollectLoss or add_loss, 
len(out) == 1, out[1] will not be accessed.
   
   




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
us...@infra.apache.org


Reply via email to