diff options
author | 2016-11-01 15:18:39 -0800 | |
---|---|---|
committer | 2016-11-01 16:24:25 -0700 | |
commit | d8c94dba8f53626077b21f90df309577827a0f18 (patch) | |
tree | 951d8a1563e7cb1e5effa31a04f414333e13243e /tensorflow/python/framework/ops.py | |
parent | 07ae2d1d606fa3f8f3fb09bcf65b570b3c606173 (diff) |
Remove weight_parameters from OpStats and graph_metrics.
Change: 137885496
Diffstat (limited to 'tensorflow/python/framework/ops.py')
-rw-r--r-- | tensorflow/python/framework/ops.py | 12 |
1 files changed, 5 insertions, 7 deletions
diff --git a/tensorflow/python/framework/ops.py b/tensorflow/python/framework/ops.py index 3e50a357cc..ac13f57b30 100644 --- a/tensorflow/python/framework/ops.py +++ b/tensorflow/python/framework/ops.py @@ -1881,23 +1881,21 @@ class RegisterStatistics(object): Well-known types of statistics include these so far: - - weight_parameters: For operations like MatMul, Conv, and BiasAdd that take - learned weights as inputs, this statistic captures how many numerical values - are used. This is good to know because the weights take up most of the size - of a typical serialized graph on disk. - - flops: When running a graph, the bulk of the computation happens doing numerical calculations like matrix multiplications. This type allows a node to return how many floating-point operations it takes to complete. The total number of FLOPs for a graph is a good guide to its expected latency. You can add your own statistics just by picking a new type string, registering - functions for the ops you care about, and then calling something like - python/tools/graph_metrics.py with the new type as an argument. + functions for the ops you care about, and then calling get_stats_for_node_def. If a statistic for an op is registered multiple times, a KeyError will be raised. + Since the statistics is counted on a per-op basis. It is not suitable for + model parameters (capacity), which is expected to be counted only once, even + if it is shared by multiple ops. (e.g. RNN) + For example, you can define a new metric called doohickey for a Foo operation by placing this in your code: |