experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Check both output name and output slot in duplicate scope id sanity check.

Browse Source 
Before this change, we would throw an error if different outputs of a node were
committed to different scope ids.  Since that is legal, this change fixes the
bug by making the check based on both output name and output index.

PiperOrigin-RevId: 263482156
This commit is contained in:
Ayush Dubey 2019-08-14 19:39:05 -07:00 committed by TensorFlower Gardener
parent f2be0e04c6
commit dcdca11bcb
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/core/grappler/optimizers/scoped_allocator_optimizer.cc
View File

@ -312,7 +312,9 @@ class UnaryElementwiseRewriter : public ScopedAllocatorOptimizer::Rewriter {
AttrSlice n_attrs = AttrSlice(*nd.from_node_def); AttrSlice n_attrs = AttrSlice(*nd.from_node_def);
std::vector<int32> scope_ids; std::vector<int32> scope_ids;
Status ss = GetNodeAttr(n_attrs, kScopedAllocatorAttrName, &scope_ids); Status ss = GetNodeAttr(n_attrs, kScopedAllocatorAttrName, &scope_ids);
if (ss.ok()) { // Check that both output name and output slot match. It is okay to have
// different outputs of the input committed to different scope ids.
if (ss.ok() && scope_ids[0] == nd.output_slot) {
LOG(INFO) << "Abandoning ScopedAllocatorOptimizer because input " LOG(INFO) << "Abandoning ScopedAllocatorOptimizer because input "
<< nd.from_node_def->name() << " output " << scope_ids[0] << nd.from_node_def->name() << " output " << scope_ids[0]
<< " is already assigned to scope_id " << scope_ids[1]; << " is already assigned to scope_id " << scope_ids[1];