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

Relying on the output_arg and control_output lists rather than the maps to make snapshot fingerprinting (more) deterministic. The failure modes were graphs that had multiple overlapping cycles in them. The way we broke cycles became dependent on the order in which we traversed the graph.

Browse Source 
Added a test that fails with the previous code.

PiperOrigin-RevId: 292562217
Change-Id: I83464b03e347959bb2f00e6443169c419f934c27
This commit is contained in:
Rohan Jain 2020-01-31 09:46:52 -08:00 committed by TensorFlower Gardener
parent 00712963cc
commit 4e20c32249
2 changed files with 31 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
tensorflow/core/kernels/data/dataset_utils.cc
View File

@ -308,9 +308,10 @@ Status HashFunctionImpl(const FunctionDefLibrary& library,
// return argument / control return argument or whether we can relax it and
// hash the index (etc...)
uint64 ret_hash = func.ret_size();
for (const auto& ret : func.ret()) {
for (const auto& output_arg : func.signature().output_arg()) {
std::string ret_name = func.ret().at(output_arg.name());
std::pair<std::string, std::string> node_spec =
absl::StrSplit(ret.second, absl::MaxSplits(':', 1));
absl::StrSplit(ret_name, absl::MaxSplits(':', 1));
// For every return value, we need to hash the output node (and the subgraph
// rooted at the output node) to ensure that the computation graph that
// ends at the output node has not changed.
@ -321,15 +322,16 @@ Status HashFunctionImpl(const FunctionDefLibrary& library,
HashNodeImpl(node_graph, *node_def, &node_hash, visited, cache));
uint64 node_port_hash = Hash64(node_spec.second);
ret_hash = Hash64CombineUnordered(
ret_hash, Hash64Combine(Hash64(ret.first),
ret_hash = Hash64Combine(
ret_hash, Hash64Combine(Hash64(output_arg.name()),
Hash64Combine(node_hash, node_port_hash)));
}
uint64 control_ret_hash = func.control_ret_size();
for (const auto& ret : func.control_ret()) {
for (const auto& control_output : func.signature().control_output()) {
std::string control_output_name = func.control_ret().at(control_output);
std::pair<std::string, std::string> node_spec =
absl::StrSplit(ret.second, absl::MaxSplits(':', 1));
absl::StrSplit(control_output_name, absl::MaxSplits(':', 1));
const NodeDef* node_def;
TF_RETURN_IF_ERROR(FindNode(node_graph, node_spec.first, &node_def));
@ -340,7 +342,7 @@ Status HashFunctionImpl(const FunctionDefLibrary& library,
control_ret_hash = Hash64CombineUnordered(
control_ret_hash,
Hash64Combine(Hash64(ret.first),
Hash64Combine(Hash64(control_output),
Hash64Combine(node_hash, node_port_hash)));
}

22
tensorflow/core/kernels/data/dataset_utils_test.cc
View File

@ -300,6 +300,28 @@ TEST_F(DatasetHashUtilsTest, HashFunctionDifferentInternalNodeNames) {
EXPECT_EQ(GetHash(fl, *f1), GetHash(fl, *f2));
}
TEST_F(DatasetHashUtilsTest, HashFunctionWithMultipleCycles) {
uint64 hash = 0;
for (int i = 0; i < 1000; ++i) {
FunctionDefLibrary fl;
FunctionDef* f1 = fl.add_function();
*f1 = FunctionDefHelper::Create("TwoCyleGraph", {},
{"p: float", "q: float"}, {},
{{{"A"}, "Abs", {"B"}},
{{"B"}, "Add", {"C", "D"}},
{{"C"}, "Ceil", {"A"}},
{{"D"}, "Cos", {"E"}},
{{"E"}, "Floor", {"B"}}},
{{"p", "A:0"}, {"q", "D:0"}}, {});
uint64 t = GetHash(fl, *f1);
if (hash == 0) {
hash = t;
} else {
EXPECT_EQ(t, hash);
}
}
}
TEST_F(DatasetHashUtilsTest, HashNodeSameGraphDifferentNames) {
GraphDef gd;