IntenseWebs/openvino
3
0 Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

GRUCellFusion - fix handling bias in rz order (#17651)

Browse Source 
Currently GRUCellFusion treats bias as if it's always in zr order,
which sometimes is not the case. So whenever GRUCellFusion detects rz order,
it should also split bias into r and z and concat the halves into zr.

Ticket: CVS-97025
This commit is contained in:
Mateusz Tabaka authored and GitHub committed 2023-05-22 20:29:05 +04:00
1 parent c5f4406c34
commit 48f207f961
3 files changed
+31 -29
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/common/transformations/src/transformations/common_optimizations/gru_cell_fusion.cpp
+19 -17
View File
@@ -120,32 +120,34 @@ ov::pass::GRUCellFusion::GRUCellFusion() {
auto cnt_of_consumers_of_zero_out = pattern_split->get_output_target_inputs(0).size();
auto cnt_of_consumers_of_first_out = pattern_split->get_output_target_inputs(1).size();
Output<Node> Wzrh, Rzrh;
Output<Node> B;
if (pattern_map.find(add_1) != pattern_map.end()) {
B = pattern_map[add_1]->input_value(1);
} else {
B = rg.make<Constant>(WR.get_element_type(), Shape{1, static_cast<size_t>(2 * hidden_size)}, 0);
}
Output<Node> Bh;
if (pattern_map.find(add_2) != pattern_map.end()) {
Bh = pattern_map[add_2]->input_value(1);
} else {
Bh = rg.make<Constant>(WRh.get_element_type(), Shape{1, static_cast<size_t>(hidden_size)}, 0);
}
Output<Node> Wzrh, Rzrh, Bzrh;
if (cnt_of_consumers_of_zero_out == 1 && cnt_of_consumers_of_first_out == 2) {
tie(Wzrh, Rzrh) = process_weights(rg, false, WR, WRh, input_size, hidden_size, axis_0, axis_1);
auto split_B_r_z = rg.make<Split>(B, axis_1, 2);
Bzrh = rg.make<Concat>(OutputVector{split_B_r_z->output(1), split_B_r_z->output(0), Bh}, 1);
} else if (cnt_of_consumers_of_zero_out == 2 && cnt_of_consumers_of_first_out == 1) {
tie(Wzrh, Rzrh) = process_weights(rg, true, WR, WRh, input_size, hidden_size, axis_0, axis_1);
Bzrh = rg.make<Concat>(OutputVector{B, Bh}, 1);
} else {
// we can't detect the weights format
return false;
}
Output<Node> bias_add_1;
if (pattern_map.find(add_1) != pattern_map.end()) {
bias_add_1 = pattern_map[add_1]->input_value(1);
} else {
bias_add_1 = rg.make<Constant>(WR.get_element_type(), Shape{1, static_cast<size_t>(2 * hidden_size)}, 0);
}
Output<Node> bias_add_2;
if (pattern_map.find(add_2) != pattern_map.end()) {
bias_add_2 = pattern_map[add_2]->input_value(1);
} else {
bias_add_2 = rg.make<Constant>(WRh.get_element_type(), Shape{1, static_cast<size_t>(hidden_size)}, 0);
}
auto B = rg.make<Concat>(OutputVector{bias_add_1, bias_add_2}, 1);
auto squeeze_B = rg.make<Squeeze>(B, axis_0);
auto squeeze_B = rg.make<Squeeze>(Bzrh, axis_0);
string act_name_1 = pattern_map.at(activation_1)->get_type_name();
string act_name_2 = pattern_map.at(activation_2)->get_type_name();
src/common/transformations/tests/common_optimizations/gru_cell_fusion.cpp
+11 -9
View File
@@ -96,10 +96,10 @@ shared_ptr<Model> gen_reference(WeightsFormat format,
auto WRh = make_shared<Parameter>(f32, Shape{hidden_size, input_size + hidden_size});
ParameterVector params = {X, H, WR, WRh};
shared_ptr<Node> Bzr = make_shared<Constant>(f32, Shape{1, 2 * hidden_size}, 0);
shared_ptr<Node> B = make_shared<Constant>(f32, Shape{1, 2 * hidden_size}, 0);
if (use_bias_add_1) {
Bzr = make_shared<Parameter>(f32, Shape{1, 2 * hidden_size});
params.push_back(dynamic_pointer_cast<Parameter>(Bzr));
B = make_shared<Parameter>(f32, Shape{1, 2 * hidden_size});
params.push_back(dynamic_pointer_cast<Parameter>(B));
}
shared_ptr<Node> Bh = make_shared<Constant>(f32, Shape{1, hidden_size}, 0);
if (use_bias_add_2) {
@@ -112,11 +112,12 @@ shared_ptr<Model> gen_reference(WeightsFormat format,
auto split_lenghts = make_shared<Constant>(i64, Shape{2}, vector<size_t>{input_size, hidden_size});
auto split_WRh = make_shared<VariadicSplit>(WRh, axis_1, split_lenghts);
Output<Node> Wzrh, Rzrh;
Output<Node> Wzrh, Rzrh, Bzrh;
if (format == WeightsFormat::zr) {
auto split_WRzr = make_shared<VariadicSplit>(WR, axis_1, split_lenghts);
Wzrh = make_shared<Concat>(OutputVector{split_WRzr->output(0), split_WRh->output(0)}, 0);
Rzrh = make_shared<Concat>(OutputVector{split_WRzr->output(1), split_WRh->output(1)}, 0);
Bzrh = make_shared<Concat>(OutputVector{B, Bh}, 1);
} else {
auto split_WRrz = make_shared<VariadicSplit>(WR, axis_1, split_lenghts);
auto split_W_r_z = make_shared<Split>(split_WRrz->output(0), axis_0, 2);
@@ -125,10 +126,11 @@ shared_ptr<Model> gen_reference(WeightsFormat format,
make_shared<Concat>(OutputVector{split_W_r_z->output(1), split_W_r_z->output(0), split_WRh->output(0)}, 0);
Rzrh =
make_shared<Concat>(OutputVector{split_R_r_z->output(1), split_R_r_z->output(0), split_WRh->output(1)}, 0);
auto split_B_r_z = make_shared<Split>(B, axis_1, 2);
Bzrh = make_shared<Concat>(OutputVector{split_B_r_z->output(1), split_B_r_z->output(0), Bh}, 1);
}
auto B = make_shared<Concat>(OutputVector{Bzr, Bh}, 1);
auto squeeze_B = make_shared<Squeeze>(B, axis_0);
auto squeeze_B = make_shared<Squeeze>(Bzrh, axis_0);
auto cell =
make_shared<GRUCell>(X, H, Wzrh, Rzrh, squeeze_B, hidden_size, vector<string>{activation_1, activation_2});
return make_shared<Model>(OutputVector{cell}, params);
@@ -201,13 +203,13 @@ static const vector<GRUFusionParams> params = {
GRUFusionParams{WeightsFormat::zr, "tanh", "sigmoid", 2, 128, 32, true, true},
GRUFusionParams{WeightsFormat::zr, "tanh", "tanh", 2, 128, 32, true, false},
GRUFusionParams{WeightsFormat::zr, "sigmoid", "relu", 2, 128, 32, false, false},
GRUFusionParams{WeightsFormat::zr, "relu", "tanh", 2, 128, 32, false, true},
GRUFusionParams{WeightsFormat::rz, "sigmoid", "tanh", 1, 1, 1, true, true},
GRUFusionParams{WeightsFormat::rz, "tanh", "sigmoid", 2, 128, 32, true, true},
GRUFusionParams{WeightsFormat::rz, "relu", "sigmoid", 2, 128, 32, true, true},
GRUFusionParams{WeightsFormat::rz, "tanh", "tanh", 2, 128, 32, true, false},
GRUFusionParams{WeightsFormat::rz, "sigmoid", "relu", 2, 128, 32, false, false},
// 92424: accuracy issue, it looks like the test infrastructure issue, the graphs are the same.
// GRUFusionParams{WeightsFormat::zr, "relu", "tanh", 2, 128, 32, false, true},
// GRUFusionParams{WeightsFormat::rz, "relu", "tanh", 2, 128, 32, false, true},
GRUFusionParams{WeightsFormat::rz, "relu", "tanh", 2, 128, 32, false, true},
};
INSTANTIATE_TEST_SUITE_P(GRUFusionTest, GRUFusionTest, ValuesIn(params));
src/tests/ie_test_utils/common_test_utils/graph_comparator.cpp
+1 -3
View File
@@ -1039,9 +1039,7 @@ AccuracyCheckResult accuracy_check(const std::shared_ptr<ov::Model>& ref_functio
IE_ASSERT(ref_outputs.size() == outputs.size());
for (int i = 0; i < ref_outputs.size(); i++) {
ov::test::utils::compare(ref_outputs[i], outputs[i],
std::numeric_limits<double>::max(),
std::numeric_limits<double>::max());
ov::test::utils::compare(ref_outputs[i], outputs[i], 5e-4, 1e-3);
}
}
catch (const std::runtime_error &re) {