saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-29 20:14:01 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Fix parsing initial hidden state in RNN (#1626)

Browse source 
* Fix the way initial hidden state is used for reverse direction in RNN

* Add test case

* Updates
This commit is contained in:
Hariharan Seshadri 2019-08-16 10:12:46 -07:00 committed by GitHub
parent f9834105aa
commit 44a42a6a98
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 59 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
onnxruntime/core/providers/cpu/rnn/rnn.cc
View file

@ -181,8 +181,12 @@ Status RNN<float>::Compute(OpKernelContext* ctx) const {
const float* h_prev = nullptr;
if (t == 0) {
if (initial_h != nullptr)
h_prev = initial_h->template Data<float>();
if (initial_h != nullptr) {
// the shape of initial_h is [num_directions, batch_size, hidden_size]
// so pick the offset (multiple of Y_frame_size == batch_size * hidden_size_)
// based on the direction
h_prev = initial_h->template Data<float>() + (direction * Y_frame_size);
}
} else {
if (isReverse)
h_prev = Y_buffer_data_current_frame + num_directions * Y_frame_size;

54
onnxruntime/test/providers/cpu/rnn/rnn_op_test.cc
View file

@ -345,7 +345,7 @@ TEST(RNNTest, RNN_forward_direction_zigged_batch) {
test.Run();
}
TEST(RNNTest, RNN_bidirectional) {
TEST(RNNTest, RNN_bidirectional_0) {
OpTester test("RNN");
int64_t num_directions = 2, input_size = 2, hidden_size = 3, batch_size = 1, seq_length = 5;
@ -407,6 +407,58 @@ TEST(RNNTest, RNN_bidirectional) {
test.Run();
}
TEST(RNNTest, RNN_bidirectional_1) {
OpTester test("RNN");
int64_t num_directions = 2, input_size = 2, hidden_size = 2, batch_size = 1, seq_length = 1;
test.AddAttribute("activations", vector<string>(num_directions, "Tanh"));
test.AddAttribute("direction", "bidirectional");
test.AddAttribute("hidden_size", hidden_size);
std::vector<int64_t> X_dims = {seq_length, batch_size, input_size};
std::vector<float> X_data({1.0F, 1.0F});
test.AddInput<float>("X", X_dims, X_data);
std::vector<int64_t> W_dims = {num_directions, hidden_size, input_size};
std::vector<float> W_data({1.0F, 1.0F, 1.0F, 1.0F,
1.0F, 1.0F, 1.0F, 1.0F});
test.AddInput<float>("W", W_dims, W_data);
std::vector<int64_t> R_dims = {num_directions, hidden_size, hidden_size};
std::vector<float> R_data({// forward
1.0F, 1.0F,
1.0F, 1.0F,
// reverse
1.0F, 1.0F,
1.0F, 1.0F});
test.AddInput<float>("R", R_dims, R_data);
std::vector<int64_t> B_dims = {num_directions, 2 * hidden_size};
std::vector<float> B_data({0.0F, 0.0F, 0.0F, 0.0F,
0.0F, 0.0F, 0.0F, 0.0F});
test.AddInput<float>("B", B_dims, B_data);
std::vector<int64_t> sequence_lens_dims({batch_size});
std::vector<int> sequence_lens_data(batch_size, (int)seq_length);
test.AddInput<int>("sequence_lens", sequence_lens_dims, sequence_lens_data);
std::vector<int64_t> initial_h_dims = {num_directions, batch_size, hidden_size};
std::vector<float> initial_h_data({0.1F, 0.2F, 0.3F, 0.4F});
test.AddInput<float>("initial_h", initial_h_dims, initial_h_data);
std::vector<int64_t> Y_dims = {seq_length, num_directions, batch_size, hidden_size};
std::vector<float> Y_data({0.98009639F, 0.98009639F, 0.99100745F, 0.99100745F});
test.AddOutput<float>("Y", Y_dims, Y_data);
std::vector<int64_t> Y_h_dims{num_directions, batch_size, hidden_size};
std::vector<float> Y_h_data({0.98009639F, 0.98009639F, 0.99100745F, 0.99100745F});
test.AddOutput<float>("Y_h", Y_h_dims, Y_h_data);
test.Run();
}
typedef enum {
RNNOutputY,
RNNOutputY_h,