Skip to content

N
ngraph
Unverified Commit 727c55d3 authored by Scott Cyphers's avatar Scott Cyphers Committed by GitHub
Browse files
Options

Merge pull request #3118 from NervanaSystems/arogowiec/lstm_fix 

Fix fused sigmoid overflow.
parents 123d7663 5db8cbb6
Hide whitespace changes
Inline Side-by-side
Showing with 221 additions and 36 deletions
src/ngraph/runtime/cpu/kernel/sigmoid_multiply.hpp
View file @ 727c55d3
......@@ -57,28 +57,28 @@ namespace ngraph
{
case 0 /*Logistic|Logistic*/:
{
auto c = (in0.exp() * in1.exp()) / ((in0.exp() + 1.f) * (in1.exp() + 1.f));
auto c = 1.f / (((-in0).exp() + 1.f) * ((-in1).exp() + 1.f));
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 1 /*Logistic|Tanh*/:
{
auto c = (in0.exp() * in1.tanh()) / (in0.exp() + 1.f);
auto c = in1.tanh() / ((-in0).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 2 /*Logistic|Identity*/:
{
auto c = (in0.exp() * in1) / (in0.exp() + 1.f);
auto c = in1 / ((-in0).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
break;
case 3 /*Tanh|Logistic*/:
{
auto c = (in0.tanh() * in1.exp()) / (in1.exp() + 1.f);
auto c = in0.tanh() / ((-in1).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
......@@ -99,7 +99,7 @@ namespace ngraph
break;
case 6 /*Identity|Logistic*/:
{
auto c = (in0 * in1.exp()) / (in1.exp() + 1.f);
auto c = in0 / ((-in1).exp() + 1.f);
out_tm.device(
ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(arena)) = c;
}
......@@ -141,10 +141,15 @@ namespace ngraph
{
case 0 /*Logistic|Logistic*/:
{
auto i0 = delta * (in1.exp() * in0.exp()) /
((in1.exp() + 1.f) * ((in0.exp() + 1.f) * (in0.exp() + 1.f)));
auto i1 = delta * (in0.exp() * in1.exp()) /
((in0.exp() + 1.f) * ((in1.exp() + 1.f) * (in1.exp() + 1.f)));
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 = delta * in0_neg_exp /
(in1_log_denominator * in0_log_denominator * in0_log_denominator);
auto i1 = delta * in1_neg_exp /
(in0_log_denominator * in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -153,12 +158,17 @@ namespace ngraph
break;
case 1 /*Logistic|Tanh*/:
{
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 =
delta * (((in1 * 2.f).exp() - 1.f) * in0.exp()) /
(((in1 * 2.f).exp() + 1.f) * ((in0.exp() + 1.f) * (in0.exp() + 1.f)));
auto i1 = delta * (in0.exp() * (4.f * (in1 * 2.f).exp())) /
((in0.exp() + 1.f) *
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f)));
delta * ((in1_2exp - 1.f) * in0_neg_exp) /
(in1_tanh_denominator * in0_log_denominator * in0_log_denominator);
auto i1 =
delta * (4.f * in1_2exp) /
(in0_log_denominator * in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -167,9 +177,12 @@ namespace ngraph
break;
case 2 /*Logistic|Identity*/:
{
auto i0 =
delta * (in1 * in0.exp()) / ((in0.exp() + 1.f) * (in0.exp() + 1.f));
auto i1 = delta * in0.exp() / ((in0.exp() + 1.f));
auto in0_neg_exp = (-in0).exp();
auto in0_log_denominator = in0_neg_exp + 1.f;
auto i0 = delta * (in1 * in0_neg_exp) /
(in0_log_denominator * in0_log_denominator);
auto i1 = delta / in0_log_denominator;
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -178,12 +191,17 @@ namespace ngraph
break;
case 3 /*Tanh|Logistic*/:
{
auto i0 = delta * (in1.exp() * (4.f * (in0 * 2.f).exp())) /
((in1.exp() + 1.f) * ((in0 * 2.f).exp() + 1.f) *
((in0 * 2.f).exp() + 1.f));
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 =
delta * (4.f * in0_2exp) /
(in1_log_denominator * in0_tanh_denominator * in0_tanh_denominator);
auto i1 =
delta * (((in0 * 2.f).exp() - 1.f) * in1.exp()) /
(((in0 * 2.f).exp() + 1.f) * ((in1.exp() + 1.f) * (in1.exp() + 1.f)));
delta * ((in0_2exp - 1.f) * in1_neg_exp) /
(in0_tanh_denominator * in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -192,12 +210,17 @@ namespace ngraph
break;
case 4 /*Tanh|Tanh*/:
{
auto i0 = delta * (((in1 * 2.f).exp() - 1.f) * (4.f * (in0 * 2.f).exp())) /
(((in1 * 2.f).exp() + 1.f) *
(((in0 * 2.f).exp() + 1.f) * ((in0 * 2.f).exp() + 1.f)));
auto i1 = delta * (((in0 * 2.f).exp() - 1.f) * (4.f * (in1 * 2.f).exp())) /
(((in0 * 2.f).exp() + 1.f) *
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f)));
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 =
delta * (in1_2exp - 1.f) * 4.f * in0_2exp /
(in1_tanh_denominator * in0_tanh_denominator * in0_tanh_denominator);
auto i1 =
delta * (in0_2exp - 1.f) * 4.f * in1_2exp /
(in0_tanh_denominator * in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -206,9 +229,12 @@ namespace ngraph
break;
case 5 /*Tanh|Identity*/:
{
auto i0 = delta * (in1 * (4.f * (in0 * 2.f).exp())) /
(((in0 * 2.f).exp() + 1.f) * ((in0 * 2.f).exp() + 1.f));
auto i1 = delta * ((in0 * 2.f).exp() - 1.f) / ((in0 * 2.f).exp() + 1.f);
auto in0_2exp = (in0 * 2.f).exp();
auto in0_tanh_denominator = in0_2exp + 1.f;
auto i0 = delta * in1 * 4.f * in0_2exp /
(in0_tanh_denominator * in0_tanh_denominator);
auto i1 = delta * (in0_2exp - 1.f) / in0_tanh_denominator;
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -217,9 +243,12 @@ namespace ngraph
break;
case 6 /*Identity|Logistic*/:
{
auto i0 = delta * (in1.exp()) / (in1.exp() + 1.f);
auto in1_neg_exp = (-in1).exp();
auto in1_log_denominator = in1_neg_exp + 1.f;
auto i0 = delta * 1.f / in1_log_denominator;
auto i1 =
delta * (in0 * in1.exp()) / ((in1.exp() + 1.f) * (in1.exp() + 1.f));
delta * in0 * in1_neg_exp / (in1_log_denominator * in1_log_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......@@ -228,9 +257,12 @@ namespace ngraph
break;
case 7 /*Identity|Tanh*/:
{
auto i0 = delta * ((in1 * 2.f).exp() - 1.f) / ((in1 * 2.f).exp() + 1.f);
auto i1 = delta * (in0 * (4.f * (in1 * 2.f).exp())) /
(((in1 * 2.f).exp() + 1.f) * ((in1 * 2.f).exp() + 1.f));
auto in1_2exp = (in1 * 2.f).exp();
auto in1_tanh_denominator = in1_2exp + 1.f;
auto i0 = delta * (in1_2exp - 1.f) / in1_tanh_denominator;
auto i1 = delta * (in0 * (4.f * in1_2exp)) /
(in1_tanh_denominator * in1_tanh_denominator);
i0_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
arena)) = i0;
i1_delta.device(ngraph::runtime::cpu::executor::GetCPUExecutor().get_device(
......
test/models/onnx/lstm_fwd_large_batch_no_clip.prototxt 0 → 100644
View file @ 727c55d3
ir_version: 4
producer_name: "nGraph ONNX Importer"
graph {
node {
input: "X"
input: "W"
input: "R"
output: ""
output: "Y_h"
op_type: "LSTM"
attribute {
name: "clip"
f: 9999.0
type: FLOAT
}
attribute {
name: "direction"
s: "forward"
type: STRING
}
attribute {
name: "hidden_size"
i: 3
type: INT
}
}
name: "compute_graph"
input {
name: "X"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 2
}
dim {
dim_value: 32
}
dim {
dim_value: 1
}
}
}
}
}
input {
name: "W"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 12
}
dim {
dim_value: 1
}
}
}
}
}
input {
name: "R"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 12
}
dim {
dim_value: 3
}
}
}
}
}
output {
name: "Y_h"
type {
tensor_type {
elem_type: 1
shape {
dim {
dim_value: 1
}
dim {
dim_value: 32
}
dim {
dim_value: 3
}
}
}
}
}
}
opset_import {
version: 7
}
test/onnx/onnx_import_rnn.in.cpp
View file @ 727c55d3
......@@ -20,6 +20,7 @@
#include <fstream>
#include <iterator>
#include <limits>
#include <numeric>
#include <sstream>
#include <stdexcept>
#include <vector>
......@@ -203,3 +204,48 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_hardsigmoid_activation)
test_case.set_tolerance(6);
test_case.run();
}
NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_large_batch_no_clip)
{
auto function = onnx_import::import_onnx_model(
file_util::path_join(SERIALIZED_ZOO, "onnx/lstm_fwd_large_batch_no_clip.prototxt"));
auto test_case = ngraph::test::NgraphTestCase(function, "${BACKEND_NAME}");
std::size_t seq_length = 2;
std::size_t batch_size = 32;
std::size_t input_size = 1;
std::size_t hidden_size = 3;
std::vector<float> in_X(seq_length * batch_size * input_size);
std::iota(std::begin(in_X), std::end(in_X), 1.f);
std::vector<float> in_R(4 * hidden_size * hidden_size, 0.1f);
// X
test_case.add_input<float>(in_X);
// W
test_case.add_input<float>(
{0.1f, 0.2f, 0.3f, 0.4f, 1.f, 2.f, 3.f, 4.f, 10.f, 11.f, 12.f, 13.f});
// R
test_case.add_input<float>(in_R);
// Y_h_data
test_case.add_expected_output<float>(
Shape{1, batch_size, hidden_size},
{0.90387899f, 0.9135572f, 0.91772245f, 0.90897038f, 0.92132433f, 0.92825467f, 0.91365823f,
0.92815113f, 0.93676105f, 0.91799162f, 0.93406357f, 0.94344562f, 0.92199681f, 0.93912057f,
0.94859476f, 0.92569357f, 0.94340185f, 0.95250664f, 0.92909964f, 0.94699686f, 0.95545127f,
0.93223207f, 0.94999634f, 0.95765468f, 0.93510761f, 0.9524867f, 0.95929726f, 0.93774272f,
0.9545467f, 0.96051891f, 0.9401536f, 0.95624603f, 0.96142619f, 0.94235605f, 0.95764499f,
0.96209939f, 0.94436539f, 0.95879495f, 0.96259862f, 0.94619635f, 0.95973921f, 0.96296872f,
0.94786299f, 0.96051397f, 0.96324302f, 0.94937864f, 0.96114929f, 0.96344629f, 0.95075587f,
0.96167006f, 0.96359692f, 0.95200645f, 0.96209679f, 0.96370852f, 0.95314133f, 0.9624464f,
0.9637912f, 0.95417069f, 0.96273278f, 0.96385246f, 0.95510395f, 0.96296733f, 0.96389785f,
0.95594975f, 0.96315942f, 0.96393147f, 0.95671607f, 0.96331673f, 0.96395638f, 0.9574102f,
0.96344554f, 0.96397483f, 0.9580388f, 0.96355102f, 0.9639885f, 0.95860795f, 0.96363739f,
0.96399863f, 0.95912322f, 0.96370811f, 0.96400613f, 0.95958963f, 0.96376601f, 0.96401169f,
0.96001179f, 0.96381342f, 0.96401581f, 0.96039386f, 0.96385224f, 0.96401886f, 0.96073964f,
0.96388402f, 0.96402112f, 0.96105254f, 0.96391004f, 0.96402279f});
test_case.run();
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment