Skip to content

Commit 8a79d04

Browse files
favorartfavorart
committed
[NeoMLTest] Add DnnCrossEntropyLossLayerTest
Signed-off-by: Kirill Golikov <[email protected]>
1 parent ea8e833 commit 8a79d04

File tree

2 files changed

+244
-0
lines changed

2 files changed

+244
-0
lines changed

‎NeoML/test/src/CMakeLists.txt

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -12,6 +12,7 @@ target_sources(${PROJECT_NAME} INTERFACE
1212
${CMAKE_CURRENT_SOURCE_DIR}/DnnBlobTest.cpp
1313
${CMAKE_CURRENT_SOURCE_DIR}/DnnChannelwiseWith1x1BlockTest.cpp
1414
${CMAKE_CURRENT_SOURCE_DIR}/DnnConfusionMatrixLayer.cpp
15+
${CMAKE_CURRENT_SOURCE_DIR}/DnnCrossEntropyLossLayerTest.cpp
1516
${CMAKE_CURRENT_SOURCE_DIR}/DnnCtcTest.cpp
1617
${CMAKE_CURRENT_SOURCE_DIR}/DnnDistributedTest.cpp
1718
${CMAKE_CURRENT_SOURCE_DIR}/DnnDropoutTest.cpp

‎NeoML/test/src/DnnCrossEntropyLossLayerTest.cpp

Lines changed: 243 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,243 @@
1+
/* Copyright © 2024 ABBYY
2+
3+
Licensed under the Apache License, Version 2.0 (the "License");
4+
you may not use this file except in compliance with the License.
5+
You may obtain a copy of the License at
6+
7+
http://www.apache.org/licenses/LICENSE-2.0
8+
9+
Unless required by applicable law or agreed to in writing, software
10+
distributed under the License is distributed on an "AS IS" BASIS,
11+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
See the License for the specific language governing permissions and
13+
limitations under the License.
14+
--------------------------------------------------------------------------------------------------------------*/
15+
16+
#include <common.h>
17+
#pragma hdrstop
18+
19+
#include <TestFixture.h>
20+
#include <DnnSimpleTest.h>
21+
22+
using namespace NeoML;
23+
using namespace NeoMLTest;
24+
25+
namespace NeoMLTest {
26+
27+
struct CCrossEntropyLossLayerTestParams final {
28+
CCrossEntropyLossLayerTestParams( float target, float result, float lossValue ) :
29+
Target( target ), Result( result ), LossValue( lossValue ) {}
30+
31+
float Target;
32+
float Result;
33+
float LossValue;
34+
};
35+
36+
class CCrossEntropyLossLayerTest :
37+
public CNeoMLTestFixture, public ::testing::WithParamInterface<CCrossEntropyLossLayerTestParams> {
38+
public:
39+
static bool InitTestFixture() { return true; }
40+
static void DeinitTestFixture() {}
41+
};
42+
43+
} // namespace NeoMLTest
44+
45+
//---------------------------------------------------------------------------------------------------------------------
46+
47+
TEST_F( CCrossEntropyLossLayerTest, ZeroBackwardDiffTest )
48+
{
49+
const auto met = MathEngine().GetType();
50+
if( met != MET_Cpu && met != MET_Cuda ) {
51+
NEOML_HILIGHT( GTEST_LOG_( INFO ) ) << "Skipped rest of test for MathEngine type=" << met << " because no implementation.\n";
52+
return;
53+
}
54+
55+
const float rawData[8]{
56+
0.7f, 0.3f,
57+
0.3f, 0.7f,
58+
0.2f, 0.8f,
59+
0.4f, 0.6f
60+
};
61+
const float rawExpectedDiff[8]{
62+
-0.10032f, 0.10032f,
63+
0.10032f, -0.10032f,
64+
0.08859f, -0.08859f,
65+
0.0f, 0.0f
66+
};
67+
const float rawFloatLabels[8]{
68+
1.f, 0.f,
69+
0.f, 1.f,
70+
0.f, 1.f,
71+
0.f, 0.f
72+
};
73+
const int rawIntLabels[4]{ 0, 1, 1, -1 };
74+
75+
CRandom random;
76+
CDnn dnn( random, MathEngine() );
77+
78+
CPtr<CSourceLayer> data = Source( dnn, "data" );
79+
CPtr<CSourceLayer> label = Source( dnn, "label" );
80+
81+
CPtr<CDnnSimpleTestDummyLearningLayer> learn = new CDnnSimpleTestDummyLearningLayer( MathEngine() );
82+
learn->SetName( "learn" );
83+
learn->Connect( *data );
84+
dnn.AddLayer( *learn );
85+
86+
CPtr<CCrossEntropyLossLayer> loss = CrossEntropyLoss()( learn.Ptr(), label.Ptr() );
87+
88+
CPtr<CDnnBlob> dataBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 2, 2, 2 );
89+
dataBlob->CopyFrom( rawData );
90+
data->SetBlob( dataBlob );
91+
92+
CPtr<CDnnBlob> expectedDiff = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 2, 2, 2 );
93+
expectedDiff->CopyFrom( rawExpectedDiff );
94+
learn->ExpectedDiff = expectedDiff;
95+
96+
CPtr<CDnnBlob> labelBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 2, 2, 2 );
97+
labelBlob->CopyFrom( rawFloatLabels );
98+
label->SetBlob( labelBlob );
99+
100+
dnn.RunAndBackwardOnce();
101+
102+
labelBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Int, 2, 2, 1 );
103+
labelBlob->CopyFrom( rawIntLabels );
104+
label->SetBlob( labelBlob );
105+
106+
dnn.RunAndBackwardOnce();
107+
}
108+
109+
TEST_F( CCrossEntropyLossLayerTest, NoSoftmax )
110+
{
111+
const auto met = MathEngine().GetType();
112+
if( met != MET_Cpu && met != MET_Cuda ) {
113+
NEOML_HILIGHT( GTEST_LOG_( INFO ) ) << "Skipped rest of test for MathEngine type=" << met << " because no implementation.\n";
114+
return;
115+
}
116+
117+
const float resultBuff[]{ 1.f / 15, 2.f / 15, 3.f / 15, 4.f / 15, 5.f / 15,
118+
5.f / 15, 4.f / 15, 3.f / 15, 2.f / 15, 1.f / 15,
119+
0.2f, 0.2f, 0.2f, 0.2f, 0.2f,
120+
0.2f, 0.2f, 0.2f, 0.2f, 0.2f,
121+
0.2f, 0.2f, 0.2f, 0.2f, 0.2f,
122+
0.2f, 0.2f, 0.2f, 0.2f, 0.2f,
123+
0.2f, 0.2f, 0.2f, 0.2f, 0.2f
124+
};
125+
126+
CRandom random;
127+
CDnn dnn( random, MathEngine() );
128+
129+
CPtr<CSourceLayer> result = Source( dnn, "result" );
130+
CPtr<CSourceLayer> target = Source( dnn, "target" );
131+
132+
CPtr<CDnnSimpleTestDummyLearningLayer> learn = new CDnnSimpleTestDummyLearningLayer( MathEngine() );
133+
learn->Connect( *result );
134+
dnn.AddLayer( *learn );
135+
136+
CPtr<CCrossEntropyLossLayer> loss = CrossEntropyLoss( /*softmax*/false )( learn.Ptr(), target.Ptr() );
137+
138+
CPtr<CDnnBlob> resultBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 1, 7, 5 );
139+
resultBlob->CopyFrom( resultBuff );
140+
result->SetBlob( resultBlob );
141+
142+
CPtr<CDnnBlob> floatTargetBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 1, 7, 5 );
143+
floatTargetBlob->Fill( 0.f );
144+
145+
floatTargetBlob->GetData().SetValueAt( 2, 1 );
146+
floatTargetBlob->GetObjectData( 1 ).SetValueAt( 4, 1 );
147+
for( int i = 2; i < 7; ++i ) {
148+
floatTargetBlob->GetObjectData( i ).SetValueAt( 3, 1 );
149+
}
150+
target->SetBlob( floatTargetBlob );
151+
152+
CPtr<CDnnBlob> expectedDiffBLob = resultBlob->GetCopy();
153+
expectedDiffBLob->Fill( 1.f / 7 );
154+
155+
expectedDiffBLob->GetData().SetValueAt( 2, -4.f / 7 );
156+
expectedDiffBLob->GetObjectData( 1 ).SetValueAt( 4, -2 );
157+
for( int i = 2; i < 7; ++i ) {
158+
expectedDiffBLob->GetObjectData( i ).SetValueAt( 3, -4.f / 7 );
159+
}
160+
learn->ExpectedDiff = expectedDiffBLob;
161+
162+
dnn.RunAndBackwardOnce();
163+
164+
{
165+
dnn.DeleteLayer( *learn );
166+
loss->Connect( 0, *result );
167+
dnn.RunAndBackwardOnce();
168+
169+
CMemoryFile file;
170+
{
171+
CArchive archive( &file, CArchive::store );
172+
archive.Serialize( dnn );
173+
}
174+
file.SeekToBegin();
175+
{
176+
CArchive archive( &file, CArchive::load );
177+
archive.Serialize( dnn );
178+
}
179+
180+
result = CheckCast<CSourceLayer>( dnn.GetLayer( result->GetName() ) );
181+
target = CheckCast<CSourceLayer>( dnn.GetLayer( target->GetName() ) );
182+
loss = CheckCast<CCrossEntropyLossLayer>( dnn.GetLayer( loss->GetName() ) );
183+
184+
result->SetBlob( resultBlob );
185+
186+
learn->Connect( *result );
187+
loss->Connect( *learn );
188+
dnn.AddLayer( *learn );
189+
}
190+
191+
CPtr<CDnnBlob> intTargetBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Int, 1, 7, 1 );
192+
intTargetBlob->FillObject<int>( 0, 2 );
193+
intTargetBlob->FillObject<int>( 1, 4 );
194+
for( int i = 2; i < 7; ++i ) {
195+
intTargetBlob->GetObjectData<int>( i ).SetValue( 3 );
196+
}
197+
target->SetBlob( intTargetBlob );
198+
199+
dnn.RunAndBackwardOnce();
200+
}
201+
202+
TEST_P( CCrossEntropyLossLayerTest, BinaryCrossEntropyLossSignTest )
203+
{
204+
const CCrossEntropyLossLayerTestParams params = GetParam();
205+
206+
CRandom random;
207+
CDnn dnn( random, MathEngine() );
208+
209+
CTextStream debugOutput;
210+
dnn.SetLog( &debugOutput );
211+
212+
CSourceLayer* result = Source( dnn, "result" );
213+
CSourceLayer* target = Source( dnn, "target" );
214+
215+
CPtr<CBinaryCrossEntropyLossLayer> loss = BinaryCrossEntropyLoss()( result, target );
216+
217+
CPtr<CDnnBlob> resultBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 1, 1, 1 );
218+
resultBlob->Fill( params.Result );
219+
result->SetBlob( resultBlob );
220+
221+
CPtr<CDnnBlob> targetBlob = CDnnBlob::CreateDataBlob( MathEngine(), CT_Float, 1, 1, 1 );
222+
targetBlob->Fill( params.Target );
223+
target->SetBlob( targetBlob );
224+
225+
dnn.RunAndBackwardOnce();
226+
227+
EXPECT_TRUE( FloatEq( loss->GetLastLoss(), params.LossValue ) );
228+
}
229+
230+
INSTANTIATE_TEST_CASE_P( CnnCrossEntropyLossTestInstantiation, CCrossEntropyLossLayerTest,
231+
::testing::Values(
232+
CCrossEntropyLossLayerTestParams( -1.f, -2.f, 0.126928f ),
233+
CCrossEntropyLossLayerTestParams( -1.f, -0.999f, 0.313531f ),
234+
CCrossEntropyLossLayerTestParams( -1.f, -1.f, 0.313262f ),
235+
CCrossEntropyLossLayerTestParams( -1.f, 1.f, 1.313262f ),
236+
CCrossEntropyLossLayerTestParams( 1.f, 1.f, 0.313262f ),
237+
CCrossEntropyLossLayerTestParams( 1.f, -1.f, 1.313262f ),
238+
CCrossEntropyLossLayerTestParams( 1.f, 3.1415926f, 4.2306e-2f ),
239+
CCrossEntropyLossLayerTestParams( -1.f, -0.451f, 0.49286f ),
240+
CCrossEntropyLossLayerTestParams( 1.f, 10000.f, 0.f ),
241+
CCrossEntropyLossLayerTestParams( -1.f, -10000.f, 0.f )
242+
)
243+
);

0 commit comments

Comments
 (0)