mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-24 03:00:14 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

Browse Source
This commit is contained in:
Alexander Alekhin 2019-06-15 16:25:11 +00:00
commit 66d7956e67
43 changed files with 428 additions and 1716 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
CMakeLists.txt
View File

@ -946,7 +946,8 @@ if(CV_TRACE)
include(cmake/OpenCVDetectTrace.cmake)
endif()
ocv_cmake_hook(POST_DETECT_DEPENDECIES)
ocv_cmake_hook(POST_DETECT_DEPENDECIES) # typo, deprecated (2019-06)
ocv_cmake_hook(POST_DETECT_DEPENDENCIES)
# ----------------------------------------------------------------------------
# Solution folders:

6
modules/calib3d/src/undistort.cpp
View File

@ -482,6 +482,12 @@ static void cvUndistortPointsInternal( const CvMat* _src, CvMat* _dst, const CvM
break;
double r2 = x*x + y*y;
double icdist = (1 + ((k[7]*r2 + k[6])*r2 + k[5])*r2)/(1 + ((k[4]*r2 + k[1])*r2 + k[0])*r2);
if (icdist < 0) // test: undistortPoints.regression_14583
{
x = (u - cx)*ifx;
y = (v - cy)*ify;
break;
}
double deltaX = 2*k[2]*x*y + k[3]*(r2 + 2*x*x)+ k[8]*r2+k[9]*r2*r2;
double deltaY = k[2]*(r2 + 2*y*y) + 2*k[3]*x*y+ k[10]*r2+k[11]*r2*r2;
x = (x0 - deltaX)*icdist;

26
modules/calib3d/test/test_undistort_points.cpp
View File

@ -119,4 +119,30 @@ TEST(Calib3d_Undistort, stop_criteria)
ASSERT_LE(obtainedError, maxError);
}
TEST(undistortPoints, regression_14583)
{
const int col = 720;
// const int row = 540;
float camera_matrix_value[] = {
437.8995f, 0.0f, 342.9241f,
0.0f, 438.8216f, 273.7163f,
0.0f, 0.0f, 1.0f
};
cv::Mat camera_interior(3, 3, CV_32F, camera_matrix_value);
float camera_distort_value[] = {-0.34329f, 0.11431f, 0.0f, 0.0f, -0.017375f};
cv::Mat camera_distort(1, 5, CV_32F, camera_distort_value);
float distort_points_value[] = {col, 0.};
cv::Mat distort_pt(1, 1, CV_32FC2, distort_points_value);
cv::Mat undistort_pt;
cv::undistortPoints(distort_pt, undistort_pt, camera_interior,
camera_distort, cv::Mat(), camera_interior);
EXPECT_NEAR(distort_pt.at<Vec2f>(0)[0], undistort_pt.at<Vec2f>(0)[0], col / 2)
<< "distort point: " << distort_pt << std::endl
<< "undistort point: " << undistort_pt;
}
}} // namespace

164
modules/core/include/opencv2/core/hal/intrin_avx512.hpp
View File

@ -2044,7 +2044,7 @@ inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b
b = v_uint64x8(_mm512_permutex2var_epi64(ab0, mask1, ab1));
}
inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g, v_uint8x64& r )
inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& a, v_uint8x64& b, v_uint8x64& c )
{
__m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
@ -2062,22 +2062,22 @@ inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g,
77, 74, 71, 68, 65, 127, 124, 121, 118, 115, 112, 109, 106, 103, 100, 97,
94, 91, 88, 85, 82, 79, 76, 73, 70, 67, 64, 61, 58, 55, 52, 49,
46, 43, 40, 37, 34, 31, 28, 25, 22, 19, 16, 13, 10, 7, 4, 1), bgr2);
b = v_uint8x64(_mm512_mask_compress_epi8(r12b2, 0xffffffffffe00000, r0b01));
g = v_uint8x64(_mm512_mask_compress_epi8(b1g12, 0x2492492492492492, bgr0));
r = v_uint8x64(_mm512_mask_expand_epi8(r0b01, 0xffffffffffe00000, r12b2));
a = v_uint8x64(_mm512_mask_compress_epi8(r12b2, 0xffffffffffe00000, r0b01));
b = v_uint8x64(_mm512_mask_compress_epi8(b1g12, 0x2492492492492492, bgr0));
c = v_uint8x64(_mm512_mask_expand_epi8(r0b01, 0xffffffffffe00000, r12b2));
#elif CV_AVX_512VBMI
__m512i b0g0b1 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr1, bgr0);
__m512i g1r1g2 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr2, bgr1);
__m512i r2b2r0 = _mm512_mask_blend_epi8(0xb6db6db6db6db6db, bgr0, bgr2);
b = v_uint8x64(_mm512_permutex2var_epi8(b0g0b1, _v512_set_epu8(125, 122, 119, 116, 113, 110, 107, 104, 101, 98, 95, 92, 89, 86, 83, 80,
a = v_uint8x64(_mm512_permutex2var_epi8(b0g0b1, _v512_set_epu8(125, 122, 119, 116, 113, 110, 107, 104, 101, 98, 95, 92, 89, 86, 83, 80,
77, 74, 71, 68, 65, 63, 61, 60, 58, 57, 55, 54, 52, 51, 49, 48,
46, 45, 43, 42, 40, 39, 37, 36, 34, 33, 31, 30, 28, 27, 25, 24,
23, 21, 20, 18, 17, 15, 14, 12, 11, 9, 8, 6, 5, 3, 2, 0), bgr2));
g = v_uint8x64(_mm512_permutex2var_epi8(g1r1g2, _v512_set_epu8( 63, 61, 60, 58, 57, 55, 54, 52, 51, 49, 48, 46, 45, 43, 42, 40,
b = v_uint8x64(_mm512_permutex2var_epi8(g1r1g2, _v512_set_epu8( 63, 61, 60, 58, 57, 55, 54, 52, 51, 49, 48, 46, 45, 43, 42, 40,
39, 37, 36, 34, 33, 31, 30, 28, 27, 25, 24, 23, 21, 20, 18, 17,
15, 14, 12, 11, 9, 8, 6, 5, 3, 2, 0, 126, 123, 120, 117, 114,
111, 108, 105, 102, 99, 96, 93, 90, 87, 84, 81, 78, 75, 72, 69, 66), bgr0));
r = v_uint8x64(_mm512_permutex2var_epi8(r2b2r0, _v512_set_epu8( 63, 60, 57, 54, 51, 48, 45, 42, 39, 36, 33, 30, 27, 24, 21, 18,
c = v_uint8x64(_mm512_permutex2var_epi8(r2b2r0, _v512_set_epu8( 63, 60, 57, 54, 51, 48, 45, 42, 39, 36, 33, 30, 27, 24, 21, 18,
15, 12, 9, 6, 3, 0, 125, 122, 119, 116, 113, 110, 107, 104, 101, 98,
95, 92, 89, 86, 83, 80, 77, 74, 71, 68, 65, 62, 59, 56, 53, 50,
47, 44, 41, 38, 35, 32, 29, 26, 23, 20, 17, 14, 11, 8, 5, 2), bgr1));
@ -2092,13 +2092,13 @@ inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g,
__m512i r0b1 = _mm512_permutex2var_epi16(bgr1, _v512_set_epu16(42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 29, 26, 23, 20, 17,
14, 11, 8, 5, 2, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43), g20r0);
__m512i g1r1 = _mm512_alignr_epi32(r12b2, g20r0, 11);
b = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, b0g0, r0b1));
r = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, r0b1, g1r1));
g = v_uint8x64(_mm512_shuffle_epi8(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1r1, b0g0), _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001)));
a = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, b0g0, r0b1));
c = v_uint8x64(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, r0b1, g1r1));
b = v_uint8x64(_mm512_shuffle_epi8(_mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1r1, b0g0), _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001)));
#endif
}
inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& b, v_uint16x32& g, v_uint16x32& r )
inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& a, v_uint16x32& b, v_uint16x32& c )
{
__m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
@ -2110,13 +2110,13 @@ inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& b, v_uint16x32&
__m512i r12b2 = _mm512_permutex2var_epi16(bgr1, mask0, bgr2);
__m512i g20r0 = _mm512_permutex2var_epi16(bgr2, mask0, bgr0);
b = v_uint16x32(_mm512_mask_blend_epi32(0xf800, b01g1, r12b2));
g = v_uint16x32(_mm512_permutex2var_epi16(bgr1, _v512_set_epu16(42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 29, 26, 23, 20, 17,
a = v_uint16x32(_mm512_mask_blend_epi32(0xf800, b01g1, r12b2));
b = v_uint16x32(_mm512_permutex2var_epi16(bgr1, _v512_set_epu16(42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 29, 26, 23, 20, 17,
14, 11, 8, 5, 2, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43), g20r0));
r = v_uint16x32(_mm512_alignr_epi32(r12b2, g20r0, 11));
c = v_uint16x32(_mm512_alignr_epi32(r12b2, g20r0, 11));
}
inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& b, v_uint32x16& g, v_uint32x16& r )
inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& a, v_uint32x16& b, v_uint32x16& c )
{
__m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
@ -2127,12 +2127,12 @@ inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& b, v_uint32x1
__m512i g12b2 = _mm512_permutex2var_epi32(bgr1, mask0, bgr2);
__m512i r20g0 = _mm512_permutex2var_epi32(bgr2, mask0, bgr0);
b = v_uint32x16(_mm512_mask_blend_epi32(0xf800, b01r1, g12b2));
g = v_uint32x16(_mm512_alignr_epi32(g12b2, r20g0, 11));
r = v_uint32x16(_mm512_permutex2var_epi32(bgr1, _v512_set_epu32(21, 20, 19, 18, 17, 16, 13, 10, 7, 4, 1, 26, 25, 24, 23, 22), r20g0));
a = v_uint32x16(_mm512_mask_blend_epi32(0xf800, b01r1, g12b2));
b = v_uint32x16(_mm512_alignr_epi32(g12b2, r20g0, 11));
c = v_uint32x16(_mm512_permutex2var_epi32(bgr1, _v512_set_epu32(21, 20, 19, 18, 17, 16, 13, 10, 7, 4, 1, 26, 25, 24, 23, 22), r20g0));
}
inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& b, v_uint64x8& g, v_uint64x8& r )
inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b, v_uint64x8& c )
{
__m512i bgr0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgr1 = _mm512_loadu_si512((const __m512i*)(ptr + 8));
@ -2143,12 +2143,12 @@ inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& b, v_uint64x8& g
__m512i r12b2 = _mm512_permutex2var_epi64(bgr1, mask0, bgr2);
__m512i g20r0 = _mm512_permutex2var_epi64(bgr2, mask0, bgr0);
b = v_uint64x8(_mm512_mask_blend_epi64(0xc0, b01g1, r12b2));
r = v_uint64x8(_mm512_alignr_epi64(r12b2, g20r0, 6));
g = v_uint64x8(_mm512_permutex2var_epi64(bgr1, _v512_set_epu64(10, 9, 8, 5, 2, 13, 12, 11), g20r0));
a = v_uint64x8(_mm512_mask_blend_epi64(0xc0, b01g1, r12b2));
c = v_uint64x8(_mm512_alignr_epi64(r12b2, g20r0, 6));
b = v_uint64x8(_mm512_permutex2var_epi64(bgr1, _v512_set_epu64(10, 9, 8, 5, 2, 13, 12, 11), g20r0));
}
inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g, v_uint8x64& r, v_uint8x64& a )
inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& a, v_uint8x64& b, v_uint8x64& c, v_uint8x64& d )
{
__m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 64));
@ -2170,10 +2170,10 @@ inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g,
__m512i br23 = _mm512_permutex2var_epi8(bgra2, mask0, bgra3);
__m512i ga23 = _mm512_permutex2var_epi8(bgra2, mask1, bgra3);
b = v_uint8x64(_mm512_permutex2var_epi8(br01, mask0, br23));
r = v_uint8x64(_mm512_permutex2var_epi8(br01, mask1, br23));
g = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask0, ga23));
a = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask1, ga23));
a = v_uint8x64(_mm512_permutex2var_epi8(br01, mask0, br23));
c = v_uint8x64(_mm512_permutex2var_epi8(br01, mask1, br23));
b = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask0, ga23));
d = v_uint8x64(_mm512_permutex2var_epi8(ga01, mask1, ga23));
#else
__m512i mask = _mm512_set4_epi32(0x0f0b0703, 0x0e0a0602, 0x0d090501, 0x0c080400);
__m512i b0g0r0a0 = _mm512_shuffle_epi8(bgra0, mask);
@ -2189,14 +2189,14 @@ inline void v_load_deinterleave( const uchar* ptr, v_uint8x64& b, v_uint8x64& g,
__m512i br23 = _mm512_permutex2var_epi32(b2g2r2a2, mask0, b3g3r3a3);
__m512i ga23 = _mm512_permutex2var_epi32(b2g2r2a2, mask1, b3g3r3a3);
b = v_uint8x64(_mm512_permutex2var_epi32(br01, mask0, br23));
r = v_uint8x64(_mm512_permutex2var_epi32(br01, mask1, br23));
g = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask0, ga23));
a = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask1, ga23));
a = v_uint8x64(_mm512_permutex2var_epi32(br01, mask0, br23));
c = v_uint8x64(_mm512_permutex2var_epi32(br01, mask1, br23));
b = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask0, ga23));
d = v_uint8x64(_mm512_permutex2var_epi32(ga01, mask1, ga23));
#endif
}
inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& b, v_uint16x32& g, v_uint16x32& r, v_uint16x32& a )
inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& a, v_uint16x32& b, v_uint16x32& c, v_uint16x32& d )
{
__m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 32));
@ -2213,13 +2213,13 @@ inline void v_load_deinterleave( const ushort* ptr, v_uint16x32& b, v_uint16x32&
__m512i br23 = _mm512_permutex2var_epi16(bgra2, mask0, bgra3);
__m512i ga23 = _mm512_permutex2var_epi16(bgra2, mask1, bgra3);
b = v_uint16x32(_mm512_permutex2var_epi16(br01, mask0, br23));
r = v_uint16x32(_mm512_permutex2var_epi16(br01, mask1, br23));
g = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask0, ga23));
a = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask1, ga23));
a = v_uint16x32(_mm512_permutex2var_epi16(br01, mask0, br23));
c = v_uint16x32(_mm512_permutex2var_epi16(br01, mask1, br23));
b = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask0, ga23));
d = v_uint16x32(_mm512_permutex2var_epi16(ga01, mask1, ga23));
}
inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& b, v_uint32x16& g, v_uint32x16& r, v_uint32x16& a )
inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& a, v_uint32x16& b, v_uint32x16& c, v_uint32x16& d )
{
__m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 16));
@ -2234,13 +2234,13 @@ inline void v_load_deinterleave( const unsigned* ptr, v_uint32x16& b, v_uint32x1
__m512i br23 = _mm512_permutex2var_epi32(bgra2, mask0, bgra3);
__m512i ga23 = _mm512_permutex2var_epi32(bgra2, mask1, bgra3);
b = v_uint32x16(_mm512_permutex2var_epi32(br01, mask0, br23));
r = v_uint32x16(_mm512_permutex2var_epi32(br01, mask1, br23));
g = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask0, ga23));
a = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask1, ga23));
a = v_uint32x16(_mm512_permutex2var_epi32(br01, mask0, br23));
c = v_uint32x16(_mm512_permutex2var_epi32(br01, mask1, br23));
b = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask0, ga23));
d = v_uint32x16(_mm512_permutex2var_epi32(ga01, mask1, ga23));
}
inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& b, v_uint64x8& g, v_uint64x8& r, v_uint64x8& a )
inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& a, v_uint64x8& b, v_uint64x8& c, v_uint64x8& d )
{
__m512i bgra0 = _mm512_loadu_si512((const __m512i*)ptr);
__m512i bgra1 = _mm512_loadu_si512((const __m512i*)(ptr + 8));
@ -2255,10 +2255,10 @@ inline void v_load_deinterleave( const uint64* ptr, v_uint64x8& b, v_uint64x8& g
__m512i br23 = _mm512_permutex2var_epi64(bgra2, mask0, bgra3);
__m512i ga23 = _mm512_permutex2var_epi64(bgra2, mask1, bgra3);
b = v_uint64x8(_mm512_permutex2var_epi64(br01, mask0, br23));
r = v_uint64x8(_mm512_permutex2var_epi64(br01, mask1, br23));
g = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask0, ga23));
a = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask1, ga23));
a = v_uint64x8(_mm512_permutex2var_epi64(br01, mask0, br23));
c = v_uint64x8(_mm512_permutex2var_epi64(br01, mask1, br23));
b = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask0, ga23));
d = v_uint64x8(_mm512_permutex2var_epi64(ga01, mask1, ga23));
}
///////////////////////////// store interleave /////////////////////////////////////
@ -2351,7 +2351,7 @@ inline void v_store_interleave( uint64* ptr, const v_uint64x8& x, const v_uint64
}
}
inline void v_store_interleave( uchar* ptr, const v_uint8x64& b, const v_uint8x64& g, const v_uint8x64& r,
inline void v_store_interleave( uchar* ptr, const v_uint8x64& a, const v_uint8x64& b, const v_uint8x64& c,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
#if CV_AVX_512VBMI
@ -2367,18 +2367,18 @@ inline void v_store_interleave( uchar* ptr, const v_uint8x64& b, const v_uint8x6
122, 58, 100, 121, 57, 99, 120, 56, 98, 119, 55, 97, 118, 54, 96, 117,
53, 95, 116, 52, 94, 115, 51, 93, 114, 50, 92, 113, 49, 91, 112, 48,
90, 111, 47, 89, 110, 46, 88, 109, 45, 87, 108, 44, 86, 107, 43, 85);
__m512i r2g0r0 = _mm512_permutex2var_epi8(g.val, mask0, r.val);
__m512i b0r1b1 = _mm512_permutex2var_epi8(b.val, mask1, r.val);
__m512i g1b2g2 = _mm512_permutex2var_epi8(b.val, mask2, g.val);
__m512i r2g0r0 = _mm512_permutex2var_epi8(b.val, mask0, c.val);
__m512i b0r1b1 = _mm512_permutex2var_epi8(a.val, mask1, c.val);
__m512i g1b2g2 = _mm512_permutex2var_epi8(a.val, mask2, b.val);
__m512i bgr0 = _mm512_mask_blend_epi8(0x9249249249249249, r2g0r0, b0r1b1);
__m512i bgr1 = _mm512_mask_blend_epi8(0x9249249249249249, b0r1b1, g1b2g2);
__m512i bgr2 = _mm512_mask_blend_epi8(0x9249249249249249, g1b2g2, r2g0r0);
#else
__m512i g1g0 = _mm512_shuffle_epi8(g.val, _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001));
__m512i b0g0 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, b.val, g1g0);
__m512i r0b1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, r.val, b.val);
__m512i g1r1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1g0, r.val);
__m512i g1g0 = _mm512_shuffle_epi8(b.val, _mm512_set4_epi32(0x0e0f0c0d, 0x0a0b0809, 0x06070405, 0x02030001));
__m512i b0g0 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, a.val, g1g0);
__m512i r0b1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, c.val, a.val);
__m512i g1r1 = _mm512_mask_blend_epi8(0xAAAAAAAAAAAAAAAA, g1g0, c.val);
__m512i mask0 = _v512_set_epu16(42, 10, 31, 41, 9, 30, 40, 8, 29, 39, 7, 28, 38, 6, 27, 37,
5, 26, 36, 4, 25, 35, 3, 24, 34, 2, 23, 33, 1, 22, 32, 0);
@ -2415,7 +2415,7 @@ inline void v_store_interleave( uchar* ptr, const v_uint8x64& b, const v_uint8x6
}
}
inline void v_store_interleave( ushort* ptr, const v_uint16x32& b, const v_uint16x32& g, const v_uint16x32& r,
inline void v_store_interleave( ushort* ptr, const v_uint16x32& a, const v_uint16x32& b, const v_uint16x32& c,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
__m512i mask0 = _v512_set_epu16(42, 10, 31, 41, 9, 30, 40, 8, 29, 39, 7, 28, 38, 6, 27, 37,
@ -2424,9 +2424,9 @@ inline void v_store_interleave( ushort* ptr, const v_uint16x32& b, const v_uint1
47, 36, 15, 46, 35, 14, 45, 34, 13, 44, 33, 12, 43, 32, 11, 42);
__m512i mask2 = _v512_set_epu16(63, 31, 20, 62, 30, 19, 61, 29, 18, 60, 28, 17, 59, 27, 16, 58,
26, 15, 57, 25, 14, 56, 24, 13, 55, 23, 12, 54, 22, 11, 53, 21);
__m512i b0g0b2 = _mm512_permutex2var_epi16(b.val, mask0, g.val);
__m512i r1b1r0 = _mm512_permutex2var_epi16(b.val, mask1, r.val);
__m512i g2r2g1 = _mm512_permutex2var_epi16(g.val, mask2, r.val);
__m512i b0g0b2 = _mm512_permutex2var_epi16(a.val, mask0, b.val);
__m512i r1b1r0 = _mm512_permutex2var_epi16(a.val, mask1, c.val);
__m512i g2r2g1 = _mm512_permutex2var_epi16(b.val, mask2, c.val);
__m512i bgr0 = _mm512_mask_blend_epi16(0x24924924, b0g0b2, r1b1r0);
__m512i bgr1 = _mm512_mask_blend_epi16(0x24924924, r1b1r0, g2r2g1);
@ -2452,15 +2452,15 @@ inline void v_store_interleave( ushort* ptr, const v_uint16x32& b, const v_uint1
}
}
inline void v_store_interleave( unsigned* ptr, const v_uint32x16& b, const v_uint32x16& g, const v_uint32x16& r,
inline void v_store_interleave( unsigned* ptr, const v_uint32x16& a, const v_uint32x16& b, const v_uint32x16& c,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
__m512i mask0 = _v512_set_epu32(26, 31, 15, 25, 30, 14, 24, 29, 13, 23, 28, 12, 22, 27, 11, 21);
__m512i mask1 = _v512_set_epu32(31, 10, 25, 30, 9, 24, 29, 8, 23, 28, 7, 22, 27, 6, 21, 26);
__m512i g1b2g2 = _mm512_permutex2var_epi32(b.val, mask0, g.val);
__m512i r2r1b1 = _mm512_permutex2var_epi32(b.val, mask1, r.val);
__m512i g1b2g2 = _mm512_permutex2var_epi32(a.val, mask0, b.val);
__m512i r2r1b1 = _mm512_permutex2var_epi32(a.val, mask1, c.val);
__m512i bgr0 = _mm512_mask_expand_epi32(_mm512_mask_expand_epi32(_mm512_maskz_expand_epi32(0x9249, b.val), 0x2492, g.val), 0x4924, r.val);
__m512i bgr0 = _mm512_mask_expand_epi32(_mm512_mask_expand_epi32(_mm512_maskz_expand_epi32(0x9249, a.val), 0x2492, b.val), 0x4924, c.val);
__m512i bgr1 = _mm512_mask_blend_epi32(0x9249, r2r1b1, g1b2g2);
__m512i bgr2 = _mm512_mask_blend_epi32(0x9249, g1b2g2, r2r1b1);
@ -2484,15 +2484,15 @@ inline void v_store_interleave( unsigned* ptr, const v_uint32x16& b, const v_uin
}
}
inline void v_store_interleave( uint64* ptr, const v_uint64x8& b, const v_uint64x8& g, const v_uint64x8& r,
inline void v_store_interleave( uint64* ptr, const v_uint64x8& a, const v_uint64x8& b, const v_uint64x8& c,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
__m512i mask0 = _v512_set_epu64( 5, 12, 7, 4, 11, 6, 3, 10);
__m512i mask1 = _v512_set_epu64(15, 7, 4, 14, 6, 3, 13, 5);
__m512i r1b1b2 = _mm512_permutex2var_epi64(b.val, mask0, r.val);
__m512i g2r2g1 = _mm512_permutex2var_epi64(g.val, mask1, r.val);
__m512i r1b1b2 = _mm512_permutex2var_epi64(a.val, mask0, c.val);
__m512i g2r2g1 = _mm512_permutex2var_epi64(b.val, mask1, c.val);
__m512i bgr0 = _mm512_mask_expand_epi64(_mm512_mask_expand_epi64(_mm512_maskz_expand_epi64(0x49, b.val), 0x92, g.val), 0x24, r.val);
__m512i bgr0 = _mm512_mask_expand_epi64(_mm512_mask_expand_epi64(_mm512_maskz_expand_epi64(0x49, a.val), 0x92, b.val), 0x24, c.val);
__m512i bgr1 = _mm512_mask_blend_epi64(0xdb, g2r2g1, r1b1b2);
__m512i bgr2 = _mm512_mask_blend_epi64(0xdb, r1b1b2, g2r2g1);
@ -2516,13 +2516,13 @@ inline void v_store_interleave( uint64* ptr, const v_uint64x8& b, const v_uint64
}
}
inline void v_store_interleave( uchar* ptr, const v_uint8x64& b, const v_uint8x64& g,
const v_uint8x64& r, const v_uint8x64& a,
inline void v_store_interleave( uchar* ptr, const v_uint8x64& a, const v_uint8x64& b,
const v_uint8x64& c, const v_uint8x64& d,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
v_uint8x64 br01, br23, ga01, ga23;
v_zip(b, r, br01, br23);
v_zip(g, a, ga01, ga23);
v_zip(a, c, br01, br23);
v_zip(b, d, ga01, ga23);
v_uint8x64 bgra0, bgra1, bgra2, bgra3;
v_zip(br01, ga01, bgra0, bgra1);
v_zip(br23, ga23, bgra2, bgra3);
@ -2550,13 +2550,13 @@ inline void v_store_interleave( uchar* ptr, const v_uint8x64& b, const v_uint8x6
}
}
inline void v_store_interleave( ushort* ptr, const v_uint16x32& b, const v_uint16x32& g,
const v_uint16x32& r, const v_uint16x32& a,
inline void v_store_interleave( ushort* ptr, const v_uint16x32& a, const v_uint16x32& b,
const v_uint16x32& c, const v_uint16x32& d,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
v_uint16x32 br01, br23, ga01, ga23;
v_zip(b, r, br01, br23);
v_zip(g, a, ga01, ga23);
v_zip(a, c, br01, br23);
v_zip(b, d, ga01, ga23);
v_uint16x32 bgra0, bgra1, bgra2, bgra3;
v_zip(br01, ga01, bgra0, bgra1);
v_zip(br23, ga23, bgra2, bgra3);
@ -2584,13 +2584,13 @@ inline void v_store_interleave( ushort* ptr, const v_uint16x32& b, const v_uint1
}
}
inline void v_store_interleave( unsigned* ptr, const v_uint32x16& b, const v_uint32x16& g,
const v_uint32x16& r, const v_uint32x16& a,
inline void v_store_interleave( unsigned* ptr, const v_uint32x16& a, const v_uint32x16& b,
const v_uint32x16& c, const v_uint32x16& d,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
v_uint32x16 br01, br23, ga01, ga23;
v_zip(b, r, br01, br23);
v_zip(g, a, ga01, ga23);
v_zip(a, c, br01, br23);
v_zip(b, d, ga01, ga23);
v_uint32x16 bgra0, bgra1, bgra2, bgra3;
v_zip(br01, ga01, bgra0, bgra1);
v_zip(br23, ga23, bgra2, bgra3);
@ -2618,13 +2618,13 @@ inline void v_store_interleave( unsigned* ptr, const v_uint32x16& b, const v_uin
}
}
inline void v_store_interleave( uint64* ptr, const v_uint64x8& b, const v_uint64x8& g,
const v_uint64x8& r, const v_uint64x8& a,
inline void v_store_interleave( uint64* ptr, const v_uint64x8& a, const v_uint64x8& b,
const v_uint64x8& c, const v_uint64x8& d,
hal::StoreMode mode=hal::STORE_UNALIGNED )
{
v_uint64x8 br01, br23, ga01, ga23;
v_zip(b, r, br01, br23);
v_zip(g, a, ga01, ga23);
v_zip(a, c, br01, br23);
v_zip(b, d, ga01, ga23);
v_uint64x8 bgra0, bgra1, bgra2, bgra3;
v_zip(br01, ga01, bgra0, bgra1);
v_zip(br23, ga23, bgra2, bgra3);

5
modules/dnn/include/opencv2/dnn/all_layers.hpp
View File

@ -492,10 +492,7 @@ CV__DNN_INLINE_NS_BEGIN
class CV_EXPORTS CropLayer : public Layer
{
public:
int startAxis;
std::vector<int> offset;
static Ptr<CropLayer> create(const LayerParams &params);
static Ptr<Layer> create(const LayerParams &params);
};
class CV_EXPORTS EltwiseLayer : public Layer

97
modules/dnn/src/dnn.cpp
View File

@ -735,20 +735,9 @@ struct DataLayer : public Layer
}
biases->set(biasesVec);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer = InferenceEngine::Builder::ScaleShiftLayer(name);
addConstantData("weights", weights, ieLayer);
addConstantData("biases", biases, ieLayer);
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "ScaleShift";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ScaleShiftLayer> ieLayer(new InferenceEngine::ScaleShiftLayer(lp));
ieLayer->_weights = weights;
ieLayer->_biases = biases;
#endif
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
@ -1488,11 +1477,7 @@ struct Net::Impl
if (layerNet != ieInpNode->net)
{
// layerNet is empty or nodes are from different graphs.
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
ieInpNode->net->addOutput(ieInpNode->layer.getName());
#else
ieInpNode->net->addOutput(ieInpNode->layer->name);
#endif
}
}
}
@ -1642,25 +1627,6 @@ struct Net::Impl
}
}
#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2018R5)
// The same blobs wrappers cannot be shared between two Inference Engine
// networks because of explicit references between layers and blobs.
// So we need to rewrap all the external blobs.
for (int i = 0; i < ld.inputBlobsId.size(); ++i)
{
LayerPin inPin = ld.inputBlobsId[i];
auto it = netBlobsWrappers.find(inPin);
if (it == netBlobsWrappers.end())
{
ld.inputBlobsWrappers[i] = InfEngineBackendWrapper::create(ld.inputBlobsWrappers[i]);
netBlobsWrappers[inPin] = ld.inputBlobsWrappers[i];
}
else
ld.inputBlobsWrappers[i] = it->second;
}
netBlobsWrappers[LayerPin(ld.id, 0)] = ld.outputBlobsWrappers[0];
#endif // IE < R5
Ptr<BackendNode> node;
if (!net.empty())
{
@ -1691,7 +1657,6 @@ struct Net::Impl
ieNode->net = net;
// Convert weights in FP16 for specific targets.
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
if ((preferableTarget == DNN_TARGET_OPENCL_FP16 ||
preferableTarget == DNN_TARGET_MYRIAD ||
preferableTarget == DNN_TARGET_FPGA) && !fused)
@ -1733,47 +1698,6 @@ struct Net::Impl
net->addBlobs(ld.inputBlobsWrappers);
net->addBlobs(ld.outputBlobsWrappers);
addInfEngineNetOutputs(ld);
#else // IE >= R5
auto weightableLayer = std::dynamic_pointer_cast<InferenceEngine::WeightableLayer>(ieNode->layer);
if ((preferableTarget == DNN_TARGET_OPENCL_FP16 ||
preferableTarget == DNN_TARGET_MYRIAD ||
preferableTarget == DNN_TARGET_FPGA) && !fused)
{
ieNode->layer->precision = InferenceEngine::Precision::FP16;
if (weightableLayer)
{
if (weightableLayer->_weights)
weightableLayer->_weights = convertFp16(weightableLayer->_weights);
if (weightableLayer->_biases)
weightableLayer->_biases = convertFp16(weightableLayer->_biases);
}
else
{
for (const auto& weights : {"weights", "biases"})
{
auto it = ieNode->layer->blobs.find(weights);
if (it != ieNode->layer->blobs.end())
it->second = convertFp16(it->second);
}
}
}
if (weightableLayer)
{
if (weightableLayer->_weights)
weightableLayer->blobs["weights"] = weightableLayer->_weights;
if (weightableLayer->_biases)
weightableLayer->blobs["biases"] = weightableLayer->_biases;
}
ieNode->connect(ld.inputBlobsWrappers, ld.outputBlobsWrappers);
net->addBlobs(ld.inputBlobsWrappers);
net->addBlobs(ld.outputBlobsWrappers);
if (!fused)
net->addLayer(ieNode->layer);
addInfEngineNetOutputs(ld);
#endif // IE >= R5
}
// Initialize all networks.
@ -1795,23 +1719,6 @@ struct Net::Impl
if (!ieNode->net->isInitialized())
{
#if INF_ENGINE_VER_MAJOR_EQ(INF_ENGINE_RELEASE_2018R4)
// For networks which is built in runtime we need to specify a
// version of it's hyperparameters.
std::string versionTrigger = "<net name=\"TestInput\" version=\"3\" batch=\"1\">"
"<layers>"
"<layer name=\"data\" type=\"Input\" precision=\"FP32\" id=\"0\">"
"<output>"
"<port id=\"0\">"
"<dim>1</dim>"
"</port>"
"</output>"
"</layer>"
"</layers>"
"</net>";
InferenceEngine::CNNNetReader reader;
reader.ReadNetwork(versionTrigger.data(), versionTrigger.size());
#endif
ieNode->net->init(preferableTarget);
ld.skip = false;
}
@ -2693,11 +2600,7 @@ Net Net::readFromModelOptimizer(const String& xml, const String& bin)
Net cvNet;
cvNet.setInputsNames(inputsNames);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
Ptr<InfEngineBackendNode> backendNode(new InfEngineBackendNode(InferenceEngine::Builder::Layer("")));
#else
Ptr<InfEngineBackendNode> backendNode(new InfEngineBackendNode(0));
#endif
backendNode->net = Ptr<InfEngineBackendNet>(new InfEngineBackendNet(ieNet));
for (auto& it : ieNet.getOutputsInfo())
{

19
modules/dnn/src/layers/batch_norm_layer.cpp
View File

@ -351,31 +351,16 @@ public:
}
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer = InferenceEngine::Builder::ScaleShiftLayer(name);
const size_t numChannels = weights_.total();
addConstantData("weights", wrapToInfEngineBlob(weights_, {numChannels}, InferenceEngine::Layout::C), ieLayer);
addConstantData("biases", wrapToInfEngineBlob(bias_, {numChannels}, InferenceEngine::Layout::C), ieLayer);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "ScaleShift";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ScaleShiftLayer> ieLayer(new InferenceEngine::ScaleShiftLayer(lp));
const size_t numChannels = weights_.total();
ieLayer->_weights = wrapToInfEngineBlob(weights_, {numChannels}, InferenceEngine::Layout::C);
ieLayer->_biases = wrapToInfEngineBlob(bias_, {numChannels}, InferenceEngine::Layout::C);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

21
modules/dnn/src/layers/blank_layer.cpp
View File

@ -107,12 +107,12 @@ public:
inputs[i].copyTo(outputs[i]);
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
CV_Assert(!input->dims.empty());
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
if (preferableTarget == DNN_TARGET_MYRIAD)
@ -122,7 +122,7 @@ public:
else
{
ieLayer.setType("Split");
ieLayer.getParameters()["axis"] = (size_t)0;
ieLayer.getParameters()["axis"] = input->dims.size() - 1;
ieLayer.getParameters()["out_sizes"] = input->dims[0];
}
std::vector<size_t> shape(input->dims);
@ -130,21 +130,8 @@ public:
ieLayer.setInputPorts({InferenceEngine::Port(shape)});
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Split";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::SplitLayer> ieLayer(new InferenceEngine::SplitLayer(lp));
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
ieLayer->params["axis"] = format("%d", (int)input->dims.size() - 1);
ieLayer->params["out_sizes"] = format("%d", (int)input->dims[0]);
#endif
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
};
Ptr<Layer> BlankLayer::create(const LayerParams& params)

16
modules/dnn/src/layers/concat_layer.cpp
View File

@ -310,29 +310,17 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::Builder::ConcatLayer ieLayer(name);
ieLayer.setAxis(clamp(axis, input->dims.size()));
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(inputs.size()));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Concat";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ConcatLayer> ieLayer(new InferenceEngine::ConcatLayer(lp));
ieLayer->_axis = clamp(axis, input->dims.size());
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
};
Ptr<ConcatLayer> ConcatLayer::create(const LayerParams& params)

98
modules/dnn/src/layers/convolution_layer.cpp
View File

@ -253,8 +253,7 @@ public:
{
if (kernel_size.size() == 3)
return preferableTarget == DNN_TARGET_CPU;
return INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R4) ||
(preferableTarget != DNN_TARGET_MYRIAD || dilation.width == dilation.height);
return (preferableTarget != DNN_TARGET_MYRIAD || dilation.width == dilation.height);
}
else
#endif
@ -534,9 +533,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
CV_Assert(input->dims.size() == 4 || input->dims.size() == 5);
@ -575,7 +574,6 @@ public:
ieBiases = wrapToInfEngineBlob(biasesMat, {(size_t)outCn}, InferenceEngine::Layout::C);
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ConvolutionLayer ieLayer(name);
ieLayer.setKernel(kernel_size);
@ -595,51 +593,8 @@ public:
l.getParameters()["auto_pad"] = padMode == "VALID" ? std::string("valid") : std::string("same_upper");
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Convolution";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ConvolutionLayer> ieLayer(new InferenceEngine::ConvolutionLayer(lp));
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
ieLayer->_kernel.insert(InferenceEngine::X_AXIS, kernel.width);
ieLayer->_kernel.insert(InferenceEngine::Y_AXIS, kernel.height);
ieLayer->_stride.insert(InferenceEngine::X_AXIS, stride.width);
ieLayer->_stride.insert(InferenceEngine::Y_AXIS, stride.height);
ieLayer->_padding.insert(InferenceEngine::X_AXIS, pad.width);
ieLayer->_padding.insert(InferenceEngine::Y_AXIS, pad.height);
ieLayer->_pads_end.insert(InferenceEngine::X_AXIS, pad.width);
ieLayer->_pads_end.insert(InferenceEngine::Y_AXIS, pad.height);
ieLayer->_dilation.insert(InferenceEngine::X_AXIS, dilation.width);
ieLayer->_dilation.insert(InferenceEngine::Y_AXIS, dilation.height);
ieLayer->params["output"] = format("%d", outCn);
ieLayer->params["kernel"] = format("%d,%d,%d,%d", outCn, inpGroupCn, kernel.height, kernel.width);
ieLayer->params["pads_begin"] = format("%d,%d", pad.height, pad.width);
ieLayer->params["pads_end"] = format("%d,%d", pad.height, pad.width);
ieLayer->params["strides"] = format("%d,%d", stride.height, stride.width);
ieLayer->params["dilations"] = format("%d,%d", dilation.height, dilation.width);
#else
ieLayer->_kernel_x = kernel.width;
ieLayer->_kernel_y = kernel.height;
ieLayer->_stride_x = stride.width;
ieLayer->_stride_y = stride.height;
ieLayer->_padding_x = pad.width;
ieLayer->_padding_y = pad.height;
ieLayer->_dilation_x = dilation.width;
ieLayer->_dilation_y = dilation.height;
#endif
ieLayer->_out_depth = outCn;
ieLayer->_group = group;
ieLayer->_weights = ieWeights;
if (ieBiases)
ieLayer->_biases = ieBiases;
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
class ParallelConv : public cv::ParallelLoopBody
{
@ -1258,7 +1213,7 @@ public:
if (kernel_size.size() == 3)
CV_Error(Error::StsNotImplemented, "Unsupported deconvolution3D layer");
if (INF_ENGINE_RELEASE >= 2018050000 && (adjustPad.height || adjustPad.width))
if (adjustPad.height || adjustPad.width)
{
if (padMode.empty())
{
@ -1867,9 +1822,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> > &) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
auto ieWeights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW);
if (fusedWeights)
{
@ -1883,7 +1838,6 @@ public:
transpose(weightsMat, newWeights);
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW layout
const int group = numOutput / outGroupCn;
@ -1911,50 +1865,8 @@ public:
if (hasBias())
addConstantData("biases", wrapToInfEngineBlob(biasesMat, {(size_t)numOutput}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW layout
const int group = numOutput / outGroupCn;
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Deconvolution";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::DeconvolutionLayer> ieLayer(new InferenceEngine::DeconvolutionLayer(lp));
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
ieLayer->_kernel.insert(InferenceEngine::X_AXIS, kernel.width);
ieLayer->_kernel.insert(InferenceEngine::Y_AXIS, kernel.height);
ieLayer->_stride.insert(InferenceEngine::X_AXIS, stride.width);
ieLayer->_stride.insert(InferenceEngine::Y_AXIS, stride.height);
ieLayer->_padding.insert(InferenceEngine::X_AXIS, pad.width);
ieLayer->_padding.insert(InferenceEngine::Y_AXIS, pad.height);
ieLayer->_pads_end.insert(InferenceEngine::X_AXIS, pad.width);
ieLayer->_pads_end.insert(InferenceEngine::Y_AXIS, pad.height);
ieLayer->_dilation.insert(InferenceEngine::X_AXIS, dilation.width);
ieLayer->_dilation.insert(InferenceEngine::Y_AXIS, dilation.height);
#else
ieLayer->_kernel_x = kernel.width;
ieLayer->_kernel_y = kernel.height;
ieLayer->_stride_x = stride.width;
ieLayer->_stride_y = stride.height;
ieLayer->_padding_x = pad.width;
ieLayer->_padding_y = pad.height;
ieLayer->_dilation_x = dilation.width;
ieLayer->_dilation_y = dilation.height;
#endif
ieLayer->_out_depth = numOutput;
ieLayer->_group = group;
ieLayer->_weights = wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW);
if (hasBias())
{
ieLayer->_biases = wrapToInfEngineBlob(blobs[1], {(size_t)numOutput}, InferenceEngine::Layout::C);
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

205
modules/dnn/src/layers/crop_layer.cpp
View File

@ -1,205 +0,0 @@
/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
// Copyright (C) 2017, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of the copyright holders may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "../precomp.hpp"
#include "../op_inf_engine.hpp"
#include "layers_common.hpp"
namespace cv
{
namespace dnn
{
class CropLayerImpl CV_FINAL : public CropLayer
{
public:
CropLayerImpl(const LayerParams& params)
{
setParamsFrom(params);
startAxis = params.get<int>("axis", 2);
const DictValue *paramOffset = params.ptr("offset");
if (paramOffset)
{
for (int i = 0; i < paramOffset->size(); i++)
offset.push_back(paramOffset->get<int>(i));
}
}
virtual bool supportBackend(int backendId) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
return INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2018R5) && crop_ranges.size() == 4;
else
#endif
return backendId == DNN_BACKEND_OPENCV;
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
CV_Assert(inputs.size() == 2);
MatShape dstShape = inputs[0];
int start = clamp(startAxis, dstShape);
for (int i = start; i < dstShape.size(); i++)
{
dstShape[i] = inputs[1][i];
}
outputs.resize(1, dstShape);
return false;
}
void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE
{
std::vector<Mat> inputs;
inputs_arr.getMatVector(inputs);
CV_Assert(2 == inputs.size());
const Mat &inpBlob = inputs[0];
const Mat &inpSzBlob = inputs[1];
int dims = inpBlob.dims;
int start_axis = clamp(startAxis, dims);
std::vector<int> offset_final(dims, 0);
if (offset.size() == 1)
{
for (int i = start_axis; i < dims; i++)
offset_final[i] = offset[0];
}
else if (offset.size() > 1)
{
if ((int)offset.size() != dims - start_axis)
CV_Error(Error::StsBadArg, "number of offset values specified must be "
"equal to the number of dimensions following axis.");
for (int i = start_axis; i < dims; i++)
offset_final[i] = offset[i - start_axis];
}
crop_ranges.resize(dims);
for (int i = 0; i < start_axis; i++)
{
crop_ranges[i] = Range(0, inpBlob.size[i]);
}
for (int i = start_axis; i < dims; i++)
{
if (offset_final[i] < 0 || offset_final[i] + inpSzBlob.size[i] > inpBlob.size[i])
CV_Error(Error::StsBadArg, "invalid crop parameters or blob sizes");
crop_ranges[i] = Range(offset_final[i], offset_final[i] + inpSzBlob.size[i]);
}
}
void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE
{
CV_TRACE_FUNCTION();
CV_TRACE_ARG_VALUE(name, "name", name.c_str());
std::vector<Mat> inputs, outputs;
inputs_arr.getMatVector(inputs);
outputs_arr.getMatVector(outputs);
Mat &input = inputs[0];
input(&crop_ranges[0]).copyTo(outputs[0]);
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Crop";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CropLayer> ieLayer(new InferenceEngine::CropLayer(lp));
CV_Assert(crop_ranges.size() == 4);
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
for (int i = 0; i < 4; ++i)
{
ieLayer->axis.push_back(i);
ieLayer->offset.push_back(crop_ranges[i].start);
ieLayer->dim.push_back(crop_ranges[i].end - crop_ranges[i].start);
}
#else
ieLayer->axis.push_back(0); // batch
ieLayer->offset.push_back(crop_ranges[0].start);
ieLayer->dim.push_back(crop_ranges[0].end - crop_ranges[0].start);
ieLayer->axis.push_back(1); // channels
ieLayer->offset.push_back(crop_ranges[1].start);
ieLayer->dim.push_back(crop_ranges[1].end - crop_ranges[1].start);
ieLayer->axis.push_back(3); // height
ieLayer->offset.push_back(crop_ranges[2].start);
ieLayer->dim.push_back(crop_ranges[2].end - crop_ranges[2].start);
ieLayer->axis.push_back(2); // width
ieLayer->offset.push_back(crop_ranges[3].start);
ieLayer->dim.push_back(crop_ranges[3].end - crop_ranges[3].start);
#endif
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
return Ptr<BackendNode>();
#endif // IE < R5
}
#endif
std::vector<Range> crop_ranges;
};
Ptr<CropLayer> CropLayer::create(const LayerParams& params)
{
return Ptr<CropLayer>(new CropLayerImpl(params));
}
}
}

25
modules/dnn/src/layers/detection_output_layer.cpp
View File

@ -918,10 +918,9 @@ public:
}
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::DetectionOutputLayer ieLayer(name);
ieLayer.setNumClasses(_numClasses);
@ -939,28 +938,8 @@ public:
l.getParameters()["eta"] = std::string("1.0");
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "DetectionOutput";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["num_classes"] = format("%d", _numClasses);
ieLayer->params["share_location"] = _shareLocation ? "1" : "0";
ieLayer->params["background_label_id"] = format("%d", _backgroundLabelId);
ieLayer->params["nms_threshold"] = format("%f", _nmsThreshold);
ieLayer->params["top_k"] = format("%d", _topK);
ieLayer->params["keep_top_k"] = format("%d", _keepTopK);
ieLayer->params["eta"] = "1.0";
ieLayer->params["confidence_threshold"] = format("%f", _confidenceThreshold);
ieLayer->params["variance_encoded_in_target"] = _varianceEncodedInTarget ? "1" : "0";
ieLayer->params["code_type"] = "caffe.PriorBoxParameter." + _codeType;
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
};
float util::caffe_box_overlap(const util::NormalizedBBox& a, const util::NormalizedBBox& b)

114
modules/dnn/src/layers/elementwise_layers.cpp
View File

@ -150,22 +150,14 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer = func.initInfEngineBuilderAPI();
ieLayer.setName(this->name);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = this->name;
lp.precision = InferenceEngine::Precision::FP32;
return Ptr<BackendNode>(new InfEngineBackendNode(func.initInfEngine(lp)));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual Ptr<BackendNode> initVkCom(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
@ -364,21 +356,10 @@ struct ReLUFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::ReLULayer("").setNegativeSlope(slope);
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "ReLU";
std::shared_ptr<InferenceEngine::ReLULayer> ieLayer(new InferenceEngine::ReLULayer(lp));
ieLayer->negative_slope = slope;
ieLayer->params["negative_slope"] = format("%f", slope);
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -488,23 +469,10 @@ struct ReLU6Functor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::ClampLayer("").setMinValue(minValue).setMaxValue(maxValue);
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "Clamp";
std::shared_ptr<InferenceEngine::ClampLayer> ieLayer(new InferenceEngine::ClampLayer(lp));
ieLayer->min_value = minValue;
ieLayer->max_value = maxValue;
ieLayer->params["min"] = format("%f", minValue);
ieLayer->params["max"] = format("%f", maxValue);
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -581,19 +549,10 @@ struct TanHFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::TanHLayer("");
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "TanH";
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -670,19 +629,10 @@ struct SigmoidFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::SigmoidLayer("");
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "Sigmoid";
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -761,18 +711,10 @@ struct ELUFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::ELULayer("");
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "ELU";
return InferenceEngine::CNNLayerPtr(new InferenceEngine::CNNLayer(lp));
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -852,21 +794,10 @@ struct AbsValFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::ReLULayer("").setNegativeSlope(-1);
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "ReLU";
std::shared_ptr<InferenceEngine::ReLULayer> ieLayer(new InferenceEngine::ReLULayer(lp));
ieLayer->negative_slope = -1;
ieLayer->params["negative_slope"] = "-1.0";
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -922,18 +853,10 @@ struct BNLLFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
CV_Error(Error::StsNotImplemented, "");
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
CV_Error(Error::StsNotImplemented, "BNLL");
return InferenceEngine::CNNLayerPtr();
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -1046,34 +969,12 @@ struct PowerFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
return InferenceEngine::Builder::PowerLayer("").setPower(power)
.setScale(scale)
.setShift(shift);
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
if (power == 1.0f && scale == 1.0f && shift == 0.0f)
{
// It looks like there is a bug in Inference Engine for DNN_TARGET_OPENCL and DNN_TARGET_OPENCL_FP16
// if power layer do nothing so we replace it to Identity.
lp.type = "Split";
return std::shared_ptr<InferenceEngine::SplitLayer>(new InferenceEngine::SplitLayer(lp));
}
else
{
lp.type = "Power";
std::shared_ptr<InferenceEngine::PowerLayer> ieLayer(new InferenceEngine::PowerLayer(lp));
ieLayer->power = power;
ieLayer->scale = scale;
ieLayer->offset = shift;
return ieLayer;
}
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN
@ -1213,7 +1114,6 @@ struct ChannelsPReLUFunctor
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
InferenceEngine::Builder::Layer l = InferenceEngine::Builder::PReLULayer("");
@ -1221,16 +1121,6 @@ struct ChannelsPReLUFunctor
addConstantData("weights", wrapToInfEngineBlob(scale, {numChannels}, InferenceEngine::Layout::C), l);
return l;
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
{
lp.type = "PReLU";
std::shared_ptr<InferenceEngine::PReLULayer> ieLayer(new InferenceEngine::PReLULayer(lp));
const size_t numChannels = scale.total();
ieLayer->_weights = wrapToInfEngineBlob(scale, {numChannels}, InferenceEngine::Layout::C);
return ieLayer;
}
#endif
#endif // HAVE_INF_ENGINE
#ifdef HAVE_VULKAN

23
modules/dnn/src/layers/eltwise_layer.cpp
View File

@ -420,10 +420,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::EltwiseLayer ieLayer(name);
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(inputs.size()));
@ -442,26 +441,8 @@ public:
l.getParameters()["coeff"] = coeffs;
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Eltwise";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::EltwiseLayer> ieLayer(new InferenceEngine::EltwiseLayer(lp));
ieLayer->coeff = coeffs;
if (op == SUM)
ieLayer->_operation = InferenceEngine::EltwiseLayer::Sum;
else if (op == PROD)
ieLayer->_operation = InferenceEngine::EltwiseLayer::Prod;
else if (op == MAX)
ieLayer->_operation = InferenceEngine::EltwiseLayer::Max;
else
CV_Error(Error::StsNotImplemented, "Unsupported eltwise operation");
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

16
modules/dnn/src/layers/flatten_layer.cpp
View File

@ -162,10 +162,9 @@ public:
}
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
ieLayer.setType("Flatten");
@ -174,19 +173,8 @@ public:
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(1));
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Flatten";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["axis"] = format("%d", _startAxis);
ieLayer->params["end_axis"] = format("%d", _endAxis);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
int _startAxis;
int _endAxis;

22
modules/dnn/src/layers/fully_connected_layer.cpp
View File

@ -439,10 +439,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::FullyConnectedLayer ieLayer(name);
const int outNum = blobs[0].size[0];
@ -454,25 +453,8 @@ public:
addConstantData("biases", wrapToInfEngineBlob(blobs[1], {(size_t)outNum}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "FullyConnected";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::FullyConnectedLayer> ieLayer(new InferenceEngine::FullyConnectedLayer(lp));
ieLayer->_out_num = blobs[0].size[0];
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
ieLayer->params["out-size"] = format("%d", blobs[0].size[0]);
#endif
ieLayer->_weights = wrapToInfEngineBlob(blobs[0], {(size_t)blobs[0].size[0], (size_t)blobs[0].size[1], 1, 1}, InferenceEngine::Layout::OIHW);
if (blobs.size() > 1)
ieLayer->_biases = wrapToInfEngineBlob(blobs[1], {(size_t)ieLayer->_out_num}, InferenceEngine::Layout::C);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

21
modules/dnn/src/layers/lrn_layer.cpp
View File

@ -391,13 +391,13 @@ public:
#endif // HAVE_HALIDE
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
float alphaSize = alpha;
if (!normBySize)
alphaSize *= (type == SPATIAL_NRM ? size*size : size);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::NormLayer ieLayer(name);
ieLayer.setSize(size);
ieLayer.setAlpha(alphaSize);
@ -407,23 +407,8 @@ public:
InferenceEngine::Builder::Layer l = ieLayer;
l.getParameters()["k"] = bias;
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Norm";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::NormLayer> ieLayer(new InferenceEngine::NormLayer(lp));
ieLayer->_size = size;
ieLayer->_k = (int)bias;
ieLayer->_beta = beta;
ieLayer->_alpha = alphaSize;
ieLayer->_isAcrossMaps = (type == CHANNEL_NRM);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

29
modules/dnn/src/layers/mvn_layer.cpp
View File

@ -118,11 +118,7 @@ public:
{
#ifdef HAVE_INF_ENGINE
if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
return !zeroDev && eps <= 1e-7f;
#else
return !zeroDev && (preferableTarget == DNN_TARGET_CPU || eps <= 1e-7f);
#endif
return !zeroDev && (preferableTarget != DNN_TARGET_MYRIAD || eps <= 1e-7f);
else
#endif // HAVE_INF_ENGINE
return backendId == DNN_BACKEND_OPENCV;
@ -351,7 +347,11 @@ public:
bias = i < shift.cols ? ((float*)shift.data)[i] : bias;
}
cv::meanStdDev(inpRow, mean, (normVariance) ? dev : noArray());
double alpha = (normVariance) ? 1/(eps + dev[0]) : 1;
double alpha = 1;
if (normVariance)
{
alpha = 1 / std::sqrt(eps + dev[0]*dev[0]);
}
double normalizationScale = 1.0;
double normalizationShift = 0.0;
if (fuse_batch_norm)
@ -369,29 +369,16 @@ public:
}
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::MVNLayer ieLayer(name);
ieLayer.setAcrossChannels(acrossChannels);
ieLayer.setNormalize(normVariance);
ieLayer.setEpsilon(eps);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "MVN";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::MVNLayer> ieLayer(new InferenceEngine::MVNLayer(lp));
ieLayer->params["across_channels"] = acrossChannels ? "1" : "0";
ieLayer->params["normalize_variance"] = normVariance ? "1" : "0";
ieLayer->params["eps"] = format("%f", eps);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

49
modules/dnn/src/layers/normalize_bbox_layer.cpp
View File

@ -257,10 +257,9 @@ public:
}
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
if (input->dims.size() == 4)
{
@ -304,54 +303,8 @@ public:
return Ptr<BackendNode>(new InfEngineBackendNode(l));
}
#else
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::LayerParams lp;
lp.name = name;
lp.precision = InferenceEngine::Precision::FP32;
if (input->dims.size() == 4)
{
const int numChannels = input->dims[2]; // NOTE: input->dims are reversed (whcn)
lp.type = "Normalize";
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
if (blobs.empty())
{
auto weights = InferenceEngine::make_shared_blob<float>(InferenceEngine::Precision::FP32,
InferenceEngine::Layout::C,
{(size_t)numChannels});
weights->allocate();
std::vector<float> ones(numChannels, 1);
weights->set(ones);
ieLayer->blobs["weights"] = weights;
ieLayer->params["channel_shared"] = "0";
}
else
{
CV_Assert(numChannels == blobs[0].total());
ieLayer->blobs["weights"] = wrapToInfEngineBlob(blobs[0], {(size_t)numChannels}, InferenceEngine::Layout::C);
ieLayer->params["channel_shared"] = blobs[0].total() == 1 ? "1" : "0";
}
ieLayer->params["eps"] = format("%f", epsilon);
ieLayer->params["across_spatial"] = acrossSpatial ? "1" : "0";
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
}
else
{
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "GRN";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["bias"] = format("%f", epsilon);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
}
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
private:
int startAxis, endAxis;

5
modules/dnn/src/layers/padding_layer.cpp
View File

@ -182,9 +182,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
ieLayer.setType("Pad");
@ -204,9 +204,8 @@ public:
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(1));
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
return Ptr<BackendNode>();
}
#endif
private:
std::vector<std::pair<int, int> > paddings; // Pairs pad before, pad after.

20
modules/dnn/src/layers/permute_layer.cpp
View File

@ -382,30 +382,14 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::PermuteLayer ieLayer(name);
ieLayer.setOrder(_order);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Permute";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
CV_Assert(!_order.empty());
ieLayer->params["order"] = format("%zu", _order[0]);
for (int i = 1; i < _order.size(); ++i)
ieLayer->params["order"] += format(",%zu", _order[i]);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
size_t _count;
std::vector<size_t> _order;

64
modules/dnn/src/layers/pooling_layer.cpp
View File

@ -320,10 +320,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
if (type == MAX || type == AVE)
{
InferenceEngine::Builder::PoolingLayer ieLayer(name);
@ -366,69 +365,8 @@ public:
else
CV_Error(Error::StsNotImplemented, "Unsupported pooling type");
return Ptr<BackendNode>();
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer;
if (type == MAX || type == AVE)
{
lp.type = "Pooling";
InferenceEngine::PoolingLayer* poolLayer = new InferenceEngine::PoolingLayer(lp);
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
poolLayer->_kernel.insert(InferenceEngine::X_AXIS, kernel.width);
poolLayer->_kernel.insert(InferenceEngine::Y_AXIS, kernel.height);
poolLayer->_stride.insert(InferenceEngine::X_AXIS, stride.width);
poolLayer->_stride.insert(InferenceEngine::Y_AXIS, stride.height);
poolLayer->_padding.insert(InferenceEngine::X_AXIS, pad_l);
poolLayer->_padding.insert(InferenceEngine::Y_AXIS, pad_t);
poolLayer->_pads_end.insert(InferenceEngine::X_AXIS, pad_r);
poolLayer->_pads_end.insert(InferenceEngine::Y_AXIS, pad_b);
poolLayer->params["kernel"] = format("%d,%d", kernel.height, kernel.width);
poolLayer->params["pads_begin"] = format("%d,%d", pad_t, pad_l);
poolLayer->params["pads_end"] = format("%d,%d", pad_b, pad_r);
poolLayer->params["strides"] = format("%d,%d", stride.height, stride.width);
#else
poolLayer->_kernel_x = kernel.width;
poolLayer->_kernel_y = kernel.height;
poolLayer->_stride_x = stride.width;
poolLayer->_stride_y = stride.height;
poolLayer->_padding_x = pad_l;
poolLayer->_padding_y = pad_t;
poolLayer->params["pad-r"] = format("%d", pad_r);
poolLayer->params["pad-b"] = format("%d", pad_b);
#endif
poolLayer->_exclude_pad = type == AVE && padMode == "SAME";
poolLayer->params["rounding-type"] = ceilMode ? "ceil" : "floor";
poolLayer->_type = type == MAX ? InferenceEngine::PoolingLayer::PoolType::MAX :
InferenceEngine::PoolingLayer::PoolType::AVG;
ieLayer = std::shared_ptr<InferenceEngine::CNNLayer>(poolLayer);
}
else if (type == ROI)
{
lp.type = "ROIPooling";
ieLayer = std::shared_ptr<InferenceEngine::CNNLayer>(new InferenceEngine::CNNLayer(lp));
ieLayer->params["pooled_w"] = format("%d", pooledSize.width);
ieLayer->params["pooled_h"] = format("%d", pooledSize.height);
ieLayer->params["spatial_scale"] = format("%f", spatialScale);
}
else if (type == PSROI)
{
lp.type = "PSROIPooling";
ieLayer = std::shared_ptr<InferenceEngine::CNNLayer>(new InferenceEngine::CNNLayer(lp));
ieLayer->params["output_dim"] = format("%d", psRoiOutChannels);
ieLayer->params["group_size"] = format("%d", pooledSize.width);
ieLayer->params["spatial_scale"] = format("%f", spatialScale);
}
else
CV_Error(Error::StsNotImplemented, "Unsupported pooling type");
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
class PoolingInvoker : public ParallelLoopBody

61
modules/dnn/src/layers/prior_box_layer.cpp
View File

@ -495,10 +495,9 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
if (_explicitSizes)
{
InferenceEngine::Builder::PriorBoxClusteredLayer ieLayer(name);
@ -556,66 +555,8 @@ public:
l.getParameters()["variance"] = _variance;
return Ptr<BackendNode>(new InfEngineBackendNode(l));
}
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = _explicitSizes ? "PriorBoxClustered" : "PriorBox";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
if (_explicitSizes)
{
CV_Assert(!_boxWidths.empty()); CV_Assert(!_boxHeights.empty());
CV_Assert(_boxWidths.size() == _boxHeights.size());
ieLayer->params["width"] = format("%f", _boxWidths[0]);
ieLayer->params["height"] = format("%f", _boxHeights[0]);
for (int i = 1; i < _boxWidths.size(); ++i)
{
ieLayer->params["width"] += format(",%f", _boxWidths[i]);
ieLayer->params["height"] += format(",%f", _boxHeights[i]);
}
}
else
{
ieLayer->params["min_size"] = format("%f", _minSize);
ieLayer->params["max_size"] = _maxSize > 0 ? format("%f", _maxSize) : "";
if (!_aspectRatios.empty())
{
ieLayer->params["aspect_ratio"] = format("%f", _aspectRatios[0]);
for (int i = 1; i < _aspectRatios.size(); ++i)
ieLayer->params["aspect_ratio"] += format(",%f", _aspectRatios[i]);
}
}
ieLayer->params["flip"] = "0"; // We already flipped aspect ratios.
ieLayer->params["clip"] = _clip ? "1" : "0";
CV_Assert(!_variance.empty());
ieLayer->params["variance"] = format("%f", _variance[0]);
for (int i = 1; i < _variance.size(); ++i)
ieLayer->params["variance"] += format(",%f", _variance[i]);
if (_stepX == _stepY)
{
ieLayer->params["step"] = format("%f", _stepX);
ieLayer->params["step_h"] = "0.0";
ieLayer->params["step_w"] = "0.0";
}
else
{
ieLayer->params["step"] = "0.0";
ieLayer->params["step_h"] = format("%f", _stepY);
ieLayer->params["step_w"] = format("%f", _stepX);
}
CV_CheckEQ(_offsetsX.size(), (size_t)1, ""); CV_CheckEQ(_offsetsY.size(), (size_t)1, ""); CV_CheckEQ(_offsetsX[0], _offsetsY[0], "");
ieLayer->params["offset"] = format("%f", _offsetsX[0]);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

31
modules/dnn/src/layers/proposal_layer.cpp
View File

@ -322,10 +322,9 @@ public:
layerOutputs[0].col(2).copyTo(dst);
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ProposalLayer ieLayer(name);
ieLayer.setBaseSize(baseSize);
@ -346,36 +345,8 @@ public:
ieLayer.setRatio(ratiosVec);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Proposal";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["base_size"] = format("%d", baseSize);
ieLayer->params["feat_stride"] = format("%d", featStride);
ieLayer->params["min_size"] = "16";
ieLayer->params["nms_thresh"] = format("%f", nmsThreshold);
ieLayer->params["post_nms_topn"] = format("%d", keepTopAfterNMS);
ieLayer->params["pre_nms_topn"] = format("%d", keepTopBeforeNMS);
if (ratios.size())
{
ieLayer->params["ratio"] = format("%f", ratios.get<float>(0));
for (int i = 1; i < ratios.size(); ++i)
ieLayer->params["ratio"] += format(",%f", ratios.get<float>(i));
}
if (scales.size())
{
ieLayer->params["scale"] = format("%f", scales.get<float>(0));
for (int i = 1; i < scales.size(); ++i)
ieLayer->params["scale"] += format(",%f", scales.get<float>(i));
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
private:
// A first half of channels are background scores. We need only a second one.

15
modules/dnn/src/layers/reorg_layer.cpp
View File

@ -178,25 +178,14 @@ public:
permute->forward(inputs, outputs, internals_arr);
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ReorgYoloLayer ieLayer(name);
ieLayer.setStride(reorgStride);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "ReorgYolo";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["stride"] = format("%d", reorgStride);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

21
modules/dnn/src/layers/reshape_layer.cpp
View File

@ -257,34 +257,15 @@ public:
}
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ReshapeLayer ieLayer(name);
CV_Assert(outShapes.size() == 1);
ieLayer.setDims(outShapes[0]);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Reshape";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ReshapeLayer> ieLayer(new InferenceEngine::ReshapeLayer(lp));
if (!newShapeDesc.empty())
ieLayer->shape = newShapeDesc;
else
{
CV_Assert(inputs.size() == 2);
InferenceEngine::DataPtr shapeSrc = infEngineDataNode(inputs[1]);
// NOTE: shapeSrc->dims are reversed
ieLayer->shape = std::vector<int>(shapeSrc->dims.rbegin(), shapeSrc->dims.rend());
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
private:
std::vector<MatShape> outShapes;

47
modules/dnn/src/layers/resize_layer.cpp
View File

@ -55,7 +55,7 @@ public:
if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
{
return (interpolation == "nearest" && scaleWidth == scaleHeight) ||
(interpolation == "bilinear" && INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R4));
(interpolation == "bilinear");
}
#endif
return backendId == DNN_BACKEND_OPENCV;
@ -162,7 +162,6 @@ public:
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
if (interpolation == "nearest")
@ -188,32 +187,6 @@ public:
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(1));
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer;
if (interpolation == "nearest")
{
lp.type = "Resample";
ieLayer = std::shared_ptr<InferenceEngine::CNNLayer>(new InferenceEngine::CNNLayer(lp));
ieLayer->params["type"] = "caffe.ResampleParameter.NEAREST";
ieLayer->params["antialias"] = "0";
}
else if (interpolation == "bilinear")
{
lp.type = "Interp";
ieLayer = std::shared_ptr<InferenceEngine::CNNLayer>(new InferenceEngine::CNNLayer(lp));
ieLayer->params["pad_beg"] = "0";
ieLayer->params["pad_end"] = "0";
ieLayer->params["align_corners"] = "0";
}
else
CV_Error(Error::StsNotImplemented, "Unsupported interpolation: " + interpolation);
ieLayer->params["width"] = cv::format("%d", outWidth);
ieLayer->params["height"] = cv::format("%d", outHeight);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
@ -271,10 +244,9 @@ public:
scaleWidth = (outWidth > 1) ? (static_cast<float>(inpWidth - 1) / (outWidth - 1)) : 0.f;
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
ieLayer.setType("Interp");
@ -285,20 +257,9 @@ public:
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(1));
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Interp";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
ieLayer->params["pad_beg"] = "0";
ieLayer->params["pad_end"] = "0";
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
};
Ptr<Layer> InterpLayer::create(const LayerParams& params)

32
modules/dnn/src/layers/scale_layer.cpp
View File

@ -194,10 +194,9 @@ public:
}
#endif // HAVE_HALIDE
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer l = InferenceEngine::Builder::ScaleShiftLayer(name);
CV_Assert(!blobs.empty());
@ -219,37 +218,8 @@ public:
if (hasBias)
addConstantData("biases", wrapToInfEngineBlob(blobs.back(), {numChannels}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "ScaleShift";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::ScaleShiftLayer> ieLayer(new InferenceEngine::ScaleShiftLayer(lp));
CV_Assert(!blobs.empty());
const size_t numChannels = blobs[0].total();
if (hasWeights)
{
ieLayer->_weights = wrapToInfEngineBlob(blobs[0], {numChannels}, InferenceEngine::Layout::C);
}
else
{
auto weights = InferenceEngine::make_shared_blob<float>(InferenceEngine::Precision::FP32,
{numChannels});
weights->allocate();
std::vector<float> ones(numChannels, 1);
weights->set(ones);
ieLayer->_weights = weights;
}
if (hasBias)
ieLayer->_biases = wrapToInfEngineBlob(blobs.back(), {numChannels}, InferenceEngine::Layout::C);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
void getScaleShift(Mat& scale, Mat& shift) const CV_OVERRIDE
{

124
modules/dnn/src/layers/slice_layer.cpp
View File

@ -54,7 +54,7 @@ namespace cv
namespace dnn
{
class SliceLayerImpl CV_FINAL : public SliceLayer
class SliceLayerImpl : public SliceLayer
{
public:
SliceLayerImpl(const LayerParams& params)
@ -112,6 +112,9 @@ public:
{
return backendId == DNN_BACKEND_OPENCV ||
(backendId == DNN_BACKEND_INFERENCE_ENGINE &&
#ifdef HAVE_INF_ENGINE
INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R1) &&
#endif
sliceRanges.size() == 1 && sliceRanges[0].size() == 4);
}
@ -256,23 +259,24 @@ public:
}
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R1)
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
CV_Assert(sliceRanges.size() == 1);
CV_Assert_N(sliceRanges.size() == 1, inputs.size() <= 2);
std::vector<size_t> axes, offsets, dims;
int from, to, step;
int numDims = sliceRanges[0].size();
if (preferableTarget == DNN_TARGET_MYRIAD)
{
from = 1;
from = axis;
to = numDims;
step = 1;
}
else
{
from = numDims - 1;
to = -1;
to = axis - 1;
step = -1;
}
for (int i = from; i != to; i += step)
@ -282,11 +286,6 @@ public:
dims.push_back(sliceRanges[0][i].size());
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
std::vector<size_t> outShape(numDims);
for (int i = 0; i < numDims; ++i)
outShape[numDims - 1 - i] = sliceRanges[0][i].size();
InferenceEngine::Builder::Layer ieLayer(name);
ieLayer.setName(name);
ieLayer.setType("Crop");
@ -295,6 +294,13 @@ public:
ieLayer.getParameters()["offset"] = offsets;
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(2));
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
if (inputs.size() != 2)
{
std::vector<size_t> outShape(numDims);
for (int i = 0; i < numDims; ++i)
outShape[numDims - 1 - i] = sliceRanges[0][i].size();
ieLayer.getInputPorts()[1].setParameter("type", "weights");
// Fake blob which will be moved to inputs (as weights).
@ -303,22 +309,91 @@ public:
InferenceEngine::Layout::ANY, outShape);
shapeSource->allocate();
addConstantData("weights", shapeSource, ieLayer);
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "Crop";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CropLayer> ieLayer(new InferenceEngine::CropLayer(lp));
ieLayer->axis = axes;
ieLayer->offset = offsets;
ieLayer->dim = dims;
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif // IE < R5
return Ptr<BackendNode>();
}
#endif
#endif
};
class CropLayerImpl CV_FINAL : public SliceLayerImpl
{
public:
CropLayerImpl(const LayerParams& params) : SliceLayerImpl(LayerParams())
{
setParamsFrom(params);
axis = params.get<int>("axis", 2);
const DictValue *paramOffset = params.ptr("offset");
if (paramOffset)
{
for (int i = 0; i < paramOffset->size(); i++)
offset.push_back(paramOffset->get<int>(i));
}
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
const int requiredOutputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const CV_OVERRIDE
{
CV_Assert(inputs.size() == 2);
MatShape dstShape = inputs[0];
int start = clamp(axis, dstShape);
for (int i = start; i < dstShape.size(); i++)
{
dstShape[i] = inputs[1][i];
}
outputs.resize(1, dstShape);
return false;
}
void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays) CV_OVERRIDE
{
std::vector<Mat> inputs;
inputs_arr.getMatVector(inputs);
CV_Assert(2 == inputs.size());
const Mat &inpBlob = inputs[0];
const Mat &inpSzBlob = inputs[1];
int dims = inpBlob.dims;
int start_axis = clamp(axis, dims);
std::vector<int> offset_final(dims, 0);
if (offset.size() == 1)
{
for (int i = start_axis; i < dims; i++)
offset_final[i] = offset[0];
}
else if (offset.size() > 1)
{
if ((int)offset.size() != dims - start_axis)
CV_Error(Error::StsBadArg, "number of offset values specified must be "
"equal to the number of dimensions following axis.");
for (int i = start_axis; i < dims; i++)
offset_final[i] = offset[i - start_axis];
}
sliceRanges.resize(1);
sliceRanges[0].resize(dims);
for (int i = 0; i < start_axis; i++)
{
sliceRanges[0][i] = Range(0, inpBlob.size[i]);
}
for (int i = start_axis; i < dims; i++)
{
if (offset_final[i] < 0 || offset_final[i] + inpSzBlob.size[i] > inpBlob.size[i])
CV_Error(Error::StsBadArg, "invalid crop parameters or blob sizes");
sliceRanges[0][i] = Range(offset_final[i], offset_final[i] + inpSzBlob.size[i]);
}
}
private:
std::vector<int> offset;
};
Ptr<SliceLayer> SliceLayer::create(const LayerParams& params)
@ -326,5 +401,10 @@ Ptr<SliceLayer> SliceLayer::create(const LayerParams& params)
return Ptr<SliceLayer>(new SliceLayerImpl(params));
}
Ptr<Layer> CropLayer::create(const LayerParams& params)
{
return Ptr<Layer>(new CropLayerImpl(params));
}
}
}

17
modules/dnn/src/layers/softmax_layer.cpp
View File

@ -323,29 +323,16 @@ public:
return Ptr<BackendNode>();
}
#ifdef HAVE_INF_ENGINE
virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::Builder::SoftMaxLayer ieLayer(name);
ieLayer.setAxis(clamp(axisRaw, input->dims.size()));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
InferenceEngine::LayerParams lp;
lp.name = name;
lp.type = "SoftMax";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::SoftMaxLayer> ieLayer(new InferenceEngine::SoftMaxLayer(lp));
ieLayer->axis = clamp(axisRaw, input->dims.size());
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#endif
#endif // HAVE_INF_ENGINE
return Ptr<BackendNode>();
}
#endif // HAVE_INF_ENGINE
int64 getFLOPS(const std::vector<MatShape> &inputs,
const std::vector<MatShape> &outputs) const CV_OVERRIDE

390
modules/dnn/src/op_inf_engine.cpp
View File

@ -26,33 +26,8 @@ namespace cv { namespace dnn {
// we can use some predefined name.
static std::string kDefaultInpLayerName = "empty_inp_layer_name";
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InfEngineBackendNode::InfEngineBackendNode(const InferenceEngine::Builder::Layer& _layer)
: BackendNode(DNN_BACKEND_INFERENCE_ENGINE), layer(_layer) {}
#else
InfEngineBackendNode::InfEngineBackendNode(const InferenceEngine::CNNLayerPtr& _layer)
: BackendNode(DNN_BACKEND_INFERENCE_ENGINE), layer(_layer) {}
void InfEngineBackendNode::connect(std::vector<Ptr<BackendWrapper> >& inputs,
std::vector<Ptr<BackendWrapper> >& outputs)
{
layer->insData.resize(inputs.size());
for (int i = 0; i < inputs.size(); ++i)
{
InferenceEngine::DataPtr dataPtr = infEngineDataNode(inputs[i]);
layer->insData[i] = InferenceEngine::DataWeakPtr(dataPtr);
dataPtr->inputTo[layer->name] = layer;
}
CV_Assert(!outputs.empty());
layer->outData.resize(1);
InferenceEngine::DataPtr dataPtr = infEngineDataNode(outputs[0]);
dataPtr->name = layer->name;
layer->outData[0] = dataPtr;
dataPtr->creatorLayer = InferenceEngine::CNNLayerWeakPtr(layer);
}
#endif
static std::vector<Ptr<InfEngineBackendWrapper> >
infEngineWrappers(const std::vector<Ptr<BackendWrapper> >& ptrs)
@ -67,8 +42,6 @@ infEngineWrappers(const std::vector<Ptr<BackendWrapper> >& ptrs)
return wrappers;
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InfEngineBackendNet::InfEngineBackendNet() : netBuilder("")
{
hasNetOwner = false;
@ -238,8 +211,6 @@ void InfEngineBackendNet::addOutput(const std::string& name)
requestedOutputs.push_back(name);
}
#endif // IE >= R5
static InferenceEngine::Layout estimateLayout(const Mat& m)
{
if (m.dims == 4)
@ -352,349 +323,6 @@ void InfEngineBackendWrapper::setHostDirty()
}
#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2018R5)
InfEngineBackendNet::InfEngineBackendNet()
{
targetDevice = InferenceEngine::TargetDevice::eCPU;
precision = InferenceEngine::Precision::FP32;
hasNetOwner = false;
}
InfEngineBackendNet::InfEngineBackendNet(InferenceEngine::CNNNetwork& net)
{
targetDevice = InferenceEngine::TargetDevice::eCPU;
precision = InferenceEngine::Precision::FP32;
inputs = net.getInputsInfo();
outputs = net.getOutputsInfo();
layers.resize(net.layerCount()); // A hack to execute InfEngineBackendNet::layerCount correctly.
netOwner = net;
hasNetOwner = true;
}
void InfEngineBackendNet::Release() CV_NOEXCEPT
{
layers.clear();
inputs.clear();
outputs.clear();
}
void InfEngineBackendNet::setPrecision(InferenceEngine::Precision p) CV_NOEXCEPT
{
precision = p;
}
InferenceEngine::Precision InfEngineBackendNet::getPrecision() CV_NOEXCEPT
{
return hasNetOwner ? netOwner.getPrecision() : precision;
}
InferenceEngine::Precision InfEngineBackendNet::getPrecision() const CV_NOEXCEPT
{
return hasNetOwner ? netOwner.getPrecision() : precision;
}
// Assume that outputs of network is unconnected blobs.
void InfEngineBackendNet::getOutputsInfo(InferenceEngine::OutputsDataMap &outputs_) CV_NOEXCEPT
{
const_cast<const InfEngineBackendNet*>(this)->getOutputsInfo(outputs_);
}
void InfEngineBackendNet::getOutputsInfo(InferenceEngine::OutputsDataMap &outputs_) const CV_NOEXCEPT
{
outputs_ = outputs;
}
// Returns input references that aren't connected to internal outputs.
void InfEngineBackendNet::getInputsInfo(InferenceEngine::InputsDataMap &inputs_) CV_NOEXCEPT
{
const_cast<const InfEngineBackendNet*>(this)->getInputsInfo(inputs_);
}
// Returns input references that aren't connected to internal outputs.
void InfEngineBackendNet::getInputsInfo(InferenceEngine::InputsDataMap &inputs_) const CV_NOEXCEPT
{
inputs_ = inputs;
}
InferenceEngine::InputInfo::Ptr InfEngineBackendNet::getInput(const std::string &inputName) CV_NOEXCEPT
{
return const_cast<const InfEngineBackendNet*>(this)->getInput(inputName);
}
InferenceEngine::InputInfo::Ptr InfEngineBackendNet::getInput(const std::string &inputName) const CV_NOEXCEPT
{
const auto& it = inputs.find(inputName);
CV_Assert(it != inputs.end());
return it->second;
}
void InfEngineBackendNet::getName(char*, size_t) CV_NOEXCEPT
{
}
void InfEngineBackendNet::getName(char*, size_t) const CV_NOEXCEPT
{
}
const std::string& InfEngineBackendNet::getName() const CV_NOEXCEPT
{
return name;
}
InferenceEngine::StatusCode InfEngineBackendNet::serialize(const std::string&, const std::string&, InferenceEngine::ResponseDesc*) const CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
}
size_t InfEngineBackendNet::layerCount() CV_NOEXCEPT
{
return const_cast<const InfEngineBackendNet*>(this)->layerCount();
}
size_t InfEngineBackendNet::layerCount() const CV_NOEXCEPT
{
return layers.size();
}
InferenceEngine::DataPtr& InfEngineBackendNet::getData(const char *dname) CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return outputs.begin()->second; // Just return something.
}
void InfEngineBackendNet::addLayer(const InferenceEngine::CNNLayerPtr &layer) CV_NOEXCEPT
{
layers.push_back(layer);
inputs.clear();
outputs.clear();
}
InferenceEngine::StatusCode
InfEngineBackendNet::addOutput(const std::string &layerName, size_t outputIndex,
InferenceEngine::ResponseDesc *resp) CV_NOEXCEPT
{
for (const auto& l : layers)
{
for (const InferenceEngine::DataPtr& out : l->outData)
{
if (out->name == layerName)
{
outputs[out->name] = out;
return InferenceEngine::StatusCode::OK;
}
}
}
CV_Error(Error::StsObjectNotFound, "Cannot find a layer " + layerName);
return InferenceEngine::StatusCode::OK;
}
InferenceEngine::StatusCode
InfEngineBackendNet::getLayerByName(const char *layerName, InferenceEngine::CNNLayerPtr &out,
InferenceEngine::ResponseDesc *resp) CV_NOEXCEPT
{
return const_cast<const InfEngineBackendNet*>(this)->getLayerByName(layerName, out, resp);
}
InferenceEngine::StatusCode InfEngineBackendNet::getLayerByName(const char *layerName,
InferenceEngine::CNNLayerPtr &out,
InferenceEngine::ResponseDesc *resp) const CV_NOEXCEPT
{
for (auto& l : layers)
{
if (l->name == layerName)
{
out = l;
return InferenceEngine::StatusCode::OK;
}
}
CV_Error(Error::StsObjectNotFound, cv::format("Cannot find a layer %s", layerName));
return InferenceEngine::StatusCode::NOT_FOUND;
}
void InfEngineBackendNet::setTargetDevice(InferenceEngine::TargetDevice device) CV_NOEXCEPT
{
if (device != InferenceEngine::TargetDevice::eCPU &&
device != InferenceEngine::TargetDevice::eGPU &&
device != InferenceEngine::TargetDevice::eMYRIAD &&
device != InferenceEngine::TargetDevice::eFPGA)
CV_Error(Error::StsNotImplemented, "");
targetDevice = device;
}
InferenceEngine::TargetDevice InfEngineBackendNet::getTargetDevice() CV_NOEXCEPT
{
return const_cast<const InfEngineBackendNet*>(this)->getTargetDevice();
}
InferenceEngine::TargetDevice InfEngineBackendNet::getTargetDevice() const CV_NOEXCEPT
{
return targetDevice == InferenceEngine::TargetDevice::eFPGA ?
InferenceEngine::TargetDevice::eHETERO : targetDevice;
}
InferenceEngine::StatusCode InfEngineBackendNet::setBatchSize(const size_t) CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
}
InferenceEngine::StatusCode InfEngineBackendNet::setBatchSize(size_t size, InferenceEngine::ResponseDesc *responseDesc) CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
}
size_t InfEngineBackendNet::getBatchSize() const CV_NOEXCEPT
{
size_t batchSize = 0;
for (const auto& inp : inputs)
{
CV_Assert(inp.second);
std::vector<size_t> dims = inp.second->getDims();
CV_Assert(!dims.empty());
if (batchSize != 0)
CV_Assert(batchSize == dims.back());
else
batchSize = dims.back();
}
return batchSize;
}
InferenceEngine::StatusCode InfEngineBackendNet::AddExtension(const InferenceEngine::IShapeInferExtensionPtr &extension, InferenceEngine::ResponseDesc *resp) CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
}
InferenceEngine::StatusCode InfEngineBackendNet::reshape(const InferenceEngine::ICNNNetwork::InputShapes &inputShapes, InferenceEngine::ResponseDesc *resp) CV_NOEXCEPT
{
CV_Error(Error::StsNotImplemented, "");
return InferenceEngine::StatusCode::OK;
}
void InfEngineBackendNet::init(int targetId)
{
if (inputs.empty())
{
// Collect all external input blobs.
inputs.clear();
std::map<std::string, InferenceEngine::DataPtr> internalOutputs;
for (const auto& l : layers)
{
for (const InferenceEngine::DataWeakPtr& ptr : l->insData)
{
InferenceEngine::DataPtr inp(ptr);
if (internalOutputs.find(inp->name) == internalOutputs.end())
{
InferenceEngine::InputInfo::Ptr inpInfo(new InferenceEngine::InputInfo());
inpInfo->setInputData(inp);
if (inputs.find(inp->name) == inputs.end())
inputs[inp->name] = inpInfo;
}
}
for (const InferenceEngine::DataPtr& out : l->outData)
{
// TODO: Replace to uniqueness assertion.
if (internalOutputs.find(out->name) == internalOutputs.end())
internalOutputs[out->name] = out;
}
}
CV_Assert(!inputs.empty());
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R3)
for (const auto& inp : inputs)
{
InferenceEngine::LayerParams lp;
lp.name = inp.first;
lp.type = "Input";
lp.precision = InferenceEngine::Precision::FP32;
std::shared_ptr<InferenceEngine::CNNLayer> inpLayer(new InferenceEngine::CNNLayer(lp));
layers.push_back(inpLayer);
InferenceEngine::DataPtr dataPtr = inp.second->getInputData();
// TODO: remove precision dependency (see setInput.normalization tests)
if (dataPtr->precision == InferenceEngine::Precision::FP32)
{
inpLayer->outData.assign(1, dataPtr);
dataPtr->creatorLayer = InferenceEngine::CNNLayerWeakPtr(inpLayer);
}
}
#endif
}
if (outputs.empty())
{
// Add all unconnected blobs to output blobs.
InferenceEngine::OutputsDataMap unconnectedOuts;
for (const auto& l : layers)
{
if (l->type == "Input")
continue;
// Add all outputs.
for (const InferenceEngine::DataPtr& out : l->outData)
{
// TODO: Replace to uniqueness assertion.
if (unconnectedOuts.find(out->name) == unconnectedOuts.end())
unconnectedOuts[out->name] = out;
}
// Remove internally connected outputs.
for (const InferenceEngine::DataWeakPtr& inp : l->insData)
{
unconnectedOuts.erase(InferenceEngine::DataPtr(inp)->name);
}
}
CV_Assert(!unconnectedOuts.empty());
for (auto it = unconnectedOuts.begin(); it != unconnectedOuts.end(); ++it)
{
outputs[it->first] = it->second;
}
}
// Set up input blobs.
inpBlobs.clear();
for (const auto& it : inputs)
{
CV_Assert(allBlobs.find(it.first) != allBlobs.end());
inpBlobs[it.first] = allBlobs[it.first];
it.second->setPrecision(inpBlobs[it.first]->precision());
}
// Set up output blobs.
outBlobs.clear();
for (const auto& it : outputs)
{
CV_Assert(allBlobs.find(it.first) != allBlobs.end());
outBlobs[it.first] = allBlobs[it.first];
}
switch (targetId)
{
case DNN_TARGET_CPU: setTargetDevice(InferenceEngine::TargetDevice::eCPU); break;
case DNN_TARGET_OPENCL_FP16:
setPrecision(InferenceEngine::Precision::FP16);
/* Falls through. */
case DNN_TARGET_OPENCL: setTargetDevice(InferenceEngine::TargetDevice::eGPU); break;
case DNN_TARGET_MYRIAD:
{
setPrecision(InferenceEngine::Precision::FP16);
setTargetDevice(InferenceEngine::TargetDevice::eMYRIAD); break;
}
case DNN_TARGET_FPGA:
{
setPrecision(InferenceEngine::Precision::FP16);
setTargetDevice(InferenceEngine::TargetDevice::eFPGA); break;
}
default:
CV_Error(Error::StsError, format("Unknown target identifier: %d", targetId));
}
if (!isInitialized())
initPlugin(*this);
}
#endif // IE < R5
static std::map<InferenceEngine::TargetDevice, InferenceEngine::InferenceEnginePluginPtr>& getSharedPlugins()
{
@ -703,7 +331,7 @@ static std::map<InferenceEngine::TargetDevice, InferenceEngine::InferenceEngineP
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5) && !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
#if !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
static bool detectMyriadX_()
{
InferenceEngine::Builder::Network builder("");
@ -724,6 +352,7 @@ static bool detectMyriadX_()
}
builder.addLayer({InferenceEngine::PortInfo(clampId)}, InferenceEngine::Builder::OutputLayer());
#else
InferenceEngine::idx_t clampId = builder.addLayer({inpId}, InferenceEngine::Builder::ClampLayer());
builder.addLayer({InferenceEngine::PortInfo(clampId)},
InferenceEngine::Builder::OutputLayer().setPort(InferenceEngine::Port({},
@ -757,7 +386,7 @@ static bool detectMyriadX_()
}
return true;
}
#endif // >= 2018R5
#endif // !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
void InfEngineBackendNet::initPlugin(InferenceEngine::ICNNNetwork& net)
{
@ -834,9 +463,7 @@ void InfEngineBackendNet::addBlobs(const std::vector<Ptr<BackendWrapper> >& ptrs
for (const auto& wrapper : wrappers)
{
std::string name = wrapper->dataPtr->name;
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
name = name.empty() ? kDefaultInpLayerName : name;
#endif
allBlobs.insert({name, wrapper->blob});
}
}
@ -993,11 +620,7 @@ bool InfEngineBackendLayer::getMemoryShapes(const std::vector<MatShape> &inputs,
std::vector<MatShape> &outputs,
std::vector<MatShape> &internals) const
{
#if INF_ENGINE_VER_MAJOR_EQ(INF_ENGINE_RELEASE_2018R3)
InferenceEngine::ICNNNetwork::InputShapes inShapes = const_cast<InferenceEngine::CNNNetwork&>(t_net).getInputShapes();
#else
InferenceEngine::ICNNNetwork::InputShapes inShapes = t_net.getInputShapes();
#endif
InferenceEngine::ICNNNetwork::InputShapes::iterator itr;
bool equal_flag = true;
size_t i = 0;
@ -1044,7 +667,6 @@ InferenceEngine::Blob::Ptr convertFp16(const InferenceEngine::Blob::Ptr& blob)
return halfs;
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
void addConstantData(const std::string& name, InferenceEngine::Blob::Ptr data,
InferenceEngine::Builder::Layer& l)
{
@ -1054,7 +676,6 @@ void addConstantData(const std::string& name, InferenceEngine::Blob::Ptr data,
l.addConstantData(name, data);
#endif
}
#endif
#endif // HAVE_INF_ENGINE
@ -1103,7 +724,7 @@ static std::string getInferenceEngineVPUType_()
{
#if defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
param_vpu_type = OPENCV_DNN_IE_VPU_TYPE_DEFAULT;
#elif INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
#else
CV_LOG_INFO(NULL, "OpenCV-DNN: running Inference Engine VPU autodetection: Myriad2/X. In case of other accelerator types specify 'OPENCV_DNN_IE_VPU_TYPE' parameter");
try {
bool isMyriadX_ = detectMyriadX_();
@ -1121,9 +742,6 @@ static std::string getInferenceEngineVPUType_()
CV_LOG_WARNING(NULL, "OpenCV-DNN: Failed Inference Engine VPU autodetection. Specify 'OPENCV_DNN_IE_VPU_TYPE' parameter.");
param_vpu_type.clear();
}
#else
CV_LOG_WARNING(NULL, "OpenCV-DNN: VPU auto-detection is not implemented. Consider specifying VPU type via 'OPENCV_DNN_IE_VPU_TYPE' parameter");
param_vpu_type = CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_2;
#endif
}
CV_LOG_INFO(NULL, "OpenCV-DNN: Inference Engine VPU type='" << param_vpu_type << "'");

121
modules/dnn/src/op_inf_engine.hpp
View File

@ -19,8 +19,6 @@
#ifdef HAVE_INF_ENGINE
#define INF_ENGINE_RELEASE_2018R3 2018030000
#define INF_ENGINE_RELEASE_2018R4 2018040000
#define INF_ENGINE_RELEASE_2018R5 2018050000
#define INF_ENGINE_RELEASE_2019R1 2019010000
@ -46,9 +44,7 @@
#include <inference_engine.hpp>
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
#include <ie_builders.hpp>
#endif
#if defined(__GNUC__) && INF_ENGINE_VER_MAJOR_LE(INF_ENGINE_RELEASE_2019R1)
#pragma GCC visibility pop
@ -64,111 +60,6 @@ namespace cv { namespace dnn {
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_LT(INF_ENGINE_RELEASE_2018R5)
class InfEngineBackendNet : public InferenceEngine::ICNNNetwork
{
public:
InfEngineBackendNet();
InfEngineBackendNet(InferenceEngine::CNNNetwork& net);
virtual void Release() CV_NOEXCEPT CV_OVERRIDE;
void setPrecision(InferenceEngine::Precision p) CV_NOEXCEPT;
virtual InferenceEngine::Precision getPrecision() CV_NOEXCEPT;
virtual InferenceEngine::Precision getPrecision() const CV_NOEXCEPT;
virtual void getOutputsInfo(InferenceEngine::OutputsDataMap &out) CV_NOEXCEPT /*CV_OVERRIDE*/;
virtual void getOutputsInfo(InferenceEngine::OutputsDataMap &out) const CV_NOEXCEPT /*CV_OVERRIDE*/;
virtual void getInputsInfo(InferenceEngine::InputsDataMap &inputs) CV_NOEXCEPT /*CV_OVERRIDE*/;
virtual void getInputsInfo(InferenceEngine::InputsDataMap &inputs) const CV_NOEXCEPT /*CV_OVERRIDE*/;
virtual InferenceEngine::InputInfo::Ptr getInput(const std::string &inputName) CV_NOEXCEPT;
virtual InferenceEngine::InputInfo::Ptr getInput(const std::string &inputName) const CV_NOEXCEPT;
virtual InferenceEngine::StatusCode serialize(const std::string &xmlPath, const std::string &binPath, InferenceEngine::ResponseDesc* resp) const CV_NOEXCEPT;
virtual void getName(char *pName, size_t len) CV_NOEXCEPT;
virtual void getName(char *pName, size_t len) const CV_NOEXCEPT;
virtual const std::string& getName() const CV_NOEXCEPT;
virtual size_t layerCount() CV_NOEXCEPT;
virtual size_t layerCount() const CV_NOEXCEPT;
virtual InferenceEngine::DataPtr& getData(const char *dname) CV_NOEXCEPT CV_OVERRIDE;
virtual void addLayer(const InferenceEngine::CNNLayerPtr &layer) CV_NOEXCEPT CV_OVERRIDE;
virtual InferenceEngine::StatusCode addOutput(const std::string &layerName,
size_t outputIndex = 0,
InferenceEngine::ResponseDesc *resp = nullptr) CV_NOEXCEPT;
virtual InferenceEngine::StatusCode getLayerByName(const char *layerName,
InferenceEngine::CNNLayerPtr &out,
InferenceEngine::ResponseDesc *resp) CV_NOEXCEPT;
virtual InferenceEngine::StatusCode getLayerByName(const char *layerName,
InferenceEngine::CNNLayerPtr &out,
InferenceEngine::ResponseDesc *resp) const CV_NOEXCEPT;
virtual void setTargetDevice(InferenceEngine::TargetDevice device) CV_NOEXCEPT CV_OVERRIDE;
virtual InferenceEngine::TargetDevice getTargetDevice() CV_NOEXCEPT;
virtual InferenceEngine::TargetDevice getTargetDevice() const CV_NOEXCEPT;
virtual InferenceEngine::StatusCode setBatchSize(const size_t size) CV_NOEXCEPT CV_OVERRIDE;
virtual InferenceEngine::StatusCode setBatchSize(size_t size, InferenceEngine::ResponseDesc* responseDesc) CV_NOEXCEPT;
virtual size_t getBatchSize() const CV_NOEXCEPT CV_OVERRIDE;
virtual InferenceEngine::StatusCode AddExtension(const InferenceEngine::IShapeInferExtensionPtr& extension, InferenceEngine::ResponseDesc* resp) CV_NOEXCEPT CV_OVERRIDE;
virtual InferenceEngine::StatusCode reshape(const InputShapes& inputShapes, InferenceEngine::ResponseDesc* resp) CV_NOEXCEPT CV_OVERRIDE;
void init(int targetId);
void addBlobs(const std::vector<Ptr<BackendWrapper> >& wrappers);
void forward();
bool isInitialized();
private:
std::vector<InferenceEngine::CNNLayerPtr> layers;
InferenceEngine::InputsDataMap inputs;
InferenceEngine::OutputsDataMap outputs;
InferenceEngine::BlobMap inpBlobs;
InferenceEngine::BlobMap outBlobs;
InferenceEngine::BlobMap allBlobs;
InferenceEngine::TargetDevice targetDevice;
InferenceEngine::Precision precision;
InferenceEngine::InferenceEnginePluginPtr enginePtr;
InferenceEngine::InferencePlugin plugin;
InferenceEngine::ExecutableNetwork netExec;
InferenceEngine::InferRequest infRequest;
// In case of models from Model Optimizer we need to manage their lifetime.
InferenceEngine::CNNNetwork netOwner;
// There is no way to check if netOwner is initialized or not so we use
// a separate flag to determine if the model has been loaded from IR.
bool hasNetOwner;
std::string name;
void initPlugin(InferenceEngine::ICNNNetwork& net);
};
#else // IE < R5
class InfEngineBackendNet
{
public:
@ -226,28 +117,18 @@ private:
std::set<int> unconnectedLayersIds;
};
#endif // IE < R5
class InfEngineBackendNode : public BackendNode
{
public:
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InfEngineBackendNode(const InferenceEngine::Builder::Layer& layer);
#else
InfEngineBackendNode(const InferenceEngine::CNNLayerPtr& layer);
#endif
void connect(std::vector<Ptr<BackendWrapper> >& inputs,
std::vector<Ptr<BackendWrapper> >& outputs);
// Inference Engine network object that allows to obtain the outputs of this layer.
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer layer;
Ptr<InfEngineBackendNet> net;
#else
InferenceEngine::CNNLayerPtr layer;
Ptr<InfEngineBackendNet> net;
#endif
};
class InfEngineBackendWrapper : public BackendWrapper
@ -282,9 +163,7 @@ Mat infEngineBlobToMat(const InferenceEngine::Blob::Ptr& blob);
// Allocates memory for a new blob.
InferenceEngine::Blob::Ptr convertFp16(const InferenceEngine::Blob::Ptr& blob);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
void addConstantData(const std::string& name, InferenceEngine::Blob::Ptr data, InferenceEngine::Builder::Layer& l);
#endif
// This is a fake class to run networks from Model Optimizer. Objects of that
// class simulate responses of layers are imported by OpenCV and supported by

6
modules/dnn/src/opencl/mvn.cl
View File

@ -118,10 +118,10 @@ __kernel void MVN(__global const Dtype* src,
return;
Dtype mean_val = mean[x];
Dtype dev_val = sqrt(dev[x]);
Dtype dev_val = dev[x];
Dtype alpha;
#ifdef NORM_VARIANCE
alpha = 1 / (eps + dev_val);
alpha = 1 / sqrt(eps + dev_val);
#else
alpha = 1;
#endif
@ -275,7 +275,7 @@ __kernel void MVN_FUSE(__global const Dtype * tmp,
barrier(CLK_LOCAL_MEM_FENCE);
Dtype4 mean_val = convert_float4(mean[row_gid]);
Dtype4 dev_val = sqrt(work[0] * alpha_val) + (Dtype4)eps;
Dtype4 dev_val = sqrt(work[0] * alpha_val + (Dtype4)eps);
Dtype4 alpha = (Dtype4)1.f / dev_val;
Dtype4 w = (Dtype4)1.f;

87
modules/dnn/src/tensorflow/tf_graph_simplifier.cpp
View File

@ -70,13 +70,6 @@ public:
{
fusedNodeInputs = inputs_;
fusedNodeOp = op;
nodesToFuse.clear();
for (int i = 0; i < nodes.size(); ++i)
{
if (std::find(fusedNodeInputs.begin(), fusedNodeInputs.end(), i) == fusedNodeInputs.end() &&
nodes[i] != "Const")
nodesToFuse.push_back(i);
}
}
static int getInputNodeId(const tensorflow::GraphDef& net,
@ -99,15 +92,17 @@ public:
// Match TensorFlow subgraph starting from <nodeId> with a set of nodes to be fused.
// Const nodes are skipped during matching. Returns true if nodes are matched and can be fused.
virtual bool match(const tensorflow::GraphDef& net, int nodeId, std::vector<int>& matchedNodesIds)
virtual bool match(const tensorflow::GraphDef& net, int nodeId,
std::vector<int>& matchedNodesIds,
std::vector<int>& targetNodesIds)
{
matchedNodesIds.clear();
matchedNodesIds.reserve(nodesToFuse.size());
targetNodesIds.clear();
std::queue<int> nodesToMatch;
std::queue<int> targetNodes;
nodesToMatch.push(nodeId);
targetNodes.push(nodesToFuse.back());
targetNodes.push(nodes.size() - 1);
while (!nodesToMatch.empty())
{
int nodeToMatch = nodesToMatch.front();
@ -142,13 +137,25 @@ public:
return false;
}
matchedNodesIds.push_back(nodeToMatch);
targetNodesIds.push_back(targetNodeId);
}
const int n = matchedNodesIds.size();
std::vector<std::pair<int, int> > elements(n);
for (int i = 0; i < n; ++i)
elements[i] = std::make_pair(matchedNodesIds[i], targetNodesIds[i]);
std::sort(elements.begin(), elements.end());
for (int i = 0; i < n; ++i)
{
matchedNodesIds[i] = elements[i].first;
targetNodesIds[i] = elements[i].second;
}
std::sort(matchedNodesIds.begin(), matchedNodesIds.end());
return true;
}
// Fuse matched subgraph.
void replace(tensorflow::GraphDef& net, const std::vector<int>& matchedNodesIds)
void replace(tensorflow::GraphDef& net, const std::vector<int>& matchedNodesIds,
const std::vector<int>& targetNodesIds)
{
// Extract names of input nodes.
std::vector<std::string> inputsNames(fusedNodeInputs.size());
@ -159,7 +166,7 @@ public:
for (int j = 0; j < matchedNodesIds.size() && inpName.empty(); ++j)
{
const tensorflow::NodeDef &node = net.node(matchedNodesIds[j]);
std::vector<int>& inpIndices = inputs[nodesToFuse[j]];
std::vector<int>& inpIndices = inputs[targetNodesIds[j]];
CV_Assert(node.input_size() == inpIndices.size());
for (int k = 0; k < inpIndices.size(); ++k)
@ -204,7 +211,6 @@ private:
std::vector<std::vector<int> > inputs; // Connections of an every node to it's inputs.
std::string fusedNodeOp; // Operation name of resulting fused node.
std::vector<int> nodesToFuse; // Set of nodes to be fused.
std::vector<int> fusedNodeInputs; // Inputs of fused node.
};
@ -360,9 +366,11 @@ public:
setFusedNode("Relu6", input);
}
virtual bool match(const tensorflow::GraphDef& net, int nodeId, std::vector<int>& matchedNodesIds) CV_OVERRIDE
virtual bool match(const tensorflow::GraphDef& net, int nodeId,
std::vector<int>& matchedNodesIds,
std::vector<int>& targetNodesIds) CV_OVERRIDE
{
if (!Subgraph::match(net, nodeId, matchedNodesIds))
if (!Subgraph::match(net, nodeId, matchedNodesIds, targetNodesIds))
return false;
Mat maxValue = getTensorContent(net.node(matchedNodesIds.front() + 1).attr().at("value").tensor());
return maxValue.type() == CV_32FC1 && maxValue.total() == 1 && maxValue.at<float>(0) == 6;
@ -394,14 +402,16 @@ public:
setFusedNode("Reshape", ids);
}
virtual bool match(const tensorflow::GraphDef& net, int nodeId, std::vector<int>& matchedNodesIds) CV_OVERRIDE
virtual bool match(const tensorflow::GraphDef& net, int nodeId,
std::vector<int>& matchedNodesIds,
std::vector<int>& targetNodesIds) CV_OVERRIDE
{
const tensorflow::NodeDef& node = net.node(nodeId);
if (node.input_size() == 0)
return false;
inpName = node.input(0);
return Subgraph::match(net, nodeId, matchedNodesIds);
return Subgraph::match(net, nodeId, matchedNodesIds, targetNodesIds);
}
@ -693,6 +703,40 @@ public:
}
};
class KerasMVNSubgraph : public Subgraph
{
public:
KerasMVNSubgraph()
{
int input = addNodeToMatch("");
int mean = addNodeToMatch("Mean", input, addNodeToMatch("Const"));
int grad = addNodeToMatch("StopGradient", mean);
int diff = addNodeToMatch("SquaredDifference", input, grad);
int var = addNodeToMatch("Mean", diff, addNodeToMatch("Const"));
int sub = addNodeToMatch("Sub", input, mean);
int add_y = addNodeToMatch("Const");
int add = addNodeToMatch("Add", var, add_y);
int pow_y = addNodeToMatch("Const");
int powNode = addNodeToMatch("Pow", add, pow_y);
addNodeToMatch("RealDiv", sub, powNode);
setFusedNode("MVN", input, add_y);
}
virtual void finalize(tensorflow::GraphDef&, tensorflow::NodeDef* fusedNode,
std::vector<tensorflow::NodeDef*>& inputNodes) CV_OVERRIDE
{
tensorflow::AttrValue eps;
Mat epsMat = getTensorContent(inputNodes[1]->attr().at("value").tensor());
CV_CheckEQ(epsMat.total(), (size_t)1, "");
CV_CheckTypeEQ(epsMat.type(), CV_32FC1, "");
eps.set_f(epsMat.at<float>(0));
fusedNode->mutable_attr()->insert(MapPair<std::string, tensorflow::AttrValue>("eps", eps));
fusedNode->mutable_input()->RemoveLast();
}
};
void simplifySubgraphs(tensorflow::GraphDef& net)
{
std::vector<Ptr<Subgraph> > subgraphs;
@ -712,16 +756,17 @@ void simplifySubgraphs(tensorflow::GraphDef& net)
subgraphs.push_back(Ptr<Subgraph>(new SoftMaxSlimSubgraph()));
subgraphs.push_back(Ptr<Subgraph>(new SoftMaxSlimV2Subgraph()));
subgraphs.push_back(Ptr<Subgraph>(new ReshapeAsShapeSubgraph()));
subgraphs.push_back(Ptr<Subgraph>(new KerasMVNSubgraph()));
int numNodes = net.node_size();
std::vector<int> matchedNodesIds;
std::vector<int> matchedNodesIds, targetNodesIds;
for (int i = 0; i < numNodes; ++i)
{
for (int j = 0; j < subgraphs.size(); ++j)
{
if (subgraphs[j]->match(net, i, matchedNodesIds))
if (subgraphs[j]->match(net, i, matchedNodesIds, targetNodesIds))
{
subgraphs[j]->replace(net, matchedNodesIds);
subgraphs[j]->replace(net, matchedNodesIds, targetNodesIds);
numNodes -= matchedNodesIds.size() - 1; // #matchedNodes removed and one added.
break;
}

7
modules/dnn/test/test_backends.cpp
View File

@ -291,6 +291,7 @@ TEST_P(DNNTestNetwork, OpenPose_pose_mpi)
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
throw SkipTestException("Test is disabled for OpenVINO <= 2018R5 + MyriadX target");
#endif
// output range: [-0.001, 0.97]
const float l1 = (target == DNN_TARGET_MYRIAD) ? 0.012 : 0.0;
const float lInf = (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 0.16 : 0.0;
@ -309,6 +310,7 @@ TEST_P(DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages)
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
throw SkipTestException("Test is disabled for OpenVINO <= 2018R5 + MyriadX target");
#endif
// The same .caffemodel but modified .prototxt
// See https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/pose/poseParameters.cpp
processNet("dnn/openpose_pose_mpi.caffemodel", "dnn/openpose_pose_mpi_faster_4_stages.prototxt",
@ -322,9 +324,6 @@ TEST_P(DNNTestNetwork, OpenFace)
#if INF_ENGINE_VER_MAJOR_EQ(2018050000)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
throw SkipTestException("Test is disabled for Myriad targets");
#elif INF_ENGINE_VER_MAJOR_EQ(2018030000)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("Test has been fixed in OpenVINO 2018R4");
#endif
#endif
if (backend == DNN_BACKEND_HALIDE)
@ -407,7 +406,9 @@ TEST_P(DNNTestNetwork, FastNeuralStyle_eccv16)
float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.4 : 4e-5;
float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 7.45 : 2e-3;
processNet("dnn/fast_neural_style_eccv16_starry_night.t7", "", inp, "", "", l1, lInf);
#if defined(HAVE_INF_ENGINE) && INF_ENGINE_RELEASE >= 2019010000
expectNoFallbacksFromIE(net);
#endif
}
INSTANTIATE_TEST_CASE_P(/*nothing*/, DNNTestNetwork, dnnBackendsAndTargets(true, true, false, true));

8
modules/dnn/test/test_darknet_importer.cpp
View File

@ -332,10 +332,6 @@ TEST_P(Test_Darknet_nets, TinyYoloVoc)
testDarknetModel(config_file, weights_file, ref.rowRange(0, 2), scoreDiff, iouDiff);
}
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2018040000)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
throw SkipTestException("Test with 'batch size 2' is disabled for Myriad target (fixed in 2018R5)");
#endif
{
SCOPED_TRACE("batch size 2");
testDarknetModel(config_file, weights_file, ref, scoreDiff, iouDiff);
@ -389,10 +385,6 @@ INSTANTIATE_TEST_CASE_P(/**/, Test_Darknet_nets, dnnBackendsAndTargets());
TEST_P(Test_Darknet_layers, shortcut)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE < 2018040000
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_CPU)
throw SkipTestException("Test is enabled starts from OpenVINO 2018R4");
#endif
testDarknetLayer("shortcut");
}

14
modules/dnn/test/test_halide_layers.cpp
View File

@ -159,12 +159,6 @@ TEST_P(Deconvolution, Accuracy)
Backend backendId = get<0>(get<7>(GetParam()));
Target targetId = get<1>(get<7>(GetParam()));
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2018040000)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_CPU
&& hasBias && group != 1)
throw SkipTestException("Test is disabled for OpenVINO 2018R4");
#endif
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_GE(2019010000)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
@ -278,19 +272,13 @@ TEST_P(AvePooling, Accuracy)
Backend backendId = get<0>(get<4>(GetParam()));
Target targetId = get<1>(get<4>(GetParam()));
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_GE(2018050000)
#if defined(INF_ENGINE_RELEASE)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
&& kernel == Size(1, 1) && (stride == Size(1, 1) || stride == Size(2, 2)))
throw SkipTestException("Test is disabled for MyriadX target");
#endif
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2018040000)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD &&
stride == Size(3, 2) && kernel == Size(3, 3) && outSize != Size(1, 1))
throw SkipTestException("Test is fixed in OpenVINO 2018R4");
#endif
const int inWidth = (outSize.width - 1) * stride.width + kernel.width;
const int inHeight = (outSize.height - 1) * stride.height + kernel.height;

12
modules/dnn/test/test_layers.cpp
View File

@ -141,10 +141,8 @@ TEST_P(Test_Caffe_layers, Convolution)
TEST_P(Test_Caffe_layers, DeConvolution)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2018040000
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_CPU)
throw SkipTestException("Test is disabled for OpenVINO 2018R4");
#endif
throw SkipTestException("Test is disabled for DLIE/CPU");
testLayerUsingCaffeModels("layer_deconvolution", true, false);
}
@ -254,8 +252,7 @@ TEST_P(Test_Caffe_layers, Fused_Concat)
#if defined(INF_ENGINE_RELEASE)
if (backend == DNN_BACKEND_INFERENCE_ENGINE
&& (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16
|| (INF_ENGINE_RELEASE < 2018040000 && target == DNN_TARGET_CPU))
&& (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16)
)
throw SkipTestException("Test is disabled for DLIE");
#endif
@ -1045,11 +1042,6 @@ TEST_P(Test_DLDT_two_inputs_3dim, as_IR)
int secondInpType = get<1>(GetParam());
Target targetId = get<2>(GetParam());
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE < 2018040000
if (secondInpType == CV_8U)
throw SkipTestException("Test is enabled starts from OpenVINO 2018R4");
#endif
std::string suffix = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? "_fp16" : "";
Net net = readNet(_tf("net_two_inputs" + suffix + ".xml"), _tf("net_two_inputs.bin"));
std::vector<int> inpSize = get<3>(GetParam());

22
modules/dnn/test/test_onnx_importer.cpp
View File

@ -86,6 +86,9 @@ TEST_P(Test_ONNX_layers, Convolution)
TEST_P(Test_ONNX_layers, Convolution3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
testONNXModels("conv3d");
@ -94,7 +97,7 @@ TEST_P(Test_ONNX_layers, Convolution3D)
TEST_P(Test_ONNX_layers, Two_convolution)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_GE(2018050000)
#if defined(INF_ENGINE_RELEASE)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
)
@ -150,6 +153,9 @@ TEST_P(Test_ONNX_layers, AveragePooling)
TEST_P(Test_ONNX_layers, MaxPooling3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
testONNXModels("max_pool3d");
@ -157,6 +163,9 @@ TEST_P(Test_ONNX_layers, MaxPooling3D)
TEST_P(Test_ONNX_layers, AvePooling3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
testONNXModels("ave_pool3d");
@ -202,7 +211,11 @@ TEST_P(Test_ONNX_layers, Constant)
TEST_P(Test_ONNX_layers, Padding)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
testONNXModels("padding", npy, 0, 0, false, false);
#else
testONNXModels("padding");
#endif
}
TEST_P(Test_ONNX_layers, Resize)
@ -247,7 +260,11 @@ TEST_P(Test_ONNX_layers, Reshape)
TEST_P(Test_ONNX_layers, Slice)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
testONNXModels("slice", npy, 0, 0, false, false);
#else
testONNXModels("slice");
#endif
}
TEST_P(Test_ONNX_layers, Softmax)
@ -504,6 +521,9 @@ TEST_P(Test_ONNX_nets, Shufflenet)
TEST_P(Test_ONNX_nets, Resnet34_kinetics)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");

24
modules/dnn/test/test_tf_importer.cpp
View File

@ -133,6 +133,9 @@ TEST_P(Test_TensorFlow_layers, conv)
TEST_P(Test_TensorFlow_layers, Convolution3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
runTensorFlowNet("conv3d");
@ -187,6 +190,7 @@ TEST_P(Test_TensorFlow_layers, batch_norm)
runTensorFlowNet("mvn_batch_norm");
runTensorFlowNet("mvn_batch_norm_1x1");
runTensorFlowNet("switch_identity");
runTensorFlowNet("keras_batch_norm_training");
}
TEST_P(Test_TensorFlow_layers, batch_norm3D)
@ -229,6 +233,9 @@ TEST_P(Test_TensorFlow_layers, ave_pool_same)
TEST_P(Test_TensorFlow_layers, MaxPooling3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
runTensorFlowNet("max_pool3d");
@ -236,6 +243,9 @@ TEST_P(Test_TensorFlow_layers, MaxPooling3D)
TEST_P(Test_TensorFlow_layers, AvePooling3D)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LT(2019010000)
throw SkipTestException("Test is enabled starts from 2019R1");
#endif
if (backend != DNN_BACKEND_INFERENCE_ENGINE || target != DNN_TARGET_CPU)
throw SkipTestException("Only DLIE backend on CPU is supported");
runTensorFlowNet("ave_pool3d");
@ -250,6 +260,7 @@ TEST_P(Test_TensorFlow_layers, deconvolution)
runTensorFlowNet("deconvolution_adj_pad_same");
runTensorFlowNet("keras_deconv_valid");
runTensorFlowNet("keras_deconv_same");
runTensorFlowNet("keras_deconv_same_v2");
}
TEST_P(Test_TensorFlow_layers, matmul)
@ -337,10 +348,15 @@ class Test_TensorFlow_nets : public DNNTestLayer {};
TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
{
#if defined(INF_ENGINE_RELEASE)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD
&& getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
)
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
{
#if INF_ENGINE_VER_MAJOR_GE(2019010000)
if (getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
throw SkipTestException("Test is disabled for MyriadX");
#else
throw SkipTestException("Test is disabled for Myriad");
#endif
}
#endif
checkBackend();
@ -364,7 +380,9 @@ TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.0043 : default_l1;
double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.037 : default_lInf;
normAssertDetections(ref, out, "", 0.2, scoreDiff, iouDiff);
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2019010000
expectNoFallbacksFromIE(net);
#endif
}
TEST_P(Test_TensorFlow_nets, Inception_v2_SSD)

8
modules/dnn/test/test_torch_importer.cpp
View File

@ -136,10 +136,8 @@ TEST_P(Test_Torch_layers, run_reshape_change_batch_size)
TEST_P(Test_Torch_layers, run_reshape)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2018040000
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
throw SkipTestException("Test is disabled for OpenVINO 2018R4");
#endif
throw SkipTestException("Test is disabled for Myriad targets");
runTorchNet("net_reshape_batch");
runTorchNet("net_reshape_channels", "", false, true);
}
@ -220,10 +218,6 @@ TEST_P(Test_Torch_layers, net_conv_gemm_lrn)
TEST_P(Test_Torch_layers, net_inception_block)
{
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE == 2018030000
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
throw SkipTestException("");
#endif
runTorchNet("net_inception_block", "", false, true);
}

11
modules/java/generator/src/cpp/common.h
View File

@ -16,6 +16,17 @@ extern "C" {
#endif
#include <jni.h>
// make -fvisibility=hidden work with java 1.7
#if defined(__linux__) && !defined(__ANDROID__) && !defined (JNI_VERSION_1_8)
// adapted from jdk1.8/jni.h
#if (defined(__GNUC__) && ((__GNUC__ > 4) || (__GNUC__ == 4) && (__GNUC_MINOR__ > 2))) || __has_attribute(visibility)
#undef JNIEXPORT
#define JNIEXPORT __attribute__((visibility("default")))
#undef JNIIMPORT
#define JNIIMPORT __attribute__((visibility("default")))
#endif
#endif
} // extern "C"
#include "opencv_java.hpp"

8
platforms/scripts/valgrind.supp
View File

@ -227,6 +227,14 @@
fun:cvWriteFrame_FFMPEG
}
{
OpenCV-OpenEXR-ThreadPool
Memcheck:Leak
fun:_Znwm
fun:_ZN16IlmThread_opencv10ThreadPoolC1Ej
fun:_ZN16IlmThread_opencv10ThreadPool16globalThreadPoolEv
}
{
OpenCV-test-gapi-thread-tls
Memcheck:Leak