 new int8 implement,better accuracy (#749)
* add the armv7a conv3x3s1 implement without overflow,remove old codes
* fix the bug of conv3x3s2 packed int8
* new int8 implement,weight quant by perchanel,better accuracy~
* fix the bug of conv3x3s1 packed int8 neon
* add the naive c fp32 and int8 winograd F(2,3)
* add the neon intrinsic int8 winograd F(2,3)
* optimize the armv7a int8 winograd F(2,3) with neon assembly
* optimize the armv7a int8 winograd F(2,3) input transform with assembly.
* add the requantize layer and int8 relu implement.
* add graph optimize conv1x1s2 -> conv1x1s1,begin optimize int8 aarch64.
* fix int8 bugs
* add the c naive im2col with sgemm
* add aarch64 int8 winograd f23, conv3x3s2 naive implement
* add the int8 sgemm conv7x7s2 on x86/armv7a platform
* optimize the int8 sgemm by neon intrinsic and packed kernel
* optimize the int8 sgemm with packed data
* optimize the int8 sgemm with armv7a neon assembly
* add the int8 sgemm on arm64-v8a platform
* perpare to merge latest codes from master
* add the int8 param files
* In the Class Net,add the fuse_network method
7 years ago |
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108 |
- 7767517
- 106 122
- Input data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3
- Convolution conv1 1 1 data conv1_conv1_relu -23330=4,3,112,112,64 0=64 1=7 3=2 4=3 5=1 6=9408 9=1
- Pooling pool1 1 1 conv1_conv1_relu pool1 -23330=4,3,56,56,64 1=3 2=2
- Split splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1 -23330=8,3,56,56,64,3,56,56,64
- Convolution res2a_branch1 1 1 pool1_splitncnn_1 res2a_branch1_scale2a_branch1 -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384
- Convolution res2a_branch2a 1 1 pool1_splitncnn_0 res2a_branch2a_res2a_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=4096 9=1
- Convolution res2a_branch2b 1 1 res2a_branch2a_res2a_branch2a_relu res2a_branch2b_res2a_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1
- Convolution res2a_branch2c 1 1 res2a_branch2b_res2a_branch2b_relu res2a_branch2c_scale2a_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384
- Eltwise res2a 2 1 res2a_branch1_scale2a_branch1 res2a_branch2c_scale2a_branch2c res2a -23330=4,3,56,56,256 0=1
- ReLU res2a_relu 1 1 res2a res2a_res2a_relu -23330=4,3,56,56,256
- Split splitncnn_1 1 2 res2a_res2a_relu res2a_res2a_relu_splitncnn_0 res2a_res2a_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256
- Convolution res2b_branch2a 1 1 res2a_res2a_relu_splitncnn_1 res2b_branch2a_res2b_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=16384 9=1
- Convolution res2b_branch2b 1 1 res2b_branch2a_res2b_branch2a_relu res2b_branch2b_res2b_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1
- Convolution res2b_branch2c 1 1 res2b_branch2b_res2b_branch2b_relu res2b_branch2c_scale2b_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384
- Eltwise res2b 2 1 res2a_res2a_relu_splitncnn_0 res2b_branch2c_scale2b_branch2c res2b -23330=4,3,56,56,256 0=1
- ReLU res2b_relu 1 1 res2b res2b_res2b_relu -23330=4,3,56,56,256
- Split splitncnn_2 1 2 res2b_res2b_relu res2b_res2b_relu_splitncnn_0 res2b_res2b_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256
- Convolution res2c_branch2a 1 1 res2b_res2b_relu_splitncnn_1 res2c_branch2a_res2c_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=16384 9=1
- Convolution res2c_branch2b 1 1 res2c_branch2a_res2c_branch2a_relu res2c_branch2b_res2c_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1
- Convolution res2c_branch2c 1 1 res2c_branch2b_res2c_branch2b_relu res2c_branch2c_scale2c_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384
- Eltwise res2c 2 1 res2b_res2b_relu_splitncnn_0 res2c_branch2c_scale2c_branch2c res2c -23330=4,3,56,56,256 0=1
- ReLU res2c_relu 1 1 res2c res2c_res2c_relu -23330=4,3,56,56,256
- Split splitncnn_3 1 2 res2c_res2c_relu res2c_res2c_relu_splitncnn_0 res2c_res2c_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256
- Convolution res3a_branch1 1 1 res2c_res2c_relu_splitncnn_1 res3a_branch1_scale3a_branch1 -23330=4,3,28,28,512 0=512 1=1 3=2 5=1 6=131072
- Convolution res3a_branch2a 1 1 res2c_res2c_relu_splitncnn_0 res3a_branch2a_res3a_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 3=2 5=1 6=32768 9=1
- Convolution res3a_branch2b 1 1 res3a_branch2a_res3a_branch2a_relu res3a_branch2b_res3a_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1
- Convolution res3a_branch2c 1 1 res3a_branch2b_res3a_branch2b_relu res3a_branch2c_scale3a_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536
- Eltwise res3a 2 1 res3a_branch1_scale3a_branch1 res3a_branch2c_scale3a_branch2c res3a -23330=4,3,28,28,512 0=1
- ReLU res3a_relu 1 1 res3a res3a_res3a_relu -23330=4,3,28,28,512
- Split splitncnn_4 1 2 res3a_res3a_relu res3a_res3a_relu_splitncnn_0 res3a_res3a_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512
- Convolution res3b_branch2a 1 1 res3a_res3a_relu_splitncnn_1 res3b_branch2a_res3b_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1
- Convolution res3b_branch2b 1 1 res3b_branch2a_res3b_branch2a_relu res3b_branch2b_res3b_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1
- Convolution res3b_branch2c 1 1 res3b_branch2b_res3b_branch2b_relu res3b_branch2c_scale3b_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536
- Eltwise res3b 2 1 res3a_res3a_relu_splitncnn_0 res3b_branch2c_scale3b_branch2c res3b -23330=4,3,28,28,512 0=1
- ReLU res3b_relu 1 1 res3b res3b_res3b_relu -23330=4,3,28,28,512
- Split splitncnn_5 1 2 res3b_res3b_relu res3b_res3b_relu_splitncnn_0 res3b_res3b_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512
- Convolution res3c_branch2a 1 1 res3b_res3b_relu_splitncnn_1 res3c_branch2a_res3c_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1
- Convolution res3c_branch2b 1 1 res3c_branch2a_res3c_branch2a_relu res3c_branch2b_res3c_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1
- Convolution res3c_branch2c 1 1 res3c_branch2b_res3c_branch2b_relu res3c_branch2c_scale3c_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536
- Eltwise res3c 2 1 res3b_res3b_relu_splitncnn_0 res3c_branch2c_scale3c_branch2c res3c -23330=4,3,28,28,512 0=1
- ReLU res3c_relu 1 1 res3c res3c_res3c_relu -23330=4,3,28,28,512
- Split splitncnn_6 1 2 res3c_res3c_relu res3c_res3c_relu_splitncnn_0 res3c_res3c_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512
- Convolution res3d_branch2a 1 1 res3c_res3c_relu_splitncnn_1 res3d_branch2a_res3d_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1
- Convolution res3d_branch2b 1 1 res3d_branch2a_res3d_branch2a_relu res3d_branch2b_res3d_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1
- Convolution res3d_branch2c 1 1 res3d_branch2b_res3d_branch2b_relu res3d_branch2c_scale3d_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536
- Eltwise res3d 2 1 res3c_res3c_relu_splitncnn_0 res3d_branch2c_scale3d_branch2c res3d -23330=4,3,28,28,512 0=1
- ReLU res3d_relu 1 1 res3d res3d_res3d_relu -23330=4,3,28,28,512
- Split splitncnn_7 1 2 res3d_res3d_relu res3d_res3d_relu_splitncnn_0 res3d_res3d_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512
- Convolution res4a_branch1 1 1 res3d_res3d_relu_splitncnn_1 res4a_branch1_scale4a_branch1 -23330=4,3,14,14,1024 0=1024 1=1 3=2 5=1 6=524288
- Convolution res4a_branch2a 1 1 res3d_res3d_relu_splitncnn_0 res4a_branch2a_res4a_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 3=2 5=1 6=131072 9=1
- Convolution res4a_branch2b 1 1 res4a_branch2a_res4a_branch2a_relu res4a_branch2b_res4a_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4a_branch2c 1 1 res4a_branch2b_res4a_branch2b_relu res4a_branch2c_scale4a_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4a 2 1 res4a_branch1_scale4a_branch1 res4a_branch2c_scale4a_branch2c res4a -23330=4,3,14,14,1024 0=1
- ReLU res4a_relu 1 1 res4a res4a_res4a_relu -23330=4,3,14,14,1024
- Split splitncnn_8 1 2 res4a_res4a_relu res4a_res4a_relu_splitncnn_0 res4a_res4a_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res4b_branch2a 1 1 res4a_res4a_relu_splitncnn_1 res4b_branch2a_res4b_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1
- Convolution res4b_branch2b 1 1 res4b_branch2a_res4b_branch2a_relu res4b_branch2b_res4b_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4b_branch2c 1 1 res4b_branch2b_res4b_branch2b_relu res4b_branch2c_scale4b_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4b 2 1 res4a_res4a_relu_splitncnn_0 res4b_branch2c_scale4b_branch2c res4b -23330=4,3,14,14,1024 0=1
- ReLU res4b_relu 1 1 res4b res4b_res4b_relu -23330=4,3,14,14,1024
- Split splitncnn_9 1 2 res4b_res4b_relu res4b_res4b_relu_splitncnn_0 res4b_res4b_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res4c_branch2a 1 1 res4b_res4b_relu_splitncnn_1 res4c_branch2a_res4c_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1
- Convolution res4c_branch2b 1 1 res4c_branch2a_res4c_branch2a_relu res4c_branch2b_res4c_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4c_branch2c 1 1 res4c_branch2b_res4c_branch2b_relu res4c_branch2c_scale4c_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4c 2 1 res4b_res4b_relu_splitncnn_0 res4c_branch2c_scale4c_branch2c res4c -23330=4,3,14,14,1024 0=1
- ReLU res4c_relu 1 1 res4c res4c_res4c_relu -23330=4,3,14,14,1024
- Split splitncnn_10 1 2 res4c_res4c_relu res4c_res4c_relu_splitncnn_0 res4c_res4c_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res4d_branch2a 1 1 res4c_res4c_relu_splitncnn_1 res4d_branch2a_res4d_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1
- Convolution res4d_branch2b 1 1 res4d_branch2a_res4d_branch2a_relu res4d_branch2b_res4d_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4d_branch2c 1 1 res4d_branch2b_res4d_branch2b_relu res4d_branch2c_scale4d_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4d 2 1 res4c_res4c_relu_splitncnn_0 res4d_branch2c_scale4d_branch2c res4d -23330=4,3,14,14,1024 0=1
- ReLU res4d_relu 1 1 res4d res4d_res4d_relu -23330=4,3,14,14,1024
- Split splitncnn_11 1 2 res4d_res4d_relu res4d_res4d_relu_splitncnn_0 res4d_res4d_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res4e_branch2a 1 1 res4d_res4d_relu_splitncnn_1 res4e_branch2a_res4e_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1
- Convolution res4e_branch2b 1 1 res4e_branch2a_res4e_branch2a_relu res4e_branch2b_res4e_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4e_branch2c 1 1 res4e_branch2b_res4e_branch2b_relu res4e_branch2c_scale4e_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4e 2 1 res4d_res4d_relu_splitncnn_0 res4e_branch2c_scale4e_branch2c res4e -23330=4,3,14,14,1024 0=1
- ReLU res4e_relu 1 1 res4e res4e_res4e_relu -23330=4,3,14,14,1024
- Split splitncnn_12 1 2 res4e_res4e_relu res4e_res4e_relu_splitncnn_0 res4e_res4e_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res4f_branch2a 1 1 res4e_res4e_relu_splitncnn_1 res4f_branch2a_res4f_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1
- Convolution res4f_branch2b 1 1 res4f_branch2a_res4f_branch2a_relu res4f_branch2b_res4f_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1
- Convolution res4f_branch2c 1 1 res4f_branch2b_res4f_branch2b_relu res4f_branch2c_scale4f_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144
- Eltwise res4f 2 1 res4e_res4e_relu_splitncnn_0 res4f_branch2c_scale4f_branch2c res4f -23330=4,3,14,14,1024 0=1
- ReLU res4f_relu 1 1 res4f res4f_res4f_relu -23330=4,3,14,14,1024
- Split splitncnn_13 1 2 res4f_res4f_relu res4f_res4f_relu_splitncnn_0 res4f_res4f_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024
- Convolution res5a_branch1 1 1 res4f_res4f_relu_splitncnn_1 res5a_branch1_scale5a_branch1 -23330=4,3,7,7,2048 0=2048 1=1 3=2 5=1 6=2097152
- Convolution res5a_branch2a 1 1 res4f_res4f_relu_splitncnn_0 res5a_branch2a_res5a_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 3=2 5=1 6=524288 9=1
- Convolution res5a_branch2b 1 1 res5a_branch2a_res5a_branch2a_relu res5a_branch2b_res5a_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1
- Convolution res5a_branch2c 1 1 res5a_branch2b_res5a_branch2b_relu res5a_branch2c_scale5a_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576
- Eltwise res5a 2 1 res5a_branch1_scale5a_branch1 res5a_branch2c_scale5a_branch2c res5a -23330=4,3,7,7,2048 0=1
- ReLU res5a_relu 1 1 res5a res5a_res5a_relu -23330=4,3,7,7,2048
- Split splitncnn_14 1 2 res5a_res5a_relu res5a_res5a_relu_splitncnn_0 res5a_res5a_relu_splitncnn_1 -23330=8,3,7,7,2048,3,7,7,2048
- Convolution res5b_branch2a 1 1 res5a_res5a_relu_splitncnn_1 res5b_branch2a_res5b_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 5=1 6=1048576 9=1
- Convolution res5b_branch2b 1 1 res5b_branch2a_res5b_branch2a_relu res5b_branch2b_res5b_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1
- Convolution res5b_branch2c 1 1 res5b_branch2b_res5b_branch2b_relu res5b_branch2c_scale5b_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576
- Eltwise res5b 2 1 res5a_res5a_relu_splitncnn_0 res5b_branch2c_scale5b_branch2c res5b -23330=4,3,7,7,2048 0=1
- ReLU res5b_relu 1 1 res5b res5b_res5b_relu -23330=4,3,7,7,2048
- Split splitncnn_15 1 2 res5b_res5b_relu res5b_res5b_relu_splitncnn_0 res5b_res5b_relu_splitncnn_1 -23330=8,3,7,7,2048,3,7,7,2048
- Convolution res5c_branch2a 1 1 res5b_res5b_relu_splitncnn_1 res5c_branch2a_res5c_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 5=1 6=1048576 9=1
- Convolution res5c_branch2b 1 1 res5c_branch2a_res5c_branch2a_relu res5c_branch2b_res5c_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1
- Convolution res5c_branch2c 1 1 res5c_branch2b_res5c_branch2b_relu res5c_branch2c_scale5c_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576
- Eltwise res5c 2 1 res5b_res5b_relu_splitncnn_0 res5c_branch2c_scale5c_branch2c res5c -23330=4,3,7,7,2048 0=1
- ReLU res5c_relu 1 1 res5c res5c_res5c_relu -23330=4,3,7,7,2048
- Pooling pool5 1 1 res5c_res5c_relu pool5 -23330=4,3,1,1,2048 0=1 1=7
- InnerProduct fc1000 1 1 pool5 fc1000 -23330=4,1,1000,1,1 0=1000 1=1 2=2048000
- Softmax prob 1 1 fc1000 output -23330=4,1,1000,1,1
|
