Lnet (#3)

* new structure for LNet * update mini-caffe * update mini-caffe * update mini-caffe * update
luoyetx · Oct 20, 2017 · 73d4f61 · 73d4f61
1 parent f7faf2a
commit 73d4f61
Show file tree

Hide file tree

Showing 12 changed files with 28,370 additions and 392 deletions.
diff --git a/1130.log b/1130.log
diff --git a/cpp/mini-caffe b/cpp/mini-caffe
diff --git a/cpp/profile/README.md b/cpp/profile/README.md
@@ -46,3 +46,7 @@ detect img/test3.jpg costs 0.3730s
 image size = (778 x 1024), s1: 0.1960s, s2: 0.1270s, s3: 0.0400s, s4: 0.0050
 bboxes, s1: 401, s2: 21, s3: 5, s4: 5
 ```
+
+### Update
+
+LNet 利用卷积替换掉全连接，在保证精度不变的情况下，提高了速度。GTX 1070 上从 17.59ms 降到了 13.907ms。
diff --git a/img/lnet.png b/img/lnet.png
diff --git a/img/test1_stage4.jpg b/img/test1_stage4.jpg
diff --git a/img/test2_stage4.jpg b/img/test2_stage4.jpg
diff --git a/img/test3_stage4.jpg b/img/test3_stage4.jpg
diff --git a/model/l.caffemodel b/model/l.caffemodel
diff --git a/proto/l.prototxt b/proto/l.prototxt
@@ -91,157 +91,62 @@ layer {
   top: "conv3"
 }
 layer {
-  name: "fc4"
-  type: "InnerProduct"
+  name: "conv4"
+  type: "Convolution"
   bottom: "conv3"
-  top: "fc4"
-  inner_product_param {
-    num_output: 256
+  top: "conv4"
+  convolution_param {
+    num_output: 100
+    group: 5
+    kernel_size: 3
+    stride: 3
   }
 }
 layer {
   name: "prelu4"
   type: "PReLU"
-  bottom: "fc4"
-  top: "fc4"
-}
-######## net-1 output ########
-layer {
-  name: "fc5_1"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc5_1"
-  inner_product_param {
-    num_output: 64
-  }
-}
-layer {
-  name: "prelu5_1"
-  type: "PReLU"
-  bottom: "fc5_1"
-  top: "fc5_1"
-}
-layer {
-  name: "fc6_1"
-  type: "InnerProduct"
-  bottom: "fc5_1"
-  top: "fc6_1"
-  inner_product_param {
-    num_output: 2
-  }
-}
-######## net-2 output ########
-layer {
-  name: "fc5_2"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc5_2"
-  inner_product_param {
-    num_output: 64
-  }
-}
-layer {
-  name: "prelu5_2"
-  type: "PReLU"
-  bottom: "fc5_2"
-  top: "fc5_2"
-}
-layer {
-  name: "fc6_2"
-  type: "InnerProduct"
-  bottom: "fc5_2"
-  top: "fc6_2"
-  inner_product_param {
-    num_output: 2
-  }
-}
-######## net-3 output ########
-layer {
-  name: "fc5_3"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc5_3"
-  inner_product_param {
-    num_output: 64
-  }
-}
-layer {
-  name: "prelu5_3"
-  type: "PReLU"
-  bottom: "fc5_3"
-  top: "fc5_3"
-}
-layer {
-  name: "fc6_3"
-  type: "InnerProduct"
-  bottom: "fc5_3"
-  top: "fc6_3"
-  inner_product_param {
-    num_output: 2
-  }
-}
-######## net-4 output ########
-layer {
-  name: "fc5_4"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc5_4"
-  inner_product_param {
-    num_output: 64
-  }
-}
-layer {
-  name: "prelu5_4"
-  type: "PReLU"
-  bottom: "fc5_4"
-  top: "fc5_4"
-}
-layer {
-  name: "fc6_4"
-  type: "InnerProduct"
-  bottom: "fc5_4"
-  top: "fc6_4"
-  inner_product_param {
-    num_output: 2
-  }
+  bottom: "conv4"
+  top: "conv4"
 }
-######## net-5 output ########
 layer {
-  name: "fc5_5"
-  type: "InnerProduct"
-  bottom: "fc4"
-  top: "fc5_5"
-  inner_product_param {
-    num_output: 64
+  name: "conv5"
+  type: "Convolution"
+  bottom: "conv4"
+  top: "conv5"
+  convolution_param {
+    num_output: 50
+    group: 5
+    kernel_size: 1
+    stride: 1
   }
 }
 layer {
-  name: "prelu5_5"
+  name: "prelu5"
   type: "PReLU"
-  bottom: "fc5_5"
-  top: "fc5_5"
+  bottom: "conv5"
+  top: "conv5"
 }
 layer {
-  name: "fc6_5"
-  type: "InnerProduct"
-  bottom: "fc5_5"
-  top: "fc6_5"
-  inner_product_param {
-    num_output: 2
+  name: "conv6"
+  type: "Convolution"
+  bottom: "conv5"
+  top: "conv6"
+  convolution_param {
+    num_output: 10
+    group: 5
+    kernel_size: 1
+    stride: 1
   }
 }
-
-######## final output ########
 layer {
-  name: "landmark_offset"
-  type: "Concat"
-  bottom: "fc6_1"
-  bottom: "fc6_2"
-  bottom: "fc6_3"
-  bottom: "fc6_4"
-  bottom: "fc6_5"
+  name: "landmark"
+  type: "Reshape"
+  bottom: "conv6"
   top: "landmark_offset"
-  concat_param {
-    axis: 1
+  reshape_param {
+    shape {
+      dim: 0
+      dim: 10
+    }
   }
 }
+2 −0		.gitignore
+1 −0		CMakeLists.txt
+8 −2		appveyor.yml
+3 −1		example/.gitignore
+3 −0		example/CMakeLists.txt
+4 −0		example/README.md
+1 −0		example/deeplandmark/landmark.cpp
+3,790 −0		example/models/r-fcn/test_agnostic.prototxt
+ −		example/r-fcn/000456.jpg
+ −		example/r-fcn/000542.jpg
+ −		example/r-fcn/001150.jpg
+ −		example/r-fcn/001763.jpg
+ −		example/r-fcn/004545.jpg
+188 −0		example/r-fcn/main.cpp
+3 −2		example/wgan/main.cpp
+1 −1		generatepb.bat
+1 −1		generatepb.sh
+13 −8		include/caffe/base.hpp
+4 −3		include/caffe/blob.hpp
+10 −0		include/caffe/c_api.h
+11 −0		include/caffe/logging.hpp
+4 −0		include/caffe/net.hpp
+12 −0		java/src/main/java/com/luoyetx/minicaffe/Net.java
+1 −0		java/src/test/java/MiniCaffeTest.java
+11 −0		python/minicaffe/net.py
+1 −0		python/tests/test.py
+12 −12		src/blob.cpp
+13 −48		src/c_api.cpp
+3 −2		src/common.cpp
+13 −3		src/common.hpp
+11 −0		src/jni/jni.c
+4 −0		src/layer.hpp
+3 −0		src/layers/base_conv_layer.hpp
+5 −0		src/layers/batch_norm_layer.hpp
+4 −3		src/layers/bn_layer.cpp
+4 −0		src/layers/bn_layer.hpp
+10 −8		src/layers/conv_layer.cpp
+11 −34		src/layers/cudnn/cudnn_conv_layer.cpp
+5 −1		src/layers/cudnn/cudnn_conv_layer.hpp
+12 −11		src/layers/lrn_layer.cpp
+5 −0		src/layers/mvn_layer.hpp
+4 −3		src/layers/pooling_layer.cpp
+315 −0		src/layers/proposal_layer.cpp
+278 −0		src/layers/proposal_layer.cu
+47 −0		src/layers/proposal_layer.hpp
+114 −0		src/layers/psroi_pooling_layer.cpp
+89 −0		src/layers/psroi_pooling_layer.cu
+55 −0		src/layers/psroi_pooling_layer.hpp
+4 −3		src/layers/relu_layer.cpp
+2 −4		src/layers/roi_pooling_layer.cpp
+4 −6		src/layers/roi_pooling_layer.cu
+4 −3		src/layers/sigmoid_layer.cpp
+4 −3		src/layers/softmax_layer.cpp
+4 −3		src/layers/tanh_layer.cpp
+31 −4		src/net.cpp
+54 −35		src/proto/caffe.proto
+256 −61		src/syncedmem.cpp
+71 −44		src/syncedmem.hpp
+54 −0		src/thread_local.hpp
+32 −0		tests/run_net.cpp
+34 −0		tools/benchmark.cpp
+199 −0		tools/fuse_bn.py
+132 −0		tools/parse_mem.py
+3 −0		tools/tools.cmake
+2 −0		travis.sh