From 08e261f42301b8f326bd6db28a35e84e0212142f Mon Sep 17 00:00:00 2001
From: nihui <shuizhuyuanluo@126.com>
Date: Thu, 18 Jan 2018 23:10:30 +0800
Subject: [PATCH] innerproduct produce continous blob, fix #236

---
 examples/fasterrcnn.cpp                      | 16 ++++-----
 examples/squeezencnn/jni/squeezencnn_jni.cpp |  2 +-
 src/layer/arm/innerproduct_arm.cpp           | 36 ++------------------
 src/layer/innerproduct.cpp                   |  5 ++-
 src/layer/scale.cpp                          |  4 +--
 5 files changed, 13 insertions(+), 50 deletions(-)

diff --git a/examples/fasterrcnn.cpp b/examples/fasterrcnn.cpp
index c3091f20e..6b82de8a5 100644
--- a/examples/fasterrcnn.cpp
+++ b/examples/fasterrcnn.cpp
@@ -174,7 +174,7 @@ static int detect_fasterrcnn(const cv::Mat& bgr, std::vector<Object>& objects)
         ex2.extract("bbox_pred", bbox_pred);
         ex2.extract("cls_prob", cls_prob);
 
-        int num_class = cls_prob.c;
+        int num_class = cls_prob.w;
         class_candidates.resize(num_class);
 
         // find class id with highest score
@@ -182,7 +182,7 @@ static int detect_fasterrcnn(const cv::Mat& bgr, std::vector<Object>& objects)
         float score = 0.f;
         for (int i=0; i<num_class; i++)
         {
-            float class_score = cls_prob.channel(i)[0];
+            float class_score = cls_prob[i];
             if (class_score > score)
             {
                 label = i;
@@ -206,14 +206,10 @@ static int detect_fasterrcnn(const cv::Mat& bgr, std::vector<Object>& objects)
         float pb_h = y2 - y1 + 1;
 
         // apply bbox regression
-        const float* bbox_xptr = bbox_pred.channel(label * 4);
-        const float* bbox_yptr = bbox_pred.channel(label * 4 + 1);
-        const float* bbox_wptr = bbox_pred.channel(label * 4 + 2);
-        const float* bbox_hptr = bbox_pred.channel(label * 4 + 3);
-        float dx = bbox_xptr[0];
-        float dy = bbox_yptr[0];
-        float dw = bbox_wptr[0];
-        float dh = bbox_hptr[0];
+        float dx = bbox_pred[label * 4];
+        float dy = bbox_pred[label * 4 + 1];
+        float dw = bbox_pred[label * 4 + 2];
+        float dh = bbox_pred[label * 4 + 3];
 
         float cx = x1 + pb_w * 0.5f;
         float cy = y1 + pb_h * 0.5f;
diff --git a/examples/squeezencnn/jni/squeezencnn_jni.cpp b/examples/squeezencnn/jni/squeezencnn_jni.cpp
index 036f14358..7ff436b5a 100644
--- a/examples/squeezencnn/jni/squeezencnn_jni.cpp
+++ b/examples/squeezencnn/jni/squeezencnn_jni.cpp
@@ -146,7 +146,7 @@ JNIEXPORT jstring JNICALL Java_com_tencent_squeezencnn_SqueezeNcnn_Detect(JNIEnv
         cls_scores.resize(out.c);
         for (int j=0; j<out.c; j++)
         {
-            const float* prob = out.data + out.cstep * j;
+            const float* prob = out.channel(j);
             cls_scores[j] = prob[0];
         }
     }
diff --git a/src/layer/arm/innerproduct_arm.cpp b/src/layer/arm/innerproduct_arm.cpp
index 3b1f1e48c..59941d294 100644
--- a/src/layer/arm/innerproduct_arm.cpp
+++ b/src/layer/arm/innerproduct_arm.cpp
@@ -29,47 +29,15 @@ int InnerProduct_arm::forward(const Mat& bottom_blob, Mat& top_blob) const
     int channels = bottom_blob.c;
     int size = w * h;
 
-    top_blob.create(1, 1, num_output);
+    top_blob.create(num_output);
     if (top_blob.empty())
         return -100;
 
-    if (size == 1)
-    {
-        // num_output
-        const float* weight_data_ptr = weight_data;
-        #pragma omp parallel for
-        for (int p=0; p<num_output; p++)
-        {
-            float* outptr = top_blob.channel(p);
-            float sum = 0.f;
-
-            if (bias_term)
-                sum = bias_data[p];
-
-            const float* w = weight_data_ptr + channels * p;
-
-            // channels
-            const float* m = bottom_blob;
-            for (int q=0; q<channels; q++)
-            {
-                sum += *m * *w;
-
-                m += 4;
-                w++;
-            }
-
-            outptr[0] = sum;
-        }
-
-        return 0;
-    }
-
     // num_output
     const float* weight_data_ptr = weight_data;
     #pragma omp parallel for
     for (int p=0; p<num_output; p++)
     {
-        float* outptr = top_blob.channel(p);
         float sum = 0.f;
 
         if (bias_term)
@@ -157,7 +125,7 @@ int InnerProduct_arm::forward(const Mat& bottom_blob, Mat& top_blob) const
 #endif // __aarch64__
 #endif // __ARM_NEON
 
-        outptr[0] = sum;
+        top_blob[p] = sum;
     }
 
     return 0;
diff --git a/src/layer/innerproduct.cpp b/src/layer/innerproduct.cpp
index 88f15671a..91d8b4049 100644
--- a/src/layer/innerproduct.cpp
+++ b/src/layer/innerproduct.cpp
@@ -56,7 +56,7 @@ int InnerProduct::forward(const Mat& bottom_blob, Mat& top_blob) const
     int channels = bottom_blob.c;
     int size = w * h;
 
-    top_blob.create(1, 1, num_output);
+    top_blob.create(num_output);
     if (top_blob.empty())
         return -100;
 
@@ -64,7 +64,6 @@ int InnerProduct::forward(const Mat& bottom_blob, Mat& top_blob) const
     #pragma omp parallel for
     for (int p=0; p<num_output; p++)
     {
-        float* outptr = top_blob.channel(p);
         float sum = 0.f;
 
         if (bias_term)
@@ -82,7 +81,7 @@ int InnerProduct::forward(const Mat& bottom_blob, Mat& top_blob) const
             }
         }
 
-        outptr[0] = sum;
+        top_blob[p] = sum;
     }
 
     return 0;
diff --git a/src/layer/scale.cpp b/src/layer/scale.cpp
index 0a222e4ea..a1ad85e17 100644
--- a/src/layer/scale.cpp
+++ b/src/layer/scale.cpp
@@ -71,7 +71,7 @@ int Scale::forward_inplace(std::vector<Mat>& bottom_top_blobs) const
         {
             float* ptr = bottom_top_blob.channel(q);
 
-            float s = scale_blob.channel(q)[0];
+            float s = scale_blob[q];
             float bias = bias_data[q];
 
             for (int i=0; i<size; i++)
@@ -87,7 +87,7 @@ int Scale::forward_inplace(std::vector<Mat>& bottom_top_blobs) const
         {
             float* ptr = bottom_top_blob.channel(q);
 
-            float s = scale_blob.channel(q)[0];
+            float s = scale_blob[q];
 
             for (int i=0; i<size; i++)
             {