From 076009ebe308fde0156304e701f36e8bb04e4d6b Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 17 Jul 2014 17:14:59 +0000
Subject: [PATCH] Fixed batch stuff in conv layer

---
 src/network.c             |    2 
 src/convolutional_layer.c |   47 ++++++++++-----
 src/col2im.c              |   32 ++++------
 src/mini_blas.h           |    4 
 src/cnn.c                 |    9 +-
 src/im2col.c              |   33 ++++++++++
 6 files changed, 83 insertions(+), 44 deletions(-)

diff --git a/src/cnn.c b/src/cnn.c
index 73d172d..0acc842 100644
--- a/src/cnn.c
+++ b/src/cnn.c
@@ -48,11 +48,10 @@
 
 	image edge = make_image((dog.h-size)/stride+1, (dog.w-size)/stride+1, n);
 
-
 	int i;
 	clock_t start = clock(), end;
 	for(i = 0; i < 1000; ++i){
-		im2col_cpu(dog.data,  1, dog.c,  dog.h,  dog.w,  size,  stride, 0, matrix);
+		im2col_cpu(dog.data, dog.c,  dog.h,  dog.w,  size,  stride, 0, matrix);
 		gemm(0,0,n,mw,mh,1,filters,mh,matrix,mw,1,edge.data,mw);
 	}
 	end = clock();
@@ -317,8 +316,8 @@
 		clock_t start = clock(), end;
 		float loss = train_network_sgd(net, train, iters, lr, momentum, decay);
 		end = clock();
-		//float test_acc = network_accuracy(net, test);
-        float test_acc = 0;
+		float test_acc = network_accuracy(net, test);
+        //float test_acc = 0;
 		printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay);
 
 		//printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*1000, loss, lr, momentum, decay);
@@ -434,7 +433,7 @@
 	float *matrix = calloc(msize, sizeof(float));
 	int i;
 	for(i = 0; i < 1000; ++i){
-		im2col_cpu(test.data, 1, c,  h,  w,  size,  stride, 0, matrix);
+		im2col_cpu(test.data,  c,  h,  w,  size,  stride, 0, matrix);
 		//image render = float_to_image(mh, mw, mc, matrix);
 	}
 }
diff --git a/src/col2im.c b/src/col2im.c
index 0520567..bc15b7b 100644
--- a/src/col2im.c
+++ b/src/col2im.c
@@ -10,10 +10,10 @@
 }
 //This one might be too, can't remember.
 void col2im_cpu(float* data_col,
-        const int batch, const int channels, const int height, const int width,
+        const int channels, const int height, const int width,
         const int ksize, const int stride, int pad, float* data_im) 
 {
-    int c,h,w,b;
+    int c,h,w;
     int height_col = (height - ksize) / stride + 1;
     int width_col = (width - ksize) / stride + 1;
     if (pad){
@@ -22,25 +22,19 @@
         pad = ksize/2;
     }
     int channels_col = channels * ksize * ksize;
-    int im_size = height*width*channels;
-    int col_size = height_col*width_col*channels_col;
-    for (b = 0; b < batch; ++b) {
-        for (c = 0; c < channels_col; ++c) {
-            int w_offset = c % ksize;
-            int h_offset = (c / ksize) % ksize;
-            int c_im = c / ksize / ksize;
-            for (h = 0; h < height_col; ++h) {
-                for (w = 0; w < width_col; ++w) {
-                    int im_row = h_offset + h * stride;
-                    int im_col = w_offset + w * stride;
-                    double val = data_col[(c * height_col + h) * width_col + w];
-                    col2im_set_pixel(data_im, height, width, channels,
-                                    im_row, im_col, c_im, pad, val);
-                }
+    for (c = 0; c < channels_col; ++c) {
+        int w_offset = c % ksize;
+        int h_offset = (c / ksize) % ksize;
+        int c_im = c / ksize / ksize;
+        for (h = 0; h < height_col; ++h) {
+            for (w = 0; w < width_col; ++w) {
+                int im_row = h_offset + h * stride;
+                int im_col = w_offset + w * stride;
+                double val = data_col[(c * height_col + h) * width_col + w];
+                col2im_set_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad, val);
             }
         }
-        data_im += im_size;
-        data_col+= col_size;
     }
 }
 
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 7571e7a..44e9244 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -79,7 +79,7 @@
     layer->bias_updates_cl = cl_make_array(layer->bias_updates, n);
     layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n);
 
-    layer->col_image_cl = cl_make_array(layer->col_image, layer->batch*out_h*out_w*size*size*c);
+    layer->col_image_cl = cl_make_array(layer->col_image, layer.batch*out_h*out_w*size*size*c);
     layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n);
     layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*n);
     #endif
@@ -124,24 +124,32 @@
 {
     int out_h = convolutional_out_height(layer);
     int out_w = convolutional_out_width(layer);
+    int i;
+
+    bias_output(layer);
 
     int m = layer.n;
     int k = layer.size*layer.size*layer.c;
-    int n = out_h*out_w*layer.batch;
+    int n = out_h*out_w;
 
     float *a = layer.filters;
     float *b = layer.col_image;
     float *c = layer.output;
-    im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w, 
-        layer.size, layer.stride, layer.pad, b);
-    bias_output(layer);
-    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+
+    for(i = 0; i < layer.batch; ++i){
+        im2col_cpu(in, layer.c, layer.h, layer.w, 
+            layer.size, layer.stride, layer.pad, b);
+        gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+        c += n*m;
+        in += layer.h*layer.w*layer.c;
+        b += k*n;
+    }
     /*
     int i;
     for(i = 0; i < m*n; ++i) printf("%f, ", layer.output[i]);
     printf("\n");
     */
-    activate_array(layer.output, m*n, layer.activation, 0.);
+    activate_array(layer.output, m*n*layer.batch, layer.activation, 0.);
 }
 
 #ifdef GPU
@@ -178,35 +186,42 @@
 
 void backward_convolutional_layer(convolutional_layer layer, float *delta)
 {
+    int i;
     int m = layer.n;
     int n = layer.size*layer.size*layer.c;
     int k = convolutional_out_height(layer)*
-        convolutional_out_width(layer)*
-        layer.batch;
-    gradient_array(layer.output, m*k, layer.activation, layer.delta);
+        convolutional_out_width(layer);
+    gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
     learn_bias_convolutional_layer(layer);
 
     float *a = layer.delta;
     float *b = layer.col_image;
     float *c = layer.filter_updates;
 
-    gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+    for(i = 0; i < layer.batch; ++i){
+        gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+        a += m*k;
+        b += k*n;
+    }
 
     if(delta){
         m = layer.size*layer.size*layer.c;
         k = layer.n;
         n = convolutional_out_height(layer)*
-            convolutional_out_width(layer)*
-            layer.batch;
+            convolutional_out_width(layer);
 
         a = layer.filters;
         b = layer.delta;
         c = layer.col_image;
 
-        gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
-
         memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
-        col2im_cpu(c, layer.batch,  layer.c,  layer.h,  layer.w,  layer.size,  layer.stride, layer.pad, delta);
+
+        for(i = 0; i < layer.batch; ++i){
+            gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
+            col2im_cpu(c, layer.c,  layer.h,  layer.w,  layer.size,  layer.stride, layer.pad, delta);
+            c += k*n;
+            delta += layer.h*layer.w*layer.c;
+        }
     }
 }
 
diff --git a/src/im2col.c b/src/im2col.c
index 304bee7..89748c9 100644
--- a/src/im2col.c
+++ b/src/im2col.c
@@ -14,7 +14,7 @@
 
 //From Berkeley Vision's Caffe!
 //https://github.com/BVLC/caffe/blob/master/LICENSE
-void im2col_cpu(float* data_im,
+void im2col_cpu_batch(float* data_im,
     const int batch, const int channels, const int height, const int width,
     const int ksize, const int stride, int pad, float* data_col) 
 {
@@ -49,6 +49,37 @@
     }
 }
 
+//From Berkeley Vision's Caffe!
+//https://github.com/BVLC/caffe/blob/master/LICENSE
+void im2col_cpu(float* data_im,
+    const int channels, const int height, const int width,
+    const int ksize, const int stride, int pad, float* data_col) 
+{
+    int c,h,w;
+    int height_col = (height - ksize) / stride + 1;
+    int width_col = (width - ksize) / stride + 1;
+    if (pad){
+        height_col = 1 + (height-1) / stride;
+        width_col = 1 + (width-1) / stride;
+        pad = ksize/2;
+    }
+    int channels_col = channels * ksize * ksize;
+    for (c = 0; c < channels_col; ++c) {
+        int w_offset = c % ksize;
+        int h_offset = (c / ksize) % ksize;
+        int c_im = c / ksize / ksize;
+        for (h = 0; h < height_col; ++h) {
+            for (w = 0; w < width_col; ++w) {
+                int im_row = h_offset + h * stride;
+                int im_col = w_offset + w * stride;
+                int col_index = (c * height_col + h) * width_col + w;
+                data_col[col_index] = im2col_get_pixel(data_im, height, width, channels,
+                        im_row, im_col, c_im, pad);
+            }
+        }
+    }
+}
+
 
 #ifdef GPU
 
diff --git a/src/mini_blas.h b/src/mini_blas.h
index bf5debb..95e924b 100644
--- a/src/mini_blas.h
+++ b/src/mini_blas.h
@@ -26,11 +26,11 @@
 #endif
 
 void im2col_cpu(float* data_im,
-    const int batch, const int channels, const int height, const int width,
+    const int channels, const int height, const int width,
     const int ksize, const int stride, int pad, float* data_col);
 
 void col2im_cpu(float* data_col,
-        const int batch, const int channels, const int height, const int width,
+        const int channels, const int height, const int width,
         const int ksize, const int stride, int pad, float* data_im);
 void test_blas();
 
diff --git a/src/network.c b/src/network.c
index 6855c55..7088398 100644
--- a/src/network.c
+++ b/src/network.c
@@ -274,7 +274,7 @@
         //printf("%5.2f %5.2f, ", out[i], truth[i]);
         //if(i == get_network_output_size(net)) printf("\n");
         delta[i] = truth[i] - out[i];
-        //printf("%f, ", delta[i]);
+        //printf("%.10f, ", out[i]);
         sum += delta[i]*delta[i];
     }
     //printf("\n");

--
Gitblit v1.10.0