Minor fix for mAP calculation during training

src/detector.c

@@ -67,6 +67,7 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
         }
 #ifdef GPU
         cuda_free(net_map.workspace);
+        cuda_free(net_map.input_state_gpu);
         if (*net_map.input16_gpu) cuda_free(*net_map.input16_gpu);
         if (*net_map.output16_gpu) cuda_free(*net_map.output16_gpu);
 #else
@@ -245,7 +246,7 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
 #ifdef OPENCV
         if (!dont_show) {
             int draw_precision = 0;
-            int calc_map_for_each = 4 * train_images_num / (net.batch * net.subdivisions);
+            int calc_map_for_each = 4 * train_images_num / (net.batch * net.subdivisions);  // calculate mAP for each 4 Epochs
             if (calc_map && (i >= (iter_map + calc_map_for_each) || i == net.max_batches) && i >= net.burn_in && i >= 1000) {
                 if (l.random) {
                     printf("Resizing to initial size: %d x %d \n", init_w, init_h);
@@ -262,32 +263,7 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
                 }
 
                 // combine Training and Validation networks
-                network net_combined = make_network(net.n);
-                layer *old_layers = net_combined.layers;
-                net_combined = net;
-                net_combined.layers = old_layers;
-                net_combined.batch = 1;
-
-                int k;
-                for (k = 0; k < net.n; ++k) {
-                    layer *l = &(net.layers[k]);
-                    net_combined.layers[k] = net.layers[k];
-                    net_combined.layers[k].batch = 1;
-
-                    if (l->type == CONVOLUTIONAL) {
-#ifdef CUDNN
-                        net_combined.layers[k].normTensorDesc = net_map.layers[k].normTensorDesc;
-                        net_combined.layers[k].normDstTensorDesc = net_map.layers[k].normDstTensorDesc;
-                        net_combined.layers[k].normDstTensorDescF16 = net_map.layers[k].normDstTensorDescF16;
-
-                        net_combined.layers[k].srcTensorDesc = net_map.layers[k].srcTensorDesc;
-                        net_combined.layers[k].dstTensorDesc = net_map.layers[k].dstTensorDesc;
-
-                        net_combined.layers[k].srcTensorDesc16 = net_map.layers[k].srcTensorDesc16;
-                        net_combined.layers[k].dstTensorDesc16 = net_map.layers[k].dstTensorDesc16;
-#endif // CUDNN
-                    }
-                }
+                network net_combined = combine_train_valid_networks(net, net_map);
 
                 iter_map = i;
                 mean_average_precision = validate_detector_map(datacfg, cfgfile, weightfile, 0.25, 0.5, &net_combined);

src/image.c

@@ -710,7 +710,7 @@ IplImage* draw_train_chart(float max_img_loss, int max_batches, int number_of_li
     char max_batches_buff[100];
     sprintf(max_batches_buff, "in cfg max_batches=%d", max_batches);
     cvPutText(img, max_batches_buff, cvPoint(draw_size - 195, img_size - 10), &font, CV_RGB(0, 0, 0));
-    cvPutText(img, "Press 's' to save: chart.jpg", cvPoint(5, img_size - 10), &font, CV_RGB(0, 0, 0));
+    cvPutText(img, "Press 's' to save: chart.png", cvPoint(5, img_size - 10), &font, CV_RGB(0, 0, 0));
     printf(" If error occurs - run training with flag: -dont_show \n");
     cvNamedWindow("average loss", CV_WINDOW_NORMAL);
     cvMoveWindow("average loss", 0, 0);
@@ -770,7 +770,11 @@ void draw_train_loss(IplImage* img, int img_size, float avg_loss, float max_img_
     cvShowImage("average loss", img);
     int k = cvWaitKey(20);
     if (k == 's' || current_batch == (max_batches - 1) || current_batch % 100 == 0) {
-        cvSaveImage("chart.jpg", img, 0);
+        //cvSaveImage("chart.jpg", img, 0);
+        IplImage* img_rgb = cvCreateImage(cvSize(img->width, img->height), 8, 3);
+        cvCvtColor(img, img_rgb, CV_RGB2BGR);
+        stbi_write_png("chart.png", img_rgb->width, img_rgb->height, 3, (char *)img_rgb->imageData, 0);
+        cvRelease(&img_rgb);
         cvPutText(img, "- Saved", cvPoint(250, img_size - 10), &font, CV_RGB(255, 0, 0));
     }
     else

src/network.c

@@ -878,4 +878,36 @@ void calculate_binary_weights(network net)
     }
     //printf("\n calculate_binary_weights Done! \n");
 
+}
+
+// combine Training and Validation networks
+network combine_train_valid_networks(network net_train, network net_map)
+{
+    network net_combined = make_network(net_train.n);
+    layer *old_layers = net_combined.layers;
+    net_combined = net_train;
+    net_combined.layers = old_layers;
+    net_combined.batch = 1;
+
+    int k;
+    for (k = 0; k < net_train.n; ++k) {
+        layer *l = &(net_train.layers[k]);
+        net_combined.layers[k] = net_train.layers[k];
+        net_combined.layers[k].batch = 1;
+
+        if (l->type == CONVOLUTIONAL) {
+#ifdef CUDNN
+            net_combined.layers[k].normTensorDesc = net_map.layers[k].normTensorDesc;
+            net_combined.layers[k].normDstTensorDesc = net_map.layers[k].normDstTensorDesc;
+            net_combined.layers[k].normDstTensorDescF16 = net_map.layers[k].normDstTensorDescF16;
+
+            net_combined.layers[k].srcTensorDesc = net_map.layers[k].srcTensorDesc;
+            net_combined.layers[k].dstTensorDesc = net_map.layers[k].dstTensorDesc;
+
+            net_combined.layers[k].srcTensorDesc16 = net_map.layers[k].srcTensorDesc16;
+            net_combined.layers[k].dstTensorDesc16 = net_map.layers[k].dstTensorDesc16;
+#endif // CUDNN
+        }
+    }
+    return net_combined;
 }

src/network.h

@@ -156,6 +156,7 @@ int get_network_nuisance(network net);
 int get_network_background(network net);
 YOLODLL_API void fuse_conv_batchnorm(network net);
 YOLODLL_API void calculate_binary_weights(network net);
+network combine_train_valid_networks(network net_train, network net_map);
 
 #ifdef __cplusplus
 }