I WISH I HAD SOME TESTS THOUGH

examples/detector.py

@@ -7,11 +7,12 @@ import sys, os
 sys.path.append(os.path.join(os.getcwd(),'python/'))
 
 import darknet as dn
+import pdb
 
 dn.set_gpu(0)
-net = dn.load_net("cfg/tiny-yolo.cfg", "tiny-yolo.weights", 0)
-meta = dn.load_meta("cfg/coco.data")
-r = dn.detect(net, meta, "data/dog.jpg")
+net = dn.load_net("cfg/yolo-thor.cfg", "/home/pjreddie/backup/yolo-thor_final.weights", 0)
+meta = dn.load_meta("cfg/thor.data")
+r = dn.detect(net, meta, "data/bedroom.jpg")
 print r
 
 # And then down here you could detect a lot more images like:

examples/nightmare.c

@@ -62,7 +62,7 @@ void optimize_picture(network *net, image orig, int max_layer, float scale, floa
     cuda_free(net->delta_gpu);
     net->delta_gpu = 0;
 #else
-    net->input = im.data;
+    copy_cpu(net->inputs, im.data, 1, net->input, 1);
     net->delta = delta.data;
     forward_network(net);
     copy_cpu(last.outputs, last.output, 1, last.delta, 1);

python/darknet.py

@@ -13,7 +13,9 @@ def sample(probs):
     return len(probs)-1
 
 def c_array(ctype, values):
-    return (ctype * len(values))(*values)
+    arr = (ctype*len(values))()
+    arr[:] = values
+    return arr
 
 class BOX(Structure):
     _fields_ = [("x", c_float),

src/convolutional_kernels.cu

@@ -236,7 +236,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network net)
             float *b = net.workspace;
             float *c = l.weight_updates_gpu + j*l.nweights/l.groups;
 
-            float *im = net.input+(i*l.groups + j)*l.c/l.groups*l.h*l.w;
+            float *im = net.input_gpu+(i*l.groups + j)*l.c/l.groups*l.h*l.w;
 
             im2col_gpu(im, l.c/l.groups, l.h, l.w,
                     l.size, l.stride, l.pad, b);

src/convolutional_layer.c

@@ -150,24 +150,24 @@ void cudnn_convolutional_setup(layer *l)
             l->weightDesc,
             l->convDesc,
             l->dstTensorDesc,
-            CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
-            0,
+            CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
+            4000000000,
             &l->fw_algo);
     cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
             l->weightDesc,
             l->ddstTensorDesc,
             l->convDesc,
             l->dsrcTensorDesc,
-            CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
-            0,
+            CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
+            4000000000,
             &l->bd_algo);
     cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
             l->srcTensorDesc,
             l->ddstTensorDesc,
             l->convDesc,
             l->dweightDesc,
-            CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
-            0,
+            CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
+            4000000000,
             &l->bf_algo);
 }
 #endif

src/network.c

@@ -389,6 +389,7 @@ int resize_network(network *net, int w, int h)
             error("Cannot resize this type of layer");
         }
         if(l.workspace_size > workspace_size) workspace_size = l.workspace_size;
+        if(l.workspace_size > 2000000000) assert(0);
         inputs = l.outputs;
         net->layers[i] = l;
         w = l.out_w;