Fixed weights-size for Depthwise grouped convolutional

src/convolutional_kernels.cu

@@ -518,7 +518,7 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
         printf("\n is_nan_or_inf(state.input) = %d \n", input_nan_inf);
         if (input_nan_inf) getchar();
 
-        int weights_nan_inf = is_nan_or_inf(l.weights_gpu, l.size * l.size * l.c * l.n);
+        int weights_nan_inf = is_nan_or_inf(l.weights_gpu, l.nweights);
         printf("\n is_nan_or_inf(l.weights_gpu) = %d \n", weights_nan_inf);
         if (weights_nan_inf) getchar();
         */
@@ -697,8 +697,8 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
         // calculate conv weight updates
         // Already: l.weight_updates_gpu = (l.weight_updates_gpu - l.weight*decay*batch*subdivision)*momentum
         //   so we should copy f32 to f16, or compute: f16=(w_up - w*d*b*s)*m
-        assert((l.c*l.n*l.size*l.size) > 0);
-        cuda_convert_f32_to_f16(l.weight_updates_gpu, l.c*l.n*l.size*l.size, l.weight_updates_gpu16);
+        assert((l.nweights) > 0);
+        cuda_convert_f32_to_f16(l.weight_updates_gpu, l.nweights, l.weight_updates_gpu16);
 
         CHECK_CUDNN(cudnnConvolutionBackwardFilter(cudnn_handle(),
             &one,
@@ -714,7 +714,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
             l.dweightDesc16,
             l.weight_updates_gpu16));    // l.weight_updates_gpu);
 
-        cuda_convert_f16_to_f32(l.weight_updates_gpu16, l.c*l.n*l.size*l.size, l.weight_updates_gpu);
+        cuda_convert_f16_to_f32(l.weight_updates_gpu16, l.nweights, l.weight_updates_gpu);
 
         if (state.delta) {
             if (l.binary || l.xnor) swap_binary(&l);
@@ -856,7 +856,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
         if (state.delta) {
             fix_nan_and_inf(state.delta, l.inputs * l.batch);
         }
-        int size = l.size * l.size * l.c * l.n;
+        int size = l.nweights;
         fix_nan_and_inf(l.weight_updates_gpu, size);
         fix_nan_and_inf(l.weights_gpu, size);
     }

src/convolutional_layer.c

@@ -1001,7 +1001,7 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
                     //size_t bit_input_size = intput_size / 8 + 1;
                     //char *bit_input = calloc(bit_input_size, sizeof(char));
 
-                    //size_t weights_size = k * m; //l.size*l.size*l.c*l.n;
+                    //size_t weights_size = k * m; //l.size*l.size*l.c*l.n; // l.nweights
                     //size_t bit_weights_size = weights_size / 8 + 1;
 
                     //char *bit_weights = calloc(bit_weights_size, sizeof(char));
@@ -1136,7 +1136,7 @@ void backward_convolutional_layer(convolutional_layer l, network_state state)
 
 void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
 {
-    //int size = l.size*l.size*l.c*l.n;
+    //int size = l.nweights;
     axpy_cpu(l.n, learning_rate / batch, l.bias_updates, 1, l.biases, 1);
     scal_cpu(l.n, momentum, l.bias_updates, 1);
 

src/network_kernels.cu

@@ -162,8 +162,8 @@ void forward_backward_network_gpu(network net, float *x, float *y)
         layer l = net.layers[i];
         if (net.cudnn_half){
             if (l.type == CONVOLUTIONAL && l.weights_gpu && l.weights_gpu16) {
-                assert((l.c*l.n*l.size*l.size) > 0);
-                cuda_convert_f32_to_f16(l.weights_gpu, l.c*l.n*l.size*l.size, l.weights_gpu16);
+                assert((l.nweights) > 0);
+                cuda_convert_f32_to_f16(l.weights_gpu, l.nweights, l.weights_gpu16);
             }
             else if (l.type == CRNN && l.input_layer->weights_gpu && l.input_layer->weights_gpu16) {
                 assert((l.input_layer->c*l.input_layer->n*l.input_layer->size*l.input_layer->size) > 0);