stochastic pooling: avoid nan

src/caffe/layers/pooling_layer.cu

@@ -120,7 +120,8 @@ __global__ void StoPoolForwardTest(const int nthreads,
     int hend = min(hstart + ksize, height);
     int wstart = pw * stride;
     int wend = min(wstart + ksize, width);
-    Dtype cumsum = 0.;
+    // We set cumsum to be 0 to avoid divide-by-zero problems
+    Dtype cumsum = FLT_MIN;
     Dtype cumvalues = 0.;
     bottom_data += (n * channels + c) * height * width;
     // First pass: get sum

src/caffe/net.cpp

@@ -287,7 +287,7 @@ void Net<Dtype>::CopyTrainedLayersFrom(const NetParameter& param) {
       DLOG(INFO) << "Ignoring source layer " << source_layer_name;
       continue;
     }
-    DLOG(INFO) << "Loading source layer " << source_layer_name;
+    LOG(INFO) << "Copying source layer " << source_layer_name;
     vector<shared_ptr<Blob<Dtype> > >& target_blobs =
         layers_[target_layer_id]->blobs();
     CHECK_EQ(target_blobs.size(), source_layer.blobs_size())