Auto merge of #87 - alexandermorozov:fix-sgd-init, r=hobofan

fix/sgd: initialize weight gradient history with zeroes

SGD solver used unintialized history tensors. If there were some NaNs then
whole network got poisoned after the first generation even if momentum
was set to zero. This patch prefills gradient history with zeros.

FIX: autumnai/leaf-examples#13

src/solvers/sgd/mod.rs

@@ -31,7 +31,13 @@ macro_rules! impl_isolver_sgd {
 
                 for weight_gradient in net.learnable_weights_gradients() {
                     let shape = weight_gradient.read().unwrap().desc().clone();
-                    let history_tensor = Arc::new(RwLock::new(SharedTensor::new(IBackend::device(&*self.backend), &shape).unwrap()));
+                    let mut tensor = SharedTensor::new(IBackend::device(&*self.backend),
+                                                       &shape).unwrap();
+
+                    let filler = ::weight::FillerType::Constant { value: 0f32 };
+                    filler.fill(&mut tensor);
+
+                    let history_tensor = Arc::new(RwLock::new(tensor));
                     self.history.push(history_tensor);
                 }
             }