Fix Adam optimizer unit test

unit_tests/optimizer/Test_Adam.cpp

@@ -70,7 +70,7 @@ TEST_CASE("[learning/Adam] update", "[Optimizer][Adam]") {
             val_tensors[i] = std::make_unique<float[]>(size_tensors[i]);
             val_grad_tensors[i] = std::make_unique<float[]>(size_tensors[i]);
             val_momentum1_tensors[i] = std::make_unique<float[]>(size_tensors[i]);
-			val_momentum2_tensors[i] = std::make_unique<float[]>(size_tensors[i]);
+            val_momentum2_tensors[i] = std::make_unique<float[]>(size_tensors[i]);
             for (std::size_t j = 0; j < size_tensors[i]; ++j) {
                 val_tensors[i][j] = valueDist(gen);
                 val_grad_tensors[i][j] = valueDist(gen);
@@ -92,7 +92,7 @@ TEST_CASE("[learning/Adam] update", "[Optimizer][Adam]") {
             momentum_tensors[i] = std::make_shared<Tensor>(dims);
             momentum_tensors[i]->setBackend("cpu");
             momentum_tensors[i]->getImpl()->setRawPtr(val_momentum1_tensors[i].get(), size_tensors[i]);
-			momentum_tensors[i]->getImpl()->setRawPtr(val_momentum2_tensors[i].get(), size_tensors[i]);
+            momentum_tensors[i]->getImpl()->setRawPtr(val_momentum2_tensors[i].get(), size_tensors[i]);
 
             REQUIRE((tensors[i]->hasImpl() &&
                     optim_tensors[i]->hasImpl() &&
@@ -102,7 +102,7 @@ TEST_CASE("[learning/Adam] update", "[Optimizer][Adam]") {
         // generate parameters
         float lr = paramDist(gen);
         float beta1 = paramDist(gen);
-		float beta2 = paramDist(gen);
+        float beta2 = paramDist(gen);
         float epsilon = paramDist(gen);
 
         // set Optimizer
@@ -121,13 +121,14 @@ TEST_CASE("[learning/Adam] update", "[Optimizer][Adam]") {
 
         // truth
         for (std::size_t step = 0; step < 10; ++step) {
+            float lr2 = lr * std::sqrt(1.0f - std::pow(beta1, step + 1)) / (1.0f - std::pow(beta1, step + 1));
+            float epsilon2 = epsilon * std::sqrt(1.0f - std::pow(beta2, step + 1));
             for (std::size_t t = 0; t < nb_tensors; ++t) {
                 for (std::size_t i = 0; i < size_tensors[t]; ++i) {
                     val_momentum1_tensors[t][i] = beta1 * val_momentum1_tensors[t][i] + (1.0f - beta1) * val_grad_tensors[t][i];
                     val_momentum2_tensors[t][i] = beta2 * val_momentum2_tensors[t][i] + (1.0f - beta2) * val_grad_tensors[t][i] * val_grad_tensors[t][i];
                     val_tensors[t][i] = val_tensors[t][i]
-                                      - lr * val_momentum1_tensors[t][i] / (1.0f -  std::pow(beta1, step + 1))
-                                           / (std::sqrt(val_momentum2_tensors[t][i] / (1.0f - std::pow(beta2, step + 1))) + epsilon);
+                                      - lr2 * val_momentum1_tensors[t][i] / (std::sqrt(val_momentum2_tensors[t][i]) +  epsilon2);
                 }
             }
             // optimizer