commit 4d33630096c769dd43716dd5607f151e3d5abef7
author Michele Di Giorgio <michele.digiorgio@arm.com> Fri Mar 02 09:43:54 2018 +0000
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:49:37 2018 +0000
tree 762897c2acac9553c0dad688d0c21842c8edff16
parent 1cd41495153c4e89d6195b42f870967339c1a13b

COMPMID-987: Make beta and gamma optional in BatchNormalization

Currently we have beta and gamma compulsory in Batch normalization. There are
network that might not need one or both of those. Thus these should be optional
with beta(offset) defaulting to zero and gamma(scale) to 1. Will also reduce
some memory requirements.

Change-Id: I15bf1ec14b814be2acebf1be1a4fba9c4fbd3190
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/123237
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

tests/validation/fixtures/BatchNormalizationLayerFixture.h

diff --git a/tests/validation/fixtures/BatchNormalizationLayerFixture.h b/tests/validation/fixtures/BatchNormalizationLayerFixture.h
index e02c619..4a6ac1a 100644
--- a/tests/validation/fixtures/BatchNormalizationLayerFixture.h
+++ b/tests/validation/fixtures/BatchNormalizationLayerFixture.h
@@ -45,10 +45,12 @@
 {
 public:
     template <typename...>
-    void setup(TensorShape shape0, TensorShape shape1, float epsilon, ActivationLayerInfo act_info, DataType dt, int fractional_bits)
+    void setup(TensorShape shape0, TensorShape shape1, float epsilon, bool use_beta, bool use_gamma, ActivationLayerInfo act_info, DataType dt, int fractional_bits)
     {
         _fractional_bits = fractional_bits;
         _data_type       = dt;
+        _use_beta        = use_beta;
+        _use_gamma       = use_gamma;
         _target          = compute_target(shape0, shape1, epsilon, act_info, dt, fractional_bits);
         _reference       = compute_reference(shape0, shape1, epsilon, act_info, dt, fractional_bits);
     }
@@ -67,8 +69,24 @@
             library->fill(src_tensor, distribution, 0);
             library->fill(mean_tensor, distribution, 1);
             library->fill(var_tensor, distribution_var, 0);
-            library->fill(beta_tensor, distribution, 3);
-            library->fill(gamma_tensor, distribution, 4);
+            if(_use_beta)
+            {
+                library->fill(beta_tensor, distribution, 3);
+            }
+            else
+            {
+                // Fill with default value 0.f
+                library->fill_tensor_value(beta_tensor, 0.f);
+            }
+            if(_use_gamma)
+            {
+                library->fill(gamma_tensor, distribution, 4);
+            }
+            else
+            {
+                // Fill with default value 1.f
+                library->fill_tensor_value(gamma_tensor, 1.f);
+            }
         }
         else
         {
@@ -80,8 +98,24 @@
             library->fill(src_tensor, distribution, 0);
             library->fill(mean_tensor, distribution, 1);
             library->fill(var_tensor, distribution_var, 0);
-            library->fill(beta_tensor, distribution, 3);
-            library->fill(gamma_tensor, distribution, 4);
+            if(_use_beta)
+            {
+                library->fill(beta_tensor, distribution, 3);
+            }
+            else
+            {
+                // Fill with default value 0
+                library->fill_tensor_value(beta_tensor, static_cast<T>(0));
+            }
+            if(_use_gamma)
+            {
+                library->fill(gamma_tensor, distribution, 4);
+            }
+            else
+            {
+                // Fill with default value 1
+                library->fill_tensor_value(gamma_tensor, static_cast<T>(1 << (_fractional_bits)));
+            }
         }
     }
 
@@ -97,7 +131,9 @@
 
         // Create and configure function
         FunctionType norm;
-        norm.configure(&src, &dst, &mean, &var, &beta, &gamma, epsilon, act_info);
+        TensorType *beta_ptr  = _use_beta ? &beta : nullptr;
+        TensorType *gamma_ptr = _use_gamma ? &gamma : nullptr;
+        norm.configure(&src, &dst, &mean, &var, beta_ptr, gamma_ptr, epsilon, act_info);
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -149,6 +185,8 @@
     SimpleTensor<T> _reference{};
     int             _fractional_bits{};
     DataType        _data_type{};
+    bool            _use_beta{};
+    bool            _use_gamma{};
 };
 
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
@@ -156,9 +194,9 @@
 {
 public:
     template <typename...>
-    void setup(TensorShape shape0, TensorShape shape1, float epsilon, ActivationLayerInfo act_info, DataType dt)
+    void setup(TensorShape shape0, TensorShape shape1, float epsilon, bool use_beta, bool use_gamma, ActivationLayerInfo act_info, DataType dt)
     {
-        BatchNormalizationLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, epsilon, act_info, dt, 0);
+        BatchNormalizationLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, epsilon, use_beta, use_gamma, act_info, dt, 0);
     }
 };
 } // namespace validation