[Application] Fix Resnet18

Fix it because the results of the computation of pytorch and nntrainer
are different.

padding was written as same when stride was 2 in NNtrainer Resnet code.
In pytorch, this is set to error. In addition, padding was not applied
normally in nntrainer, so the results were different. To solve this
problem, the parameters of a1 layer (conv layer) have been modified.

Additionally, momentum and epsilon were added to batch_norm layer.

Signed-off-by: SeoHyungjun <hyungjun.seo@samsung.com>

diff --git a/Applications/Resnet/jni/main.cpp b/Applications/Resnet/jni/main.cpp
index 074f4537940c33732cac485c6484febec824baf1..0d43f855c6b02f2285b462b3f726ef9382e1136f 100644 (file)
--- a/Applications/Resnet/jni/main.cpp
+++ b/Applications/Resnet/jni/main.cpp
@@ -113,15 +113,18 @@ std::vector<LayerHandle> resnetBlock(const std::string &block_name,
   };
 
   /** residual path */
-  LayerHandle a1 = create_conv("a1", 3, downsample ? 2 : 1, "same", input_name);
-  LayerHandle a2 = createLayer(
-    "batch_normalization", {with_name("a2"), withKey("activation", "relu")});
+  LayerHandle a1 = create_conv("a1", 3, downsample ? 2 : 1,
+                               downsample ? "1,1" : "same", input_name);
+  LayerHandle a2 =
+    createLayer("batch_normalization",
+                {with_name("a2"), withKey("activation", "relu"),
+                 withKey("momentum", "0.9"), withKey("epsilon", "0.00001")});
   LayerHandle a3 = create_conv("a3", 3, 1, "same", scoped_name("a2"));
 
   /** skip path */
   LayerHandle b1 = nullptr;
   if (downsample) {
-    b1 = create_conv("b1", 1, 2, "same", input_name);
+    b1 = create_conv("b1", 1, 2, "0,0", input_name);
   }
 
   const std::string skip_name = b1 ? scoped_name("b1") : input_name;
@@ -132,8 +135,8 @@ std::vector<LayerHandle> resnetBlock(const std::string &block_name,
 
   LayerHandle c2 =
     createLayer("batch_normalization",
-                {withKey("name", block_name), withKey("activation", "relu")});
-
+                {withKey("name", block_name), withKey("activation", "relu"),
+                 withKey("momentum", "0.9"), withKey("epsilon", "0.00001")});
   if (downsample) {
     return {b1, a1, a2, a3, c1, c2};
   } else {
@@ -165,9 +168,10 @@ std::vector<LayerHandle> createResnet18Graph() {
                             withKey("weight_initializer", "xavier_uniform"),
                           }));
 
-  layers.push_back(
-    createLayer("batch_normalization", {withKey("name", "first_bn_relu"),
-                                        withKey("activation", "relu")}));
+  layers.push_back(createLayer(
+    "batch_normalization",
+    {withKey("name", "first_bn_relu"), withKey("activation", "relu"),
+     withKey("momentum", "0.9"), withKey("epsilon", "0.00001")}));
 
   std::vector<std::vector<LayerHandle>> blocks;
 
@@ -184,9 +188,9 @@ std::vector<LayerHandle> createResnet18Graph() {
     layers.insert(layers.end(), block.begin(), block.end());
   }
 
-  layers.push_back(createLayer("pooling2d", {withKey("name", "last_p1"),
-                                             withKey("pooling", "average"),
-                                             withKey("pool_size", {4, 4})}));
+  layers.push_back(createLayer(
+    "pooling2d", {withKey("name", "last_p1"), withKey("pooling", "average"),
+                  withKey("pool_size", {4, 4}), withKey("stride", "4,4")}));
 
   layers.push_back(createLayer("flatten", {withKey("name", "last_f1")}));
   layers.push_back(