[split layer] make a unittest to test split input dimension by split number

 - Make a unittest to test split input dimension by split number
 - Conv2d layer is added to make model has a weight

Signed-off-by: hyeonseok lee <hs89.lee@samsung.com>

diff --git a/nntrainer/layers/common_properties.h b/nntrainer/layers/common_properties.h
index a5a9c46..f99fd96 100644 (file)
--- a/nntrainer/layers/common_properties.h
+++ b/nntrainer/layers/common_properties.h
@@ -250,8 +250,8 @@ public:
 };
 
 /**
- * @brief split number property, split number is used to split input dimension
- * in split layer
+ * @brief split number property, split number indicates how many numbers of outs
+ * are generated by spliting the input dimension
  *
  */
 class SplitNumber : public PositiveIntegerProperty {

diff --git a/nntrainer/layers/split_layer.cpp b/nntrainer/layers/split_layer.cpp
index 4e55d81..2b15f07 100644 (file)
--- a/nntrainer/layers/split_layer.cpp
+++ b/nntrainer/layers/split_layer.cpp
@@ -34,20 +34,22 @@ void SplitLayer::finalize(InitLayerContext &context) {
 
   unsigned int split_dimension = std::get<props::SplitDimension>(split_props);
 
+  const TensorDim &in_dim = context.getInputDimensions()[0];
+
   if (std::get<props::SplitNumber>(split_props).empty()) {
     std::get<props::SplitNumber>(split_props)
-      .set(context.getNumRequestedOutputs());
+      .set(in_dim.getTensorDim(split_dimension));
   }
   unsigned int split_number = std::get<props::SplitNumber>(split_props);
 
   /**
    * The split is only done along the split_dimension dimension.
+   * (Assumes input data is continous)
    * For example, consider input dimension [b,c,h,w], split_number = n
-   * 1. axis = 1, output_dim = [b,n,h,w], num_outputs = c//n
-   * 2. axis = 2, output_dim = [b,c,n,w], num_outputs = h//n
-   * 3. axis = 3, output_dim = [b,c,h,n], num_outputs = w//n
+   * 1. axis = 1, output_dim = [b,c//n,h,w], num_outputs = n
+   * 2. axis = 2, output_dim = [b,c,h//n,w], num_outputs = n
+   * 3. axis = 3, output_dim = [b,c,h,w//n], num_outputs = n
    */
-  const TensorDim &in_dim = context.getInputDimensions()[0];
   NNTR_THROW_IF(split_number != context.getNumRequestedOutputs(),
                 std::invalid_argument)
     << "Given split number does not match with number of outputs";

diff --git a/packaging/unittest_models_multiout.tar.gz b/packaging/unittest_models_multiout.tar.gz
index 689f8db..ce44df3 100644 (file)
Binary files a/packaging/unittest_models_multiout.tar.gz and b/packaging/unittest_models_multiout.tar.gz differ

diff --git a/test/input_gen/genModelsMultiout_v2.py b/test/input_gen/genModelsMultiout_v2.py
index 6fff417..20a4996 100644 (file)
--- a/test/input_gen/genModelsMultiout_v2.py
+++ b/test/input_gen/genModelsMultiout_v2.py
@@ -11,6 +11,24 @@
 from recorder_v2 import record_v2, inspect_file
 import torch
 
+class Split(torch.nn.Module):
+    def __init__(self, axis, split_number, channel):
+        super().__init__()
+        self.axis = axis
+        self.split_number = split_number
+        self.conv = torch.nn.Conv2d(channel, channel, 1)
+        self.loss = torch.nn.MSELoss()
+
+    def forward(self, inputs, labels):
+        outs = self.conv(inputs[0])
+        split_size = outs.size(self.axis) // self.split_number
+        *outs, = torch.split(outs, split_size, self.axis)
+        out = torch.clone(outs[0])
+        for i in range(1, len(outs)):
+            out += outs[i]
+
+        loss = self.loss(out, labels[0])
+        return out, loss
 
 class SplitAndJoin(torch.nn.Module):
     def __init__(self):
@@ -93,6 +111,30 @@ class OneToMany(torch.nn.Module):
 
 if __name__ == "__main__":
     record_v2(
+        Split(3, 5, 3),
+        iteration=2,
+        input_dims=[(2, 3, 4, 5)],
+        label_dims=[(2, 3, 4, 1)],
+        name="split_axis3_split_number5"
+    )
+
+    record_v2(
+        Split(2, 4, 3),
+        iteration=2,
+        input_dims=[(2, 3, 4, 5)],
+        label_dims=[(2, 3, 1, 5)],
+        name="split_axis2_split_number4"
+    )
+
+    record_v2(
+        Split(2, 2, 3),
+        iteration=2,
+        input_dims=[(2, 3, 4, 5)],
+        label_dims=[(2, 3, 2, 5)],
+        name="split_axis2_split_number2"
+    )
+
+    record_v2(
         SplitAndJoin(),
         iteration=2,
         input_dims=[(5, 3)],

diff --git a/test/unittest/models/unittest_models_multiout.cpp b/test/unittest/models/unittest_models_multiout.cpp
index 2f330b7..e262b7c 100644 (file)
--- a/test/unittest/models/unittest_models_multiout.cpp
+++ b/test/unittest/models/unittest_models_multiout.cpp
@@ -22,6 +22,65 @@
 
 using namespace nntrainer;
 
+static std::unique_ptr<NeuralNetwork> split_axis3_split_number5() {
+  std::unique_ptr<NeuralNetwork> nn(new NeuralNetwork());
+  nn->setProperty({"batch_size=2"});
+
+  auto graph = makeGraph({
+    {"conv2d",
+     {"name=conv", "input_shape=3:4:5", "filters=3", "kernel_size=1,1"}},
+    {"split", {"name=split", "input_layers=conv", "axis=3", "split_number=5"}},
+    {"addition",
+     {"name=add", "input_layers=split(0),split(1),split(2),split(3),split(4)"}},
+    {"mse", {"name=loss", "input_layers=add"}},
+  });
+  for (auto &node : graph) {
+    nn->addLayer(node);
+  }
+
+  nn->setOptimizer(ml::train::createOptimizer("sgd", {"learning_rate = 0.1"}));
+  return nn;
+}
+
+static std::unique_ptr<NeuralNetwork> split_axis2_split_number4() {
+  std::unique_ptr<NeuralNetwork> nn(new NeuralNetwork());
+  nn->setProperty({"batch_size=2"});
+
+  auto graph = makeGraph({
+    {"conv2d",
+     {"name=conv", "input_shape=3:4:5", "filters=3", "kernel_size=1,1"}},
+    {"split", {"name=split", "input_layers=conv", "axis=2", "split_number=4"}},
+    {"addition",
+     {"name=add", "input_layers=split(0),split(1),split(2),split(3)"}},
+    {"mse", {"name=loss", "input_layers=add"}},
+  });
+  for (auto &node : graph) {
+    nn->addLayer(node);
+  }
+
+  nn->setOptimizer(ml::train::createOptimizer("sgd", {"learning_rate = 0.1"}));
+  return nn;
+}
+
+static std::unique_ptr<NeuralNetwork> split_axis2_split_number2() {
+  std::unique_ptr<NeuralNetwork> nn(new NeuralNetwork());
+  nn->setProperty({"batch_size=2"});
+
+  auto graph = makeGraph({
+    {"conv2d",
+     {"name=conv", "input_shape=3:4:5", "filters=3", "kernel_size=1,1"}},
+    {"split", {"name=split", "input_layers=conv", "axis=2", "split_number=2"}},
+    {"addition", {"name=add", "input_layers=split(0),split(1)"}},
+    {"mse", {"name=loss", "input_layers=add"}},
+  });
+  for (auto &node : graph) {
+    nn->addLayer(node);
+  }
+
+  nn->setOptimizer(ml::train::createOptimizer("sgd", {"learning_rate = 0.1"}));
+  return nn;
+}
+
 /// A has two output tensor a1, a2 and B, C takes it
 ///     A
 /// (a0, a1)
@@ -176,6 +235,12 @@ static std::unique_ptr<NeuralNetwork> split_and_join_dangle() {
 GTEST_PARAMETER_TEST(
   multiInoutModels, nntrainerModelTest,
   ::testing::ValuesIn({
+    mkModelTc_V2(split_axis3_split_number5, "split_axis3_split_number5",
+                 ModelTestOption::ALL_V2),
+    mkModelTc_V2(split_axis2_split_number4, "split_axis2_split_number4",
+                 ModelTestOption::ALL_V2),
+    mkModelTc_V2(split_axis2_split_number2, "split_axis2_split_number2",
+                 ModelTestOption::ALL_V2),
     mkModelTc_V2(split_and_join, "split_and_join", ModelTestOption::ALL_V2),
     mkModelTc_V2(one_to_one, "one_to_one", ModelTestOption::ALL_V2),
     mkModelTc_V2(one_to_one_reversed, "one_to_one__reversed",