stochastic pooling test

diff --git a/src/caffe/test/test_gradient_check_util.hpp b/src/caffe/test/test_gradient_check_util.hpp
index 4973483..d736008 100644 (file)
--- a/src/caffe/test/test_gradient_check_util.hpp
+++ b/src/caffe/test/test_gradient_check_util.hpp
@@ -28,9 +28,7 @@ class GradientChecker {
       : stepsize_(stepsize), threshold_(threshold), seed_(seed),
         kink_(kink), kink_range_(kink_range) {}
   // Checks the gradient of a layer, with provided bottom layers and top
-  // layers. The gradient checker will check the gradient with respect to
-  // the parameters of the layer, as well as the input blobs if check_through
-  // is set True.
+  // layers.
   // Note that after the gradient check, we do not guarantee that the data
   // stored in the layer parameters and the blobs are unchanged.
   void CheckGradient(Layer<Dtype>& layer, vector<Blob<Dtype>*>& bottom,

diff --git a/src/caffe/test/test_pooling_layer.cpp b/src/caffe/test/test_pooling_layer.cpp
index a5d0c9f..67cae13 100644 (file)
--- a/src/caffe/test/test_pooling_layer.cpp
+++ b/src/caffe/test/test_pooling_layer.cpp
@@ -33,8 +33,6 @@ class PoolingLayerTest : public ::testing::Test {
     blob_top_vec_.push_back(blob_top_);
   };
   virtual ~PoolingLayerTest() { delete blob_bottom_; delete blob_top_; }
-  void ReferenceLRNForward(const Blob<Dtype>& blob_bottom,
-      const LayerParameter& layer_param, Blob<Dtype>* blob_top);
   Blob<Dtype>* const blob_bottom_;
   Blob<Dtype>* const blob_top_;
   vector<Blob<Dtype>*> blob_bottom_vec_;
@@ -71,7 +69,7 @@ TYPED_TEST(PoolingLayerTest, PrintGPUBackward) {
   }
   for (int i = 0; i < this->blob_top_->count(); ++i) {
     cout << "top data " << i << " " << this->blob_top_->cpu_data()[i] << endl;
-  }  
+  }
 
   for (int i = 0; i < this->blob_top_->count(); ++i) {
     this->blob_top_->mutable_cpu_diff()[i] = 1.;
@@ -79,7 +77,7 @@ TYPED_TEST(PoolingLayerTest, PrintGPUBackward) {
   layer.Backward(this->blob_top_vec_, true, &(this->blob_bottom_vec_));
   for (int i = 0; i < this->blob_bottom_->count(); ++i) {
     cout << "bottom diff " << i << " " << this->blob_bottom_->cpu_diff()[i] << endl;
-  }  
+  }
 }
 */
 

diff --git a/src/caffe/test/test_stochastic_pooing.cpp b/src/caffe/test/test_stochastic_pooing.cpp
new file mode 100644 (file)
index 0000000..e2b60ee
--- /dev/null
+++ b/src/caffe/test/test_stochastic_pooing.cpp
@@ -0,0 +1,162 @@
+// Copyright 2013 Yangqing Jia
+
+#include <cstring>
+#include <cuda_runtime.h>
+
+#include "gtest/gtest.h"
+#include "caffe/blob.hpp"
+#include "caffe/common.hpp"
+#include "caffe/filler.hpp"
+#include "caffe/vision_layers.hpp"
+#include "caffe/test/test_gradient_check_util.hpp"
+
+#include "caffe/test/test_caffe_main.hpp"
+
+namespace caffe {
+
+extern cudaDeviceProp CAFFE_TEST_CUDA_PROP;
+
+template <typename Dtype>
+class StochasticPoolingLayerTest : public ::testing::Test {
+ protected:
+  StochasticPoolingLayerTest()
+      : blob_bottom_(new Blob<Dtype>()),
+        blob_top_(new Blob<Dtype>()) {};
+  virtual void SetUp() {
+    Caffe::set_random_seed(1701);
+    blob_bottom_->Reshape(2, 3, 6, 5);
+    // fill the values
+    FillerParameter filler_param;
+    filler_param.set_min(0.1);
+    filler_param.set_max(1.);
+    UniformFiller<Dtype> filler(filler_param);
+    filler.Fill(this->blob_bottom_);
+    blob_bottom_vec_.push_back(blob_bottom_);
+    blob_top_vec_.push_back(blob_top_);
+  };
+
+  virtual ~StochasticPoolingLayerTest() {
+    delete blob_bottom_; delete blob_top_;
+  }
+
+  Blob<Dtype>* const blob_bottom_;
+  Blob<Dtype>* const blob_top_;
+  vector<Blob<Dtype>*> blob_bottom_vec_;
+  vector<Blob<Dtype>*> blob_top_vec_;
+};
+
+typedef ::testing::Types<float, double> Dtypes;
+TYPED_TEST_CASE(StochasticPoolingLayerTest, Dtypes);
+
+TYPED_TEST(StochasticPoolingLayerTest, TestSetup) {
+  LayerParameter layer_param;
+  layer_param.set_kernelsize(3);
+  layer_param.set_stride(2);
+  PoolingLayer<TypeParam> layer(layer_param);
+  layer.SetUp(this->blob_bottom_vec_, &(this->blob_top_vec_));
+  EXPECT_EQ(this->blob_top_->num(), this->blob_bottom_->num());
+  EXPECT_EQ(this->blob_top_->channels(), this->blob_bottom_->channels());
+  EXPECT_EQ(this->blob_top_->height(), 3);
+  EXPECT_EQ(this->blob_top_->width(), 2);
+}
+
+TYPED_TEST(StochasticPoolingLayerTest, TestStochasticGPU) {
+  Caffe::set_mode(Caffe::GPU);
+  Caffe::set_phase(Caffe::TRAIN);
+  LayerParameter layer_param;
+  layer_param.set_kernelsize(3);
+  layer_param.set_stride(2);
+
+  layer_param.set_pool(LayerParameter_PoolMethod_STOCHASTIC);
+  PoolingLayer<TypeParam> layer(layer_param);
+  layer.SetUp(this->blob_bottom_vec_, &(this->blob_top_vec_));
+  layer.Forward(this->blob_bottom_vec_, &(this->blob_top_vec_));
+
+  // Check if the output is correct - it should do random sampling
+  const TypeParam* bottom_data = this->blob_bottom_->cpu_data();
+  const TypeParam* top_data = this->blob_top_->cpu_data();
+  TypeParam total = 0;
+  for (int n = 0; n < this->blob_top_->num(); ++n) {
+    for (int c = 0; c < this->blob_top_->channels(); ++c) {
+      for (int ph = 0; ph < this->blob_top_->height(); ++ph) {
+        for (int pw = 0; pw < this->blob_top_->width(); ++pw) {
+          TypeParam pooled = top_data[this->blob_top_->offset(n, c, ph, pw)];
+          total += pooled;
+          int hstart = ph * 2;
+          int hend = min(hstart + 3, this->blob_bottom_->height());
+          int wstart = pw * 2;
+          int wend = min(wstart + 3, this->blob_bottom_->width());
+          bool has_equal = false;
+          for (int h = hstart; h < hend; ++h) {
+            for (int w = wstart; w < wend; ++w) {
+              has_equal |= (pooled == bottom_data[this->blob_bottom_->offset(n, c, h, w)]);
+            }
+          }
+          EXPECT_TRUE(has_equal);
+        }
+      }
+    }
+  }
+  // When we are doing stochastic pooling, the average we get should be higher
+  // than the simple data average since we are weighting more on higher-valued
+  // ones.
+  EXPECT_GE(total / this->blob_top_->count(), 0.55);
+}
+
+TYPED_TEST(StochasticPoolingLayerTest, TestStochasticGPUTestPhase) {
+  Caffe::set_mode(Caffe::GPU);
+  Caffe::set_phase(Caffe::TEST);
+  LayerParameter layer_param;
+  layer_param.set_kernelsize(3);
+  layer_param.set_stride(2);
+
+  layer_param.set_pool(LayerParameter_PoolMethod_STOCHASTIC);
+  PoolingLayer<TypeParam> layer(layer_param);
+  layer.SetUp(this->blob_bottom_vec_, &(this->blob_top_vec_));
+  layer.Forward(this->blob_bottom_vec_, &(this->blob_top_vec_));
+
+  // Check if the output is correct - it should do random sampling
+  const TypeParam* bottom_data = this->blob_bottom_->cpu_data();
+  const TypeParam* top_data = this->blob_top_->cpu_data();
+  for (int n = 0; n < this->blob_top_->num(); ++n) {
+    for (int c = 0; c < this->blob_top_->channels(); ++c) {
+      for (int ph = 0; ph < this->blob_top_->height(); ++ph) {
+        for (int pw = 0; pw < this->blob_top_->width(); ++pw) {
+          TypeParam pooled = top_data[this->blob_top_->offset(n, c, ph, pw)];
+          int hstart = ph * 2;
+          int hend = min(hstart + 3, this->blob_bottom_->height());
+          int wstart = pw * 2;
+          int wend = min(wstart + 3, this->blob_bottom_->width());
+          bool smaller_than_max = false;
+          for (int h = hstart; h < hend; ++h) {
+            for (int w = wstart; w < wend; ++w) {
+              smaller_than_max |= (pooled <= bottom_data[this->blob_bottom_->offset(n, c, h, w)]);
+            }
+          }
+          EXPECT_TRUE(smaller_than_max);
+        }
+      }
+    }
+  }
+}
+
+
+
+TYPED_TEST(StochasticPoolingLayerTest, TestGradientGPU) {
+  Caffe::set_mode(Caffe::GPU);
+  Caffe::set_phase(Caffe::TRAIN);
+  LayerParameter layer_param;
+  layer_param.set_kernelsize(3);
+  layer_param.set_stride(2);
+
+  layer_param.set_pool(LayerParameter_PoolMethod_STOCHASTIC);
+  PoolingLayer<TypeParam> layer(layer_param);
+  GradientChecker<TypeParam> checker(1e-2, 1e-3);
+  // it is too expensive to call curand multiple times, so we don't do an
+  // exhaustive gradient check.
+  checker.CheckGradient(layer, this->blob_bottom_vec_, this->blob_top_vec_);
+}
+
+
+
+}