data layer: allow scaling and subtraction

diff --git a/src/caffe/layers/data_layer.cpp b/src/caffe/layers/data_layer.cpp
index 5786a0b..442a359 100644 (file)
--- a/src/caffe/layers/data_layer.cpp
+++ b/src/caffe/layers/data_layer.cpp
@@ -48,12 +48,15 @@ void DataLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   Datum datum;
   Dtype* top_data = (*top)[0]->mutable_cpu_data();
   Dtype* top_label = (*top)[1]->mutable_cpu_data();
+  const Dtype scale = this->layer_param_.scale();
+  const Dtype subtraction = this->layer_param_.subtraction();
   for (int i = 0; i < this->layer_param_.batchsize(); ++i) {
     // get a blob
     datum.ParseFromString(iter_->value().ToString());
     const string& data = datum.data();
     for (int j = 0; j < datum_size_; ++j) {
-      top_data[i * datum_size_ + j] = (uint8_t)data[j];
+      top_data[i * datum_size_ + j] =
+          (static_cast<Dtype>((uint8_t)data[j]) * scale) - subtraction;
     }
     top_label[i] = datum.label();
     // go to the next iter

diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index 70c399c..012ad56 100644 (file)
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -58,8 +58,11 @@ message LayerParameter {
 
   // For data layers, specify the data source
   optional string source = 16;
+  // For data pre-processing, we can do simple scaling and constant subtraction
+  optional float scale = 17 [ default = 1 ];
+  optional float subtraction = 18 [ default = 0 ];
   // For datay layers, specify the batch size.
-  optional uint32 batchsize = 17;
+  optional uint32 batchsize = 19;
 
   // The blobs containing the numeric parameters of the layer
   repeated BlobProto blobs = 50;
@@ -90,4 +93,4 @@ message SolverParameter {
   optional float momentum = 10; // The momentum value.
 
   optional string snapshot_prefix = 11; // The prefix for the snapshot.
-}
\ No newline at end of file
+}

diff --git a/src/caffe/test/data/lenet.prototxt b/src/caffe/test/data/lenet.prototxt
index 8bb3ff7..ac5c1ca 100644 (file)
--- a/src/caffe/test/data/lenet.prototxt
+++ b/src/caffe/test/data/lenet.prototxt
@@ -1,6 +1,15 @@
 name: "LeNet"
-input: "data"
-input: "label"
+layers {
+  layer {
+    name: "mnist"
+    type: "data"
+    source: "caffe/test/data/mnist-train-leveldb"
+    batchsize: 32
+    scale: 0.00390625
+  }
+  top: "data"
+  top: "label"
+}
 layers {
   layer {
     name: "conv1"

diff --git a/src/caffe/test/lenet.hpp b/src/caffe/test/lenet.hpp
deleted file mode 100644 (file)
index 8b5b285..0000000
--- a/src/caffe/test/lenet.hpp
+++ /dev/null
@@ -1,130 +0,0 @@
-// Copyright Yangqing Jia 2013
-
-// This file is merely here so we can easily get a string of the lenet.
-// It is actually not the very original LeNet, but with the sigmoid layers
-// replaced by ReLU layers.
-
-#ifndef CAFFE_TEST_LENET_HPP_
-#define CAFFE_TEST_LENET_HPP_
-
-#include <string>
-
-namespace caffe {
-
-const char* kLENET = "name: \"LeNet\"\n"
-"input: \"data\"\n"
-"input: \"label\"\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"conv1\"\n"
-"    type: \"conv\"\n"
-"    num_output: 20\n"
-"    kernelsize: 5\n"
-"    stride: 1\n"
-"    weight_filler {\n"
-"      type: \"xavier\"\n"
-"    }\n"
-"    bias_filler {\n"
-"      type: \"constant\"\n"
-"    }\n"
-"  }\n"
-"  bottom: \"data\"\n"
-"  top: \"conv1\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"pool1\"\n"
-"    type: \"pool\"\n"
-"    kernelsize: 2\n"
-"    stride: 2\n"
-"    pool: MAX\n"
-"  }\n"
-"  bottom: \"conv1\"\n"
-"  top: \"pool1\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"conv2\"\n"
-"    type: \"conv\"\n"
-"    num_output: 50\n"
-"    kernelsize: 5\n"
-"    stride: 1\n"
-"    weight_filler {\n"
-"      type: \"xavier\"\n"
-"    }\n"
-"    bias_filler {\n"
-"      type: \"constant\"\n"
-"    }\n"
-"  }\n"
-"  bottom: \"pool1\"\n"
-"  top: \"conv2\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"pool2\"\n"
-"    type: \"pool\"\n"
-"    kernelsize: 2\n"
-"    stride: 2\n"
-"    pool: MAX\n"
-"  }\n"
-"  bottom: \"conv2\"\n"
-"  top: \"pool2\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"ip1\"\n"
-"    type: \"innerproduct\"\n"
-"    num_output: 500\n"
-"    weight_filler {\n"
-"      type: \"xavier\"\n"
-"    }\n"
-"    bias_filler {\n"
-"      type: \"constant\"\n"
-"    }\n"
-"  }\n"
-"  bottom: \"pool2\"\n"
-"  top: \"ip1\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"relu1\"\n"
-"    type: \"relu\"\n"
-"  }\n"
-"  bottom: \"ip1\"\n"
-"  top: \"relu1\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"ip2\"\n"
-"    type: \"innerproduct\"\n"
-"    num_output: 10\n"
-"    weight_filler {\n"
-"      type: \"xavier\"\n"
-"    }\n"
-"    bias_filler {\n"
-"      type: \"constant\"\n"
-"    }\n"
-"  }\n"
-"  bottom: \"relu1\"\n"
-"  top: \"ip2\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"prob\"\n"
-"    type: \"softmax\"\n"
-"  }\n"
-"  bottom: \"ip2\"\n"
-"  top: \"prob\"\n"
-"}\n"
-"layers {\n"
-"  layer {\n"
-"    name: \"loss\"\n"
-"    type: \"multinomial_logistic_loss\"\n"
-"  }\n"
-"  bottom: \"prob\"\n"
-"  bottom: \"label\"\n"
-"}";
-
-}  // namespace caffe
-
-#endif

diff --git a/src/caffe/test/test_net_proto.cpp b/src/caffe/test/test_net_proto.cpp
index cd16ed6..8cfbd77 100644 (file)
--- a/src/caffe/test/test_net_proto.cpp
+++ b/src/caffe/test/test_net_proto.cpp
@@ -15,7 +15,6 @@
 #include "caffe/proto/caffe.pb.h"
 #include "caffe/util/io.hpp"
 
-#include "caffe/test/lenet.hpp"
 #include "caffe/test/test_caffe_main.hpp"
 
 namespace caffe {
@@ -26,37 +25,17 @@ class NetProtoTest : public ::testing::Test {};
 typedef ::testing::Types<float, double> Dtypes;
 TYPED_TEST_CASE(NetProtoTest, Dtypes);
 
-TYPED_TEST(NetProtoTest, TestLoadFromText) {
-  NetParameter net_param;
-  ReadProtoFromTextFile("caffe/test/data/lenet.prototxt", &net_param);
-}
-
 TYPED_TEST(NetProtoTest, TestSetup) {
   NetParameter net_param;
-  string lenet_string(kLENET);
-  // Load the network
-  CHECK(google::protobuf::TextFormat::ParseFromString(
-      lenet_string, &net_param));
+  ReadProtoFromTextFile("caffe/test/data/lenet.prototxt", &net_param);
   // check if things are right
-  EXPECT_EQ(net_param.layers_size(), 9);
-  EXPECT_EQ(net_param.input_size(), 2);
-
-  // Now, initialize a network using the parameter
-  shared_ptr<Blob<TypeParam> > data(new Blob<TypeParam>(10, 1, 28, 28));
-  shared_ptr<Blob<TypeParam> > label(new Blob<TypeParam>(10, 1, 1, 1));
-  FillerParameter filler_param;
-  shared_ptr<Filler<TypeParam> > filler;
-  filler.reset(new ConstantFiller<TypeParam>(filler_param));
-  filler->Fill(label.get());
-  filler.reset(new UniformFiller<TypeParam>(filler_param));
-  filler->Fill(data.get());
+  EXPECT_EQ(net_param.layers_size(), 10);
+  EXPECT_EQ(net_param.input_size(), 0);
 
   vector<Blob<TypeParam>*> bottom_vec;
-  bottom_vec.push_back(data.get());
-  bottom_vec.push_back(label.get());
 
   Net<TypeParam> caffe_net(net_param, bottom_vec);
-  EXPECT_EQ(caffe_net.layer_names().size(), 9);
+  EXPECT_EQ(caffe_net.layer_names().size(), 10);
   EXPECT_EQ(caffe_net.blob_names().size(), 10);
 
   // Print a few statistics to see if things are correct
@@ -67,8 +46,15 @@ TYPED_TEST(NetProtoTest, TestSetup) {
         << caffe_net.blobs()[i]->height() << ", "
         << caffe_net.blobs()[i]->width();
   }
+  Caffe::set_mode(Caffe::CPU);
+  // Run the network without training.
+  LOG(ERROR) << "Performing Forward";
+  caffe_net.Forward(bottom_vec);
+  LOG(ERROR) << "Performing Backward";
+  LOG(ERROR) << caffe_net.Backward();
+  
+  Caffe::set_mode(Caffe::GPU);
   // Run the network without training.
-  vector<Blob<TypeParam>*> top_vec;
   LOG(ERROR) << "Performing Forward";
   caffe_net.Forward(bottom_vec);
   LOG(ERROR) << "Performing Backward";