From e7f543bb264ad8597a3edaf3b938e9c3cc57bf33 Mon Sep 17 00:00:00 2001
From: ih4cku <ih4cku@gmail.com>
Date: Wed, 17 Jun 2015 12:15:28 +0800
Subject: [PATCH 01/75] register a dummy reducer to prevent mincepie runtime
 error

---
 tools/extra/resize_and_crop_images.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tools/extra/resize_and_crop_images.py b/tools/extra/resize_and_crop_images.py
index c844f590c06..fd2c3134edb 100755
--- a/tools/extra/resize_and_crop_images.py
+++ b/tools/extra/resize_and_crop_images.py
@@ -101,7 +101,7 @@ def map(self, key, value):
         yield value, FLAGS.output_folder
 
 mapreducer.REGISTER_DEFAULT_MAPPER(ResizeCropImagesMapper)
-
+mapreducer.REGISTER_DEFAULT_REDUCER(mapreducer.NoPassReducer)
 mapreducer.REGISTER_DEFAULT_READER(mapreducer.FileReader)
 mapreducer.REGISTER_DEFAULT_WRITER(mapreducer.FileWriter)
  

From 10725393518df14b9b6976686f72fae792c3f393 Mon Sep 17 00:00:00 2001
From: Jeff Donahue <jeff.donahue@gmail.com>
Date: Mon, 5 Oct 2015 15:46:54 -0700
Subject: [PATCH 02/75] NetSpec: type-check Function inputs (they must be Top
 instances)

---
 python/caffe/net_spec.py           | 4 ++++
 python/caffe/test/test_net_spec.py | 8 ++++++++
 2 files changed, 12 insertions(+)

diff --git a/python/caffe/net_spec.py b/python/caffe/net_spec.py
index 93fc01927db..b6520627a4b 100644
--- a/python/caffe/net_spec.py
+++ b/python/caffe/net_spec.py
@@ -103,6 +103,10 @@ class Function(object):
 
     def __init__(self, type_name, inputs, params):
         self.type_name = type_name
+        for index, input in enumerate(inputs):
+            if not isinstance(input, Top):
+                raise TypeError('%s input %d is not a Top (type is %s)' %
+                                (type_name, index, type(input)))
         self.inputs = inputs
         self.params = params
         self.ntop = self.params.get('ntop', 1)
diff --git a/python/caffe/test/test_net_spec.py b/python/caffe/test/test_net_spec.py
index fee3c0aaebe..ffe71bacb08 100644
--- a/python/caffe/test/test_net_spec.py
+++ b/python/caffe/test/test_net_spec.py
@@ -79,3 +79,11 @@ def test_zero_tops(self):
         net_proto = silent_net()
         net = self.load_net(net_proto)
         self.assertEqual(len(net.forward()), 0)
+
+    def test_type_error(self):
+        """Test that a TypeError is raised when a Function input isn't a Top."""
+        data = L.DummyData(ntop=2)  # data is a 2-tuple of Tops
+        r = r"^Silence input 0 is not a Top \(type is <(type|class) 'tuple'>\)$"
+        with self.assertRaisesRegexp(TypeError, r):
+            L.Silence(data, ntop=0)  # should raise: data is a tuple, not a Top
+        L.Silence(*data, ntop=0)  # shouldn't raise: each elt of data is a Top

From 52dcf4801dddf05df3ddef238895cabbc6c4384a Mon Sep 17 00:00:00 2001
From: Azat <davletag@mail.ru>
Date: Thu, 3 Dec 2015 13:56:48 +0300
Subject: [PATCH 03/75] sigmoid fix (cu)

Previous implementation caused FP overflow for x less than -90
---
 src/caffe/layers/sigmoid_layer.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/caffe/layers/sigmoid_layer.cu b/src/caffe/layers/sigmoid_layer.cu
index 184c61ede83..8a4ea6616e0 100644
--- a/src/caffe/layers/sigmoid_layer.cu
+++ b/src/caffe/layers/sigmoid_layer.cu
@@ -8,7 +8,7 @@ namespace caffe {
 template <typename Dtype>
 __global__ void SigmoidForward(const int n, const Dtype* in, Dtype* out) {
   CUDA_KERNEL_LOOP(index, n) {
-    out[index] = 1. / (1. + exp(-in[index]));
+    out[index] = 0.5 * tanh(0.5 * in[index]) + 0.5;
   }
 }
 

From 0f61cc09467afa35835dc09617f1042e4f77c9fb Mon Sep 17 00:00:00 2001
From: Azat <davletag@mail.ru>
Date: Thu, 3 Dec 2015 14:00:08 +0300
Subject: [PATCH 04/75] sigmoid fix (cpp)

Previous implementation caused FP overflow for x less than -90
---
 src/caffe/layers/sigmoid_layer.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/caffe/layers/sigmoid_layer.cpp b/src/caffe/layers/sigmoid_layer.cpp
index 85fd9676812..f8aa769a174 100644
--- a/src/caffe/layers/sigmoid_layer.cpp
+++ b/src/caffe/layers/sigmoid_layer.cpp
@@ -7,7 +7,7 @@ namespace caffe {
 
 template <typename Dtype>
 inline Dtype sigmoid(Dtype x) {
-  return 1. / (1. + exp(-x));
+  return 0.5 * tanh(0.5 * x) + 0.5;
 }
 
 template <typename Dtype>

From 337b07589f4e44761bdb9ef4c242f83ca40c9da5 Mon Sep 17 00:00:00 2001
From: shai <shai@magisto.com>
Date: Mon, 21 Mar 2016 09:08:02 +0200
Subject: [PATCH 05/75] upgrading InfogainLoss layer: (1) incorporating Softmax
 layer to make the gradeint computation robust, much like SoftmaxWithLoss
 layer (see: http://stackoverflow.com/a/34917052/1714410 for more
 information). (2) supporting loss along axis

---
 include/caffe/layers/infogain_loss_layer.hpp |  35 ++++++
 src/caffe/layers/infogain_loss_layer.cpp     | 172 ++++++++++++++++++++++-----
 src/caffe/proto/caffe.proto                  |   1 +
 src/caffe/test/test_infogain_loss_layer.cpp  |  83 ++++++++++++-
 4 files changed, 257 insertions(+), 34 deletions(-)

diff --git a/include/caffe/layers/infogain_loss_layer.hpp b/include/caffe/layers/infogain_loss_layer.hpp
index 633f339a28e..edecde829ad 100644
--- a/include/caffe/layers/infogain_loss_layer.hpp
+++ b/include/caffe/layers/infogain_loss_layer.hpp
@@ -8,6 +8,7 @@
 #include "caffe/proto/caffe.pb.h"
 
 #include "caffe/layers/loss_layer.hpp"
+#include "caffe/layers/softmax_layer.hpp"
 
 namespace caffe {
 
@@ -60,6 +61,12 @@ class InfogainLossLayer : public LossLayer<Dtype> {
   virtual inline int MinBottomBlobs() const { return 2; }
   virtual inline int MaxBottomBlobs() const { return 3; }
 
+  // InfogainLossLayer computes softmax prob internally.
+  // optional second "top" outputs the softmax prob
+  virtual inline int ExactNumTopBlobs() const { return -1; }
+  virtual inline int MinTopBlobs() const { return 1; }
+  virtual inline int MaxTopBlobs() const { return 2; }
+
   virtual inline const char* type() const { return "InfogainLoss"; }
 
  protected:
@@ -102,7 +109,35 @@ class InfogainLossLayer : public LossLayer<Dtype> {
   virtual void Backward_cpu(const vector<Blob<Dtype>*>& top,
       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
 
+  /// Read the normalization mode parameter and compute the normalizer based
+  /// on the blob size.  If normalization_mode is VALID, the count of valid
+  /// outputs will be read from valid_count, unless it is -1 in which case
+  /// all outputs are assumed to be valid.
+  virtual Dtype get_normalizer(
+      LossParameter_NormalizationMode normalization_mode, int valid_count);
+  /// fill sum_rows_H_ according to matrix H
+  virtual void sum_rows_of_H(const Blob<Dtype>* H);
+
+  /// The internal SoftmaxLayer used to map predictions to a distribution.
+  shared_ptr<Layer<Dtype> > softmax_layer_;
+  /// prob stores the output probability predictions from the SoftmaxLayer.
+  Blob<Dtype> prob_;
+  /// bottom vector holder used in call to the underlying SoftmaxLayer::Forward
+  vector<Blob<Dtype>*> softmax_bottom_vec_;
+  /// top vector holder used in call to the underlying SoftmaxLayer::Forward
+  vector<Blob<Dtype>*> softmax_top_vec_;
+
   Blob<Dtype> infogain_;
+  Blob<Dtype> sum_rows_H_;  // cache the row sums of H.
+
+  /// Whether to ignore instances with a certain label.
+  bool has_ignore_label_;
+  /// The label indicating that an instance should be ignored.
+  int ignore_label_;
+  /// How to normalize the output loss.
+  LossParameter_NormalizationMode normalization_;
+
+  int infogain_axis_, outer_num_, inner_num_, num_labels_;
 };
 
 }  // namespace caffe
diff --git a/src/caffe/layers/infogain_loss_layer.cpp b/src/caffe/layers/infogain_loss_layer.cpp
index 624d3118124..3c3f460ec34 100644
--- a/src/caffe/layers/infogain_loss_layer.cpp
+++ b/src/caffe/layers/infogain_loss_layer.cpp
@@ -3,7 +3,8 @@
 #include <vector>
 
 #include "caffe/layers/infogain_loss_layer.hpp"
-#include "caffe/util/io.hpp"
+#include "caffe/util/io.hpp"  // for bolb reading of matrix H
+#include "caffe/util/math_functions.hpp"
 
 namespace caffe {
 
@@ -11,6 +12,31 @@ template <typename Dtype>
 void InfogainLossLayer<Dtype>::LayerSetUp(
     const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
   LossLayer<Dtype>::LayerSetUp(bottom, top);
+  // internal softmax layer
+  LayerParameter softmax_layer_param(this->layer_param_);
+  SoftmaxParameter* softmax_param = softmax_layer_param.mutable_softmax_param();
+  softmax_param->set_axis(this->layer_param_.infogain_loss_param().axis());
+  softmax_layer_param.set_type("Softmax");
+  softmax_layer_param.clear_loss_weight();
+  softmax_layer_param.add_loss_weight(1);
+  softmax_layer_ = LayerRegistry<Dtype>::CreateLayer(softmax_layer_param);
+  softmax_bottom_vec_.clear();
+  softmax_bottom_vec_.push_back(bottom[0]);
+  softmax_top_vec_.clear();
+  softmax_top_vec_.push_back(&prob_);
+  softmax_layer_->SetUp(softmax_bottom_vec_, softmax_top_vec_);
+
+  // ignore label
+  has_ignore_label_ =
+    this->layer_param_.loss_param().has_ignore_label();
+  if (has_ignore_label_) {
+    ignore_label_ = this->layer_param_.loss_param().ignore_label();
+  }
+  // normalization
+  CHECK(!this->layer_param_.loss_param().has_normalize())
+    << "normalize is deprecated. use \"normalization\"";
+  normalization_ = this->layer_param_.loss_param().normalization();
+  // matrix H
   if (bottom.size() < 3) {
     CHECK(this->layer_param_.infogain_loss_param().has_source())
         << "Infogain matrix source must be specified.";
@@ -25,28 +51,86 @@ template <typename Dtype>
 void InfogainLossLayer<Dtype>::Reshape(
     const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
   LossLayer<Dtype>::Reshape(bottom, top);
+  softmax_layer_->Reshape(softmax_bottom_vec_, softmax_top_vec_);
+  infogain_axis_ =
+    bottom[0]->CanonicalAxisIndex(
+      this->layer_param_.infogain_loss_param().axis());
+  outer_num_ = bottom[0]->count(0, infogain_axis_);
+  inner_num_ = bottom[0]->count(infogain_axis_ + 1);
+  CHECK_EQ(outer_num_ * inner_num_, bottom[1]->count())
+      << "Number of labels must match number of predictions; "
+      << "e.g., if infogain axis == 1 and prediction shape is (N, C, H, W), "
+      << "label count (number of labels) must be N*H*W, "
+      << "with integer values in {0, 1, ..., C-1}.";
+  num_labels_ = bottom[0]->shape(infogain_axis_);
   Blob<Dtype>* infogain = NULL;
   if (bottom.size() < 3) {
     infogain = &infogain_;
   } else {
     infogain = bottom[2];
   }
-  CHECK_EQ(bottom[1]->channels(), 1);
-  CHECK_EQ(bottom[1]->height(), 1);
-  CHECK_EQ(bottom[1]->width(), 1);
-  const int num = bottom[0]->num();
-  const int dim = bottom[0]->count() / num;
-  CHECK_EQ(infogain->num(), 1);
-  CHECK_EQ(infogain->channels(), 1);
-  CHECK_EQ(infogain->height(), dim);
-  CHECK_EQ(infogain->width(), dim);
+  CHECK_EQ(infogain->count(), num_labels_*num_labels_);
+  sum_rows_H_.Reshape(vector<int>(1, num_labels_));
+  if (bottom.size() == 2) {
+    // H is provided as a parameter and will not change. sum rows once
+    sum_rows_of_H(infogain);
+  }
+  if (top.size() >= 2) {
+    // softmax output
+    top[1]->ReshapeLike(*bottom[0]);
+  }
+}
+
+template <typename Dtype>
+Dtype InfogainLossLayer<Dtype>::get_normalizer(
+    LossParameter_NormalizationMode normalization_mode, int valid_count) {
+  Dtype normalizer;
+  switch (normalization_mode) {
+    case LossParameter_NormalizationMode_FULL:
+      normalizer = Dtype(outer_num_ * inner_num_);
+      break;
+    case LossParameter_NormalizationMode_VALID:
+      if (valid_count == -1) {
+        normalizer = Dtype(outer_num_ * inner_num_);
+      } else {
+        normalizer = Dtype(valid_count);
+      }
+      break;
+    case LossParameter_NormalizationMode_BATCH_SIZE:
+      normalizer = Dtype(outer_num_);
+      break;
+    case LossParameter_NormalizationMode_NONE:
+      normalizer = Dtype(1);
+      break;
+    default:
+      LOG(FATAL) << "Unknown normalization mode: "
+          << LossParameter_NormalizationMode_Name(normalization_mode);
+  }
+  // Some users will have no labels for some examples in order to 'turn off' a
+  // particular loss in a multi-task setup. The max prevents NaNs in that case.
+  return std::max(Dtype(1.0), normalizer);
 }
 
+template <typename Dtype>
+void InfogainLossLayer<Dtype>::sum_rows_of_H(const Blob<Dtype>* H) {
+  CHECK_EQ(H->count(), num_labels_*num_labels_)
+    << "H must be " << num_labels_ << "x" << num_labels_;
+  const Dtype* infogain_mat = H->cpu_data();
+  Dtype* sum = sum_rows_H_.mutable_cpu_data();
+  for ( int row = 0; row < num_labels_ ; row++ ) {
+    sum[row] = 0;
+    for ( int col = 0; col < num_labels_ ; col++ ) {
+      sum[row] += infogain_mat[row*num_labels_+col];
+    }
+  }
+}
 
 template <typename Dtype>
 void InfogainLossLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
-  const Dtype* bottom_data = bottom[0]->cpu_data();
+  // The forward pass computes the softmax prob values.
+  softmax_layer_->Forward(softmax_bottom_vec_, softmax_top_vec_);
+  const Dtype* prob_data = prob_.cpu_data();
   const Dtype* bottom_label = bottom[1]->cpu_data();
   const Dtype* infogain_mat = NULL;
   if (bottom.size() < 3) {
@@ -54,17 +138,30 @@ void InfogainLossLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   } else {
     infogain_mat = bottom[2]->cpu_data();
   }
-  int num = bottom[0]->num();
-  int dim = bottom[0]->count() / bottom[0]->num();
+  int count = 0;
   Dtype loss = 0;
-  for (int i = 0; i < num; ++i) {
-    int label = static_cast<int>(bottom_label[i]);
-    for (int j = 0; j < dim; ++j) {
-      Dtype prob = std::max(bottom_data[i * dim + j], Dtype(kLOG_THRESHOLD));
-      loss -= infogain_mat[label * dim + j] * log(prob);
+  for (int i = 0; i < outer_num_; ++i) {
+    for (int j = 0; j < inner_num_; j++) {
+      const int label_value =
+        static_cast<int>(bottom_label[i * inner_num_ + j]);
+      if (has_ignore_label_ && label_value == ignore_label_) {
+        continue;
+      }
+      DCHECK_GE(label_value, 0);
+      DCHECK_LT(label_value, num_labels_);
+      for (int l = 0; l < num_labels_; l++) {
+        loss -= infogain_mat[label_value * num_labels_ + l] *
+          log(std::max(
+                prob_data[i * inner_num_*num_labels_ + l * inner_num_ + j],
+                Dtype(kLOG_THRESHOLD)));
+      }
+      ++count;
     }
   }
-  top[0]->mutable_cpu_data()[0] = loss / num;
+  top[0]->mutable_cpu_data()[0] = loss / get_normalizer(normalization_, count);
+  if (top.size() == 2) {
+    top[1]->ShareData(prob_);
+  }
 }
 
 template <typename Dtype>
@@ -80,25 +177,44 @@ void InfogainLossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
                << " Layer cannot backpropagate to infogain inputs.";
   }
   if (propagate_down[0]) {
-    const Dtype* bottom_data = bottom[0]->cpu_data();
+    const Dtype* prob_data = prob_.cpu_data();
     const Dtype* bottom_label = bottom[1]->cpu_data();
     const Dtype* infogain_mat = NULL;
     if (bottom.size() < 3) {
       infogain_mat = infogain_.cpu_data();
     } else {
       infogain_mat = bottom[2]->cpu_data();
+      // H is provided as a "bottom" and might change. sum rows every time.
+      sum_rows_of_H(bottom[2]);
     }
+    const Dtype* sum_rows_H = sum_rows_H_.cpu_data();
     Dtype* bottom_diff = bottom[0]->mutable_cpu_diff();
-    int num = bottom[0]->num();
-    int dim = bottom[0]->count() / bottom[0]->num();
-    const Dtype scale = - top[0]->cpu_diff()[0] / num;
-    for (int i = 0; i < num; ++i) {
-      const int label = static_cast<int>(bottom_label[i]);
-      for (int j = 0; j < dim; ++j) {
-        Dtype prob = std::max(bottom_data[i * dim + j], Dtype(kLOG_THRESHOLD));
-        bottom_diff[i * dim + j] = scale * infogain_mat[label * dim + j] / prob;
+    const int dim = bottom[0]->count() / outer_num_;
+    int count = 0;
+    for (int i = 0; i < outer_num_; ++i) {
+      for (int j = 0; j < inner_num_; ++j) {
+        const int label_value =
+          static_cast<int>(bottom_label[i * inner_num_ + j]);
+        DCHECK_GE(label_value, 0);
+        DCHECK_LT(label_value, num_labels_);
+        if (has_ignore_label_ && label_value == ignore_label_) {
+          for (int l = 0; l < num_labels_; ++l) {
+            bottom_diff[i * dim + l * inner_num_ + j] = 0;
+          }
+        } else {
+          for (int l = 0; l < num_labels_; ++l) {
+            bottom_diff[i * dim + l * inner_num_ + j] =
+               prob_data[i*dim + l*inner_num_ + j]*sum_rows_H[label_value]
+               - infogain_mat[label_value * num_labels_ + l];
+          }
+          ++count;
+        }
       }
     }
+    // Scale gradient
+    Dtype loss_weight = top[0]->cpu_diff()[0] /
+                        get_normalizer(normalization_, count);
+    caffe_scal(bottom[0]->count(), loss_weight, bottom_diff);
   }
 }
 
diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index 6900bb71482..591e9647258 100644
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -794,6 +794,7 @@ message ImageDataParameter {
 message InfogainLossParameter {
   // Specify the infogain matrix source.
   optional string source = 1;
+  optional int32 axis = 2 [default = 1]; // axis of prob
 }
 
 message InnerProductParameter {
diff --git a/src/caffe/test/test_infogain_loss_layer.cpp b/src/caffe/test/test_infogain_loss_layer.cpp
index a24ac683dc5..34f21271a62 100644
--- a/src/caffe/test/test_infogain_loss_layer.cpp
+++ b/src/caffe/test/test_infogain_loss_layer.cpp
@@ -1,3 +1,4 @@
+#include <algorithm>
 #include <vector>
 
 #include "gtest/gtest.h"
@@ -18,17 +19,22 @@ class InfogainLossLayerTest : public MultiDeviceTest<TypeParam> {
 
  protected:
   InfogainLossLayerTest()
-      : blob_bottom_data_(new Blob<Dtype>(10, 5, 1, 1)),
-        blob_bottom_label_(new Blob<Dtype>(10, 1, 1, 1)),
+      : blob_bottom_data_(new Blob<Dtype>(4, 2, 5, 2)),
+        blob_bottom_label_(new Blob<Dtype>(4, 2, 1, 2)),
         blob_bottom_infogain_(new Blob<Dtype>(1, 1, 5, 5)),
-        blob_top_loss_(new Blob<Dtype>()) {
+        blob_top_loss_(new Blob<Dtype>()),
+        blob_top_prob_(new Blob<Dtype>()),
+        inner_(2), outer_(4*2), num_labels_(5) {
     Caffe::set_random_seed(1701);
     FillerParameter filler_param;
-    PositiveUnitballFiller<Dtype> filler(filler_param);
+    filler_param.set_min(-0.5);
+    filler_param.set_max(2.0);
+    UniformFiller<Dtype> filler(filler_param);
     filler.Fill(this->blob_bottom_data_);
     blob_bottom_vec_.push_back(blob_bottom_data_);
     for (int i = 0; i < blob_bottom_label_->count(); ++i) {
-      blob_bottom_label_->mutable_cpu_data()[i] = caffe_rng_rand() % 5;
+      blob_bottom_label_->mutable_cpu_data()[i] =
+        caffe_rng_rand() % num_labels_;
     }
     blob_bottom_vec_.push_back(blob_bottom_label_);
     filler_param.set_min(0.1);
@@ -37,29 +43,94 @@ class InfogainLossLayerTest : public MultiDeviceTest<TypeParam> {
     infogain_filler.Fill(this->blob_bottom_infogain_);
     blob_bottom_vec_.push_back(blob_bottom_infogain_);
     blob_top_vec_.push_back(blob_top_loss_);
+    blob_top_vec_.push_back(blob_top_prob_);
   }
   virtual ~InfogainLossLayerTest() {
     delete blob_bottom_data_;
     delete blob_bottom_label_;
     delete blob_bottom_infogain_;
     delete blob_top_loss_;
+    delete blob_top_prob_;
   }
   Blob<Dtype>* const blob_bottom_data_;
   Blob<Dtype>* const blob_bottom_label_;
   Blob<Dtype>* const blob_bottom_infogain_;
   Blob<Dtype>* const blob_top_loss_;
+  Blob<Dtype>* const blob_top_prob_;
   vector<Blob<Dtype>*> blob_bottom_vec_;
   vector<Blob<Dtype>*> blob_top_vec_;
+  int inner_, outer_, num_labels_;
 };
 
 TYPED_TEST_CASE(InfogainLossLayerTest, TestDtypesAndDevices);
 
+TYPED_TEST(InfogainLossLayerTest, TestInfogainLoss) {
+  typedef typename TypeParam::Dtype Dtype;
+  LayerParameter layer_param;
+  layer_param.mutable_infogain_loss_param()->set_axis(2);
+  layer_param.clear_loss_weight();
+  layer_param.add_loss_weight(1);
+  layer_param.add_loss_weight(0);
+  /*vector<float>* lw = layer_param.mutable_loss_weight();
+  lw->clear();
+  lw->push_back(1);
+  lw->push_back(1);*/
+  InfogainLossLayer<Dtype> layer(layer_param);
+  layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
+  layer.Forward(this->blob_bottom_vec_, this->blob_top_vec_);
+  // Now, check values
+  const Dtype* data = this->blob_bottom_vec_[0]->cpu_data();
+  const Dtype* prob = this->blob_top_vec_[1]->cpu_data();
+  const Dtype* labels = this->blob_bottom_vec_[1]->cpu_data();
+  const Dtype* H = this->blob_bottom_vec_[2]->cpu_data();
+  // first. test the prob top
+  CHECK_EQ(this->blob_bottom_vec_[0]->num_axes(),
+    this->blob_top_vec_[1]->num_axes())
+      << "prob top shape not match bottom data";
+  for (int ai = 0 ; ai < this->blob_bottom_vec_[0]->num_axes(); ai++) {
+    CHECK_EQ(this->blob_bottom_vec_[0]->shape(ai),
+      this->blob_top_vec_[1]->shape(ai))
+        << "prob top shape not match bottom data";
+  }
+  vector<Dtype> est_prob(this->num_labels_, 0);
+  for ( int i = 0 ; i < this->outer_; i++ ) {
+    for ( int j = 0; j < this->inner_; j++ ) {
+      Dtype den = 0;
+      for ( int  l = 0; l < this->num_labels_; l++ ) {
+        est_prob[l] = std::exp(
+          data[i*this->num_labels_*this->inner_ + l*this->inner_ + j]);
+        den += est_prob[l];
+      }
+      for ( int l = 0; l < this->num_labels_; l++ ) {
+        EXPECT_NEAR(prob[i*this->num_labels_*this->inner_ + l*this->inner_ + j],
+          est_prob[l]/den, 1e-6);
+      }
+    }
+  }
+  Dtype loss = 0;  // loss from prob top
+  for ( int i = 0 ; i < this->outer_; i++ ) {
+    for ( int j = 0; j < this->inner_; j++ ) {
+      int gt = static_cast<int>(labels[i*this->inner_+j]);
+      for ( int l = 0; l < this->num_labels_; l++ ) {
+        loss -= H[gt*this->num_labels_ + l] *
+          log(std::max(
+            prob[i*this->num_labels_*this->inner_ + l*this->inner_ + j],
+            Dtype(kLOG_THRESHOLD)));
+      }
+    }
+  }
+  EXPECT_NEAR(this->blob_top_loss_->cpu_data()[0],
+    loss/(this->outer_*this->inner_), 1e-6);
+}
 
 TYPED_TEST(InfogainLossLayerTest, TestGradient) {
   typedef typename TypeParam::Dtype Dtype;
   LayerParameter layer_param;
+  layer_param.mutable_infogain_loss_param()->set_axis(2);
   InfogainLossLayer<Dtype> layer(layer_param);
-  GradientChecker<Dtype> checker(1e-4, 2e-2, 1701, 1, 0.01);
+  this->blob_top_vec_.clear();  // ignore prob top.
+  this->blob_top_vec_.push_back(this->blob_top_loss_);
+  GradientChecker<Dtype> checker(1e-4, 2e-2, 1701);  // no "kink"
   checker.CheckGradientExhaustive(&layer, this->blob_bottom_vec_,
       this->blob_top_vec_, 0);
 }

From a66bea30d6c0706f106b355c7cafc9e7ffae7bb5 Mon Sep 17 00:00:00 2001
From: An Tran <tranlaman@gmail.com>
Date: Wed, 30 Mar 2016 17:32:10 +0800
Subject: [PATCH 06/75] small bug in pooling_layer.cu

---
 src/caffe/layers/pooling_layer.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/caffe/layers/pooling_layer.cu b/src/caffe/layers/pooling_layer.cu
index 1ea46cc81b1..81ead1e8686 100644
--- a/src/caffe/layers/pooling_layer.cu
+++ b/src/caffe/layers/pooling_layer.cu
@@ -138,7 +138,7 @@ __global__ void StoPoolForwardTest(const int nthreads,
     const int wstart = pw * stride_w;
     const int wend = min(wstart + kernel_w, width);
     // We set cumsum to be 0 to avoid divide-by-zero problems
-    Dtype cumsum = FLT_MIN;
+    Dtype cumsum = 0.;
     Dtype cumvalues = 0.;
     const Dtype* const bottom_slice =
         bottom_data + (n * channels + c) * height * width;

From d17fbea6aad122c3818d5ef3593487869948b4b7 Mon Sep 17 00:00:00 2001
From: An Tran <tranlaman@gmail.com>
Date: Thu, 31 Mar 2016 10:27:31 +0800
Subject: [PATCH 07/75] avoid divide by zeros, suggested by SeanBell

---
 src/caffe/layers/pooling_layer.cu | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/caffe/layers/pooling_layer.cu b/src/caffe/layers/pooling_layer.cu
index 81ead1e8686..46eddb94924 100644
--- a/src/caffe/layers/pooling_layer.cu
+++ b/src/caffe/layers/pooling_layer.cu
@@ -149,7 +149,7 @@ __global__ void StoPoolForwardTest(const int nthreads,
         cumvalues += bottom_slice[h * width + w] * bottom_slice[h * width + w];
       }
     }
-    top_data[index] = cumvalues / cumsum;
+    top_data[index] = (cumsum > 0.) ? cumvalues / cumsum : 0.;
   }
 }
 

From d4e7c93a6873f75a53d7618e82343e4b5b8a239e Mon Sep 17 00:00:00 2001
From: Aaron Schumacher <ajschumacher@gmail.com>
Date: Thu, 19 May 2016 14:04:22 -0500
Subject: [PATCH 08/75] convert non-uint8 dtypes to float; refs #2391

As recommended by @longjon, this will allow `caffe.io.array_to_datum` to handle, for example, numpy.float32 arrays.

It might be worth noting that `datum.float_data` is stored as protobuf type 2, which is float32, as opposed to protobuf type 1, which is float64. It is a little unintuitive that caffe currently requires data to be passed in as float64 but then writes float32 to LMDB. To demonstrate this:

```python
datum = caffe.io.array_to_datum(np.array([[[0.9]]]))
caffe.io.datum_to_array(datum)
# array([[[ 0.9]]])
datum_str = datum.SerializeToString()
new_datum = caffe.proto.caffe_pb2.Datum()
new_datum.ParseFromString(datum_str)
caffe.io.datum_to_array(new_datum)
# array([[[ 0.89999998]]])
```

This behavior is somewhat hidden because `datum_to_array` returns type float64, even though the data doesn't actually have that resolution if it has been stored as protobuf text anywhere (for example in LMDB).

Alternative solutions:
 * Require and return float32, consistent with the protobuf representation.
 * Change the protobuf to allow float32 or float64 and update surrounding code to support this.
---
 python/caffe/io.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/caffe/io.py b/python/caffe/io.py
index e1759beb587..966c164cffd 100644
--- a/python/caffe/io.py
+++ b/python/caffe/io.py
@@ -75,7 +75,7 @@ def array_to_datum(arr, label=None):
     if arr.dtype == np.uint8:
         datum.data = arr.tostring()
     else:
-        datum.float_data.extend(arr.flat)
+        datum.float_data.extend(arr.astype(float).flat)
     if label is not None:
         datum.label = label
     return datum

From 5d7a71ae108f86c05bc03eb542155b30bd28ca74 Mon Sep 17 00:00:00 2001
From: Lumin Zhou <CDLuminate@users.noreply.github.com>
Date: Mon, 30 May 2016 04:19:16 +0000
Subject: [PATCH 09/75] using GNUInstallDirs in root cmake file

---
 CMakeLists.txt | 1 +
 1 file changed, 1 insertion(+)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index da7142c9b3c..c765889e99c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -18,6 +18,7 @@ add_definitions(-DCAFFE_VERSION=${CAFFE_TARGET_VERSION})
 list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake/Modules)
 
 include(ExternalProject)
+include(GNUInstallDirs)
 
 include(cmake/Utils.cmake)
 include(cmake/Targets.cmake)

From 90b98ce76fe8613d345932f47a6250dc772f7b8f Mon Sep 17 00:00:00 2001
From: Lumin Zhou <CDLuminate@users.noreply.github.com>
Date: Mon, 30 May 2016 04:21:27 +0000
Subject: [PATCH 10/75] fix install path with GNUInstallDir support

---
 src/caffe/CMakeLists.txt | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/caffe/CMakeLists.txt b/src/caffe/CMakeLists.txt
index 8a80c940488..5a1b73f7493 100644
--- a/src/caffe/CMakeLists.txt
+++ b/src/caffe/CMakeLists.txt
@@ -29,9 +29,9 @@ set_target_properties(caffe PROPERTIES
  add_subdirectory(test)
 
 # ---[ Install
-install(DIRECTORY ${Caffe_INCLUDE_DIR}/caffe DESTINATION include)
-install(FILES ${proto_hdrs} DESTINATION include/caffe/proto)
-install(TARGETS caffe proto EXPORT CaffeTargets DESTINATION lib)
+install(DIRECTORY ${Caffe_INCLUDE_DIR}/caffe DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
+install(FILES ${proto_hdrs} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/caffe/proto)
+install(TARGETS caffe proto EXPORT CaffeTargets DESTINATION ${CMAKE_INSTALL_LIBDIR})
 
 file(WRITE ${PROJECT_BINARY_DIR}/__init__.py)
 list(APPEND proto_python ${PROJECT_BINARY_DIR}/__init__.py)

From 581650b18d7580df726d1d6d54d83c397d1379bb Mon Sep 17 00:00:00 2001
From: Lumin Zhou <CDLuminate@users.noreply.github.com>
Date: Mon, 30 May 2016 04:22:42 +0000
Subject: [PATCH 11/75] fix install path with GNUInstallDir support

---
 tools/CMakeLists.txt | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/tools/CMakeLists.txt b/tools/CMakeLists.txt
index 02fbd5cadd8..3789450555e 100644
--- a/tools/CMakeLists.txt
+++ b/tools/CMakeLists.txt
@@ -25,5 +25,6 @@ foreach(source ${srcs})
   endif()
 
   # Install
-  install(TARGETS ${name} DESTINATION bin)
+  install(TARGETS ${name} DESTINATION ${CMAKE_INSTALL_BINDIR})
+
 endforeach(source)

From f710ef5e89d3ec22891b24099c66b7a6e9f06c45 Mon Sep 17 00:00:00 2001
From: Lumin Zhou <CDLuminate@users.noreply.github.com>
Date: Mon, 30 May 2016 04:24:13 +0000
Subject: [PATCH 12/75] fix install path with GNUInstallDir support

---
 examples/CMakeLists.txt | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 663d7360b7d..2a2300332ad 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -19,7 +19,8 @@ foreach(source_file ${examples_srcs})
   caffe_set_solution_folder(${name} examples)
 
   # install
-  install(TARGETS ${name} DESTINATION bin)
+  install(TARGETS ${name} DESTINATION ${CMAKE_INSTALL_BINDIR})
+
 
   if(UNIX OR APPLE)
     # Funny command to make tutorials work

From 9376bde1beba649e4c522b742064223ac9d2cab4 Mon Sep 17 00:00:00 2001
From: jasjuang <jasjuang@gmail.com>
Date: Thu, 21 Jul 2016 12:04:41 -0700
Subject: [PATCH 13/75] add in sudo make uninstall for cmake

---
 CMakeLists.txt           | 11 +++++++++++
 cmake/Uninstall.cmake.in | 26 ++++++++++++++++++++++++++
 2 files changed, 37 insertions(+)
 create mode 100644 cmake/Uninstall.cmake.in

diff --git a/CMakeLists.txt b/CMakeLists.txt
index da7142c9b3c..7b8dab2bb24 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -85,8 +85,19 @@ if(BUILD_python)
   add_dependencies(pytest pycaffe)
 endif()
 
+# ---[ uninstall target
+configure_file(
+    ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Uninstall.cmake.in
+    ${CMAKE_CURRENT_BINARY_DIR}/cmake/Uninstall.cmake
+    IMMEDIATE @ONLY)
+
+add_custom_target(uninstall
+    COMMAND ${CMAKE_COMMAND} -P
+    ${CMAKE_CURRENT_BINARY_DIR}/cmake/Uninstall.cmake)
+
 # ---[ Configuration summary
 caffe_print_configuration_summary()
 
 # ---[ Export configs generation
 caffe_generate_export_configs()
+
diff --git a/cmake/Uninstall.cmake.in b/cmake/Uninstall.cmake.in
new file mode 100644
index 00000000000..bb8e2964e46
--- /dev/null
+++ b/cmake/Uninstall.cmake.in
@@ -0,0 +1,26 @@
+if(NOT EXISTS "@CMAKE_CURRENT_BINARY_DIR@/install_manifest.txt")
+  message(FATAL_ERROR "Cannot find install manifest: @CMAKE_CURRENT_BINARY_DIR@/install_manifest.txt")
+endif(NOT EXISTS "@CMAKE_CURRENT_BINARY_DIR@/install_manifest.txt")
+
+if (NOT DEFINED CMAKE_INSTALL_PREFIX)
+  set (CMAKE_INSTALL_PREFIX "@CMAKE_INSTALL_PREFIX@")
+endif ()
+ message(${CMAKE_INSTALL_PREFIX})
+
+file(READ "@CMAKE_CURRENT_BINARY_DIR@/install_manifest.txt" files)
+string(REGEX REPLACE "\n" ";" files "${files}")
+foreach(file ${files})
+  message(STATUS "Uninstalling $ENV{DESTDIR}${file}")
+  if(IS_SYMLINK "$ENV{DESTDIR}${file}" OR EXISTS "$ENV{DESTDIR}${file}")
+    exec_program(
+      "@CMAKE_COMMAND@" ARGS "-E remove \"$ENV{DESTDIR}${file}\""
+      OUTPUT_VARIABLE rm_out
+      RETURN_VALUE rm_retval
+      )
+    if(NOT "${rm_retval}" STREQUAL 0)
+      message(FATAL_ERROR "Problem when removing $ENV{DESTDIR}${file}")
+    endif(NOT "${rm_retval}" STREQUAL 0)
+  else(IS_SYMLINK "$ENV{DESTDIR}${file}" OR EXISTS "$ENV{DESTDIR}${file}")
+    message(STATUS "File $ENV{DESTDIR}${file} does not exist.")
+  endif(IS_SYMLINK "$ENV{DESTDIR}${file}" OR EXISTS "$ENV{DESTDIR}${file}")
+endforeach(file)
\ No newline at end of file

From d607858b90b645d8177c3970d782f0ab5c529558 Mon Sep 17 00:00:00 2001
From: Zhou Mo <cdluminate@gmail.com>
Date: Tue, 9 Aug 2016 15:13:47 +0000
Subject: [PATCH 14/75] Fix more float comparison precision issue

With reference to this commit:
f1a8470aa21e35a5b2bb83007f8fb7680a354815

This fix changes some EXPECT_EQ into EXPECT_FLOAT_EQ .
---
 src/caffe/test/test_convolution_layer.cpp     | 2 +-
 src/caffe/test/test_gradient_based_solver.cpp | 8 ++++----
 src/caffe/test/test_neuron_layer.cpp          | 8 ++++----
 3 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/src/caffe/test/test_convolution_layer.cpp b/src/caffe/test/test_convolution_layer.cpp
index 9bb19d13592..85c10a29483 100644
--- a/src/caffe/test/test_convolution_layer.cpp
+++ b/src/caffe/test/test_convolution_layer.cpp
@@ -695,7 +695,7 @@ TYPED_TEST(ConvolutionLayerTest, TestNDAgainst2D) {
   }
   ASSERT_EQ(backward_result_nd.count(), backward_result_2d.count());
   for (int i = 0; i < backward_result_2d.count(); ++i) {
-    EXPECT_EQ(backward_result_2d.cpu_diff()[i],
+    EXPECT_FLOAT_EQ(backward_result_2d.cpu_diff()[i],
               backward_result_nd.cpu_diff()[i]);
   }
   ASSERT_EQ(backward_weight_result_nd.count(),
diff --git a/src/caffe/test/test_gradient_based_solver.cpp b/src/caffe/test/test_gradient_based_solver.cpp
index 975a8f0f88a..9395f4e95c6 100644
--- a/src/caffe/test/test_gradient_based_solver.cpp
+++ b/src/caffe/test/test_gradient_based_solver.cpp
@@ -538,9 +538,9 @@ class GradientBasedSolverTest : public MultiDeviceTest<TypeParam> {
     const vector<Blob<Dtype>*>& params = solver_->net()->learnable_params();
     for (int i = 0; i < params.size(); ++i) {
       for (int j = 0; j < params[i]->count(); ++j) {
-        EXPECT_EQ(param_copies[i]->cpu_data()[j], params[i]->cpu_data()[j])
+        EXPECT_FLOAT_EQ(param_copies[i]->cpu_data()[j], params[i]->cpu_data()[j])
             << "param " << i << " data differed at dim " << j;
-        EXPECT_EQ(param_copies[i]->cpu_diff()[j], params[i]->cpu_diff()[j])
+        EXPECT_FLOAT_EQ(param_copies[i]->cpu_diff()[j], params[i]->cpu_diff()[j])
             << "param " << i << " diff differed at dim " << j;
       }
     }
@@ -549,9 +549,9 @@ class GradientBasedSolverTest : public MultiDeviceTest<TypeParam> {
     const vector<shared_ptr<Blob<Dtype> > >& history = solver_->history();
     for (int i = 0; i < history.size(); ++i) {
       for (int j = 0; j < history[i]->count(); ++j) {
-        EXPECT_EQ(history_copies[i]->cpu_data()[j], history[i]->cpu_data()[j])
+        EXPECT_FLOAT_EQ(history_copies[i]->cpu_data()[j], history[i]->cpu_data()[j])
             << "history blob " << i << " data differed at dim " << j;
-        EXPECT_EQ(history_copies[i]->cpu_diff()[j], history[i]->cpu_diff()[j])
+        EXPECT_FLOAT_EQ(history_copies[i]->cpu_diff()[j], history[i]->cpu_diff()[j])
             << "history blob " << i << " diff differed at dim " << j;
       }
     }
diff --git a/src/caffe/test/test_neuron_layer.cpp b/src/caffe/test/test_neuron_layer.cpp
index 342f825cec3..57bd47b3a2e 100644
--- a/src/caffe/test/test_neuron_layer.cpp
+++ b/src/caffe/test/test_neuron_layer.cpp
@@ -791,16 +791,16 @@ TYPED_TEST(NeuronLayerTest, TestPReLUInPlace) {
   ip2.Backward(blob_middle_vec_2, propagate_down, blob_bottom_vec_2);
   // Check numbers
   for (int s = 0; s < blob_bottom_2->count(); ++s) {
-    EXPECT_EQ(this->blob_bottom_->cpu_diff()[s], blob_bottom_2->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(this->blob_bottom_->cpu_diff()[s], blob_bottom_2->cpu_diff()[s]);
   }
   for (int s = 0; s < ip.blobs()[0]->count(); ++s) {
-    EXPECT_EQ(ip.blobs()[0]->cpu_diff()[s], ip2.blobs()[0]->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(ip.blobs()[0]->cpu_diff()[s], ip2.blobs()[0]->cpu_diff()[s]);
   }
   for (int s = 0; s < ip.blobs()[1]->count(); ++s) {
-    EXPECT_EQ(ip.blobs()[1]->cpu_diff()[s], ip2.blobs()[1]->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(ip.blobs()[1]->cpu_diff()[s], ip2.blobs()[1]->cpu_diff()[s]);
   }
   for (int s = 0; s < prelu.blobs()[0]->count(); ++s) {
-    EXPECT_EQ(prelu.blobs()[0]->cpu_diff()[s],
+    EXPECT_FLOAT_EQ(prelu.blobs()[0]->cpu_diff()[s],
         prelu2.blobs()[0]->cpu_diff()[s]);
   }
 }

From 42d20fe21eeb8067b09ef5e935bb4c235dbf9f3f Mon Sep 17 00:00:00 2001
From: Zhou Mo <cdluminate@gmail.com>
Date: Wed, 10 Aug 2016 14:36:33 +0000
Subject: [PATCH 15/75] Import bash completion script for caffe from Debian
 Package.

Imported from Debian Package caffe (1.0.0~rc3+20160715-g42cd785-2).
---
 scripts/caffe | 73 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 73 insertions(+)
 create mode 100644 scripts/caffe

diff --git a/scripts/caffe b/scripts/caffe
new file mode 100644
index 00000000000..8a0b22af6ac
--- /dev/null
+++ b/scripts/caffe
@@ -0,0 +1,73 @@
+# bash completion for Caffe's command line utility       -*- shell-script -*-
+# COPYRIGHT (C) 2015,2016 Zhou Mo <cdluminate@gmail.com>
+# License: BSD-2-Clause
+# Originally appeard at https://github.com/BVLC/caffe/issues/3149
+
+# Updated for caffe (1.0.0~rc3+20160715-g42cd785)
+_caffe()
+{
+  local cur prev words cword
+  _init_completion -s || return
+
+  local prototxts='@(prototxt)'
+  local caffemodels='@(caffemodel,binaryproto)'
+  local solverstates='@(solverstate)'
+  local caffefiles='@(prototxt|caffemodel|solverstate)'
+
+  local flags='-gpu -iterations -model -snapshot -solver -weights -sighup_effect -sigint_effect -level -stage -phase'
+  
+  if [[ $cword -eq 1 ]]; then
+    COMPREPLY=( $( compgen -W 'train test time device_query' -- "$cur" ) )
+    return 0
+  fi
+  
+  if [[ $cword -eq 2 ]]; then
+    case ${words[1]} in
+    train|test|device_query|time)
+      COMPREPLY=( $( compgen -W "$flags" -- "$cur") )
+      return 0
+      ;;
+    *)
+      return 0
+      ;;
+    esac
+  fi
+
+  case $prev in
+  -gpu|-iterations|-version|-level|-stage)
+    return 0
+    ;;
+  -solver|-model)
+    _filedir $prototxts
+    return 0
+    ;;
+  -weights)
+    _filedir $caffemodels
+    return 0
+    ;;
+  -snapshot)
+    _filedir $solverstates
+    return 0
+    ;;
+  -sighup_effect|-sigint_effect)
+    COMPREPLY=( $( compgen -W 'snapshot stop none' -- "$cur") )
+    return 0
+    ;;
+  -phase)
+    COMPREPLY=( $( compgen -W 'TRAIN TEST' -- "$cur") )
+    return 0
+    ;;
+  *)
+    COMPREPLY=( $( compgen -W "$flags" -- "$cur") )
+    return 0
+    ;;
+  esac
+
+  # file completion on relevant files
+  _filedir "$caffefiles"
+
+  return 0
+}
+complete -F _caffe caffe
+
+# vim

From 6382d67da1d2b5d9ebe92df8a20a8ac1947366ea Mon Sep 17 00:00:00 2001
From: An Tran <tranlaman@gmail.com>
Date: Fri, 12 Aug 2016 16:39:11 +0800
Subject: [PATCH 16/75] small improments in compute_image_mean

---
 tools/compute_image_mean.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/tools/compute_image_mean.cpp b/tools/compute_image_mean.cpp
index 2035d515195..417f5e4c622 100644
--- a/tools/compute_image_mean.cpp
+++ b/tools/compute_image_mean.cpp
@@ -22,9 +22,11 @@ DEFINE_string(backend, "lmdb",
         "The backend {leveldb, lmdb} containing the images");
 
 int main(int argc, char** argv) {
+#ifdef USE_OPENCV
   ::google::InitGoogleLogging(argv[0]);
+  // Print output to stderr (while still logging)
+  FLAGS_alsologtostderr = 1;
 
-#ifdef USE_OPENCV
 #ifndef GFLAGS_GFLAGS_H_
   namespace gflags = google;
 #endif
@@ -65,7 +67,7 @@ int main(int argc, char** argv) {
   for (int i = 0; i < size_in_datum; ++i) {
     sum_blob.add_data(0.);
   }
-  LOG(INFO) << "Starting Iteration";
+  LOG(INFO) << "Starting iteration";
   while (cursor->valid()) {
     Datum datum;
     datum.ParseFromString(cursor->value());
@@ -114,7 +116,7 @@ int main(int argc, char** argv) {
     for (int i = 0; i < dim; ++i) {
       mean_values[c] += sum_blob.data(dim * c + i);
     }
-    LOG(INFO) << "mean_value channel [" << c << "]:" << mean_values[c] / dim;
+    LOG(INFO) << "mean_value channel [" << c << "]: " << mean_values[c] / dim;
   }
 #else
   LOG(FATAL) << "This tool requires OpenCV; compile with USE_OPENCV.";

From 95a436c601a04af620a0e166393d3ff695905bc4 Mon Sep 17 00:00:00 2001
From: max argus <argus.max@gmail.com>
Date: Thu, 25 Aug 2016 09:20:24 +0000
Subject: [PATCH 17/75] Fix: made load_hd5 check blob dims by default.

Size checks are needed for loading parameters to avoid strange bugs
when loading data we continue to reshape.
---
 include/caffe/util/hdf5.hpp               |  4 ++--
 src/caffe/layers/hdf5_data_layer.cpp      |  3 ++-
 src/caffe/test/test_hdf5_output_layer.cpp | 10 +++++----
 src/caffe/test/test_hdf5data_layer.cpp    |  2 +-
 src/caffe/util/hdf5.cpp                   | 34 +++++++++++++++++++++++++------
 5 files changed, 39 insertions(+), 14 deletions(-)

diff --git a/include/caffe/util/hdf5.hpp b/include/caffe/util/hdf5.hpp
index ce568c5eb0d..71549c1cc02 100644
--- a/include/caffe/util/hdf5.hpp
+++ b/include/caffe/util/hdf5.hpp
@@ -13,12 +13,12 @@ namespace caffe {
 template <typename Dtype>
 void hdf5_load_nd_dataset_helper(
     hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
-    Blob<Dtype>* blob);
+    Blob<Dtype>* blob, bool reshape);
 
 template <typename Dtype>
 void hdf5_load_nd_dataset(
     hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
-    Blob<Dtype>* blob);
+    Blob<Dtype>* blob, bool reshape = false);
 
 template <typename Dtype>
 void hdf5_save_nd_dataset(
diff --git a/src/caffe/layers/hdf5_data_layer.cpp b/src/caffe/layers/hdf5_data_layer.cpp
index 2f13dc641df..0099129000c 100644
--- a/src/caffe/layers/hdf5_data_layer.cpp
+++ b/src/caffe/layers/hdf5_data_layer.cpp
@@ -39,8 +39,9 @@ void HDF5DataLayer<Dtype>::LoadHDF5FileData(const char* filename) {
 
   for (int i = 0; i < top_size; ++i) {
     hdf_blobs_[i] = shared_ptr<Blob<Dtype> >(new Blob<Dtype>());
+    // Allow reshape here, as we are loading data not params
     hdf5_load_nd_dataset(file_id, this->layer_param_.top(i).c_str(),
-        MIN_DATA_DIM, MAX_DATA_DIM, hdf_blobs_[i].get());
+        MIN_DATA_DIM, MAX_DATA_DIM, hdf_blobs_[i].get(), true);
   }
 
   herr_t status = H5Fclose(file_id);
diff --git a/src/caffe/test/test_hdf5_output_layer.cpp b/src/caffe/test/test_hdf5_output_layer.cpp
index 3833ebff78e..2bc2de1e647 100644
--- a/src/caffe/test/test_hdf5_output_layer.cpp
+++ b/src/caffe/test/test_hdf5_output_layer.cpp
@@ -77,10 +77,12 @@ TYPED_TEST(HDF5OutputLayerTest, TestForward) {
                           H5P_DEFAULT);
   ASSERT_GE(file_id, 0)<< "Failed to open HDF5 file" <<
       this->input_file_name_;
+  // Allow reshape here as we are loading data not params
+  bool reshape = true;
   hdf5_load_nd_dataset(file_id, HDF5_DATA_DATASET_NAME, 0, 4,
-                       this->blob_data_);
+                       this->blob_data_, reshape);
   hdf5_load_nd_dataset(file_id, HDF5_DATA_LABEL_NAME, 0, 4,
-                       this->blob_label_);
+                       this->blob_label_, reshape);
   herr_t status = H5Fclose(file_id);
   EXPECT_GE(status, 0)<< "Failed to close HDF5 file " <<
       this->input_file_name_;
@@ -105,12 +107,12 @@ TYPED_TEST(HDF5OutputLayerTest, TestForward) {
 
   Blob<Dtype>* blob_data = new Blob<Dtype>();
   hdf5_load_nd_dataset(file_id, HDF5_DATA_DATASET_NAME, 0, 4,
-                       blob_data);
+                       blob_data, reshape);
   this->CheckBlobEqual(*(this->blob_data_), *blob_data);
 
   Blob<Dtype>* blob_label = new Blob<Dtype>();
   hdf5_load_nd_dataset(file_id, HDF5_DATA_LABEL_NAME, 0, 4,
-                       blob_label);
+                       blob_label, reshape);
   this->CheckBlobEqual(*(this->blob_label_), *blob_label);
 
   status = H5Fclose(file_id);
diff --git a/src/caffe/test/test_hdf5data_layer.cpp b/src/caffe/test/test_hdf5data_layer.cpp
index 8884ce95a23..e0fd62134c5 100644
--- a/src/caffe/test/test_hdf5data_layer.cpp
+++ b/src/caffe/test/test_hdf5data_layer.cpp
@@ -70,7 +70,7 @@ TYPED_TEST(HDF5DataLayerTest, TestRead) {
   int height = 6;
   int width = 5;
 
-  // Test that the layer setup got the correct parameters.
+  // Test that the layer setup gives correct parameters.
   HDF5DataLayer<Dtype> layer(param);
   layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
   EXPECT_EQ(this->blob_top_data_->num(), batch_size);
diff --git a/src/caffe/util/hdf5.cpp b/src/caffe/util/hdf5.cpp
index 7730e76ab87..0003f1b3988 100644
--- a/src/caffe/util/hdf5.cpp
+++ b/src/caffe/util/hdf5.cpp
@@ -9,7 +9,7 @@ namespace caffe {
 template <typename Dtype>
 void hdf5_load_nd_dataset_helper(
     hid_t file_id, const char* dataset_name_, int min_dim, int max_dim,
-    Blob<Dtype>* blob) {
+    Blob<Dtype>* blob, bool reshape) {
   // Verify that the dataset exists.
   CHECK(H5LTfind_dataset(file_id, dataset_name_))
       << "Failed to find HDF5 dataset " << dataset_name_;
@@ -56,17 +56,38 @@ void hdf5_load_nd_dataset_helper(
     LOG(FATAL) << "Datatype class unknown";
   }
 
+
   vector<int> blob_dims(dims.size());
   for (int i = 0; i < dims.size(); ++i) {
     blob_dims[i] = dims[i];
   }
-  blob->Reshape(blob_dims);
+
+  if (reshape) {
+    blob->Reshape(blob_dims);
+  } else {
+    if (blob_dims != blob->shape()) {
+      // create shape string for error message
+      ostringstream stream;
+      int count = 1;
+      for (int i = 0; i < blob_dims.size(); ++i) {
+        stream << blob_dims[i] << " ";
+        count = count * blob_dims[i];
+      }
+      stream << "(" << count << ")";
+      string source_shape_string = stream.str();
+
+      CHECK(blob_dims == blob->shape()) << "Cannot load blob from hdf5; shape "
+            << "mismatch. Source shape is " << source_shape_string
+            << " target shape is " << blob->shape_string();
+    }
+  }
 }
 
 template <>
 void hdf5_load_nd_dataset<float>(hid_t file_id, const char* dataset_name_,
-        int min_dim, int max_dim, Blob<float>* blob) {
-  hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob);
+        int min_dim, int max_dim, Blob<float>* blob, bool reshape) {
+  hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob,
+                              reshape);
   herr_t status = H5LTread_dataset_float(
     file_id, dataset_name_, blob->mutable_cpu_data());
   CHECK_GE(status, 0) << "Failed to read float dataset " << dataset_name_;
@@ -74,8 +95,9 @@ void hdf5_load_nd_dataset<float>(hid_t file_id, const char* dataset_name_,
 
 template <>
 void hdf5_load_nd_dataset<double>(hid_t file_id, const char* dataset_name_,
-        int min_dim, int max_dim, Blob<double>* blob) {
-  hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob);
+        int min_dim, int max_dim, Blob<double>* blob, bool reshape) {
+  hdf5_load_nd_dataset_helper(file_id, dataset_name_, min_dim, max_dim, blob,
+                              reshape);
   herr_t status = H5LTread_dataset_double(
     file_id, dataset_name_, blob->mutable_cpu_data());
   CHECK_GE(status, 0) << "Failed to read double dataset " << dataset_name_;

From 85ab6100a122042c7dfd4adaf06f4c0b2e71148d Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Mon, 27 Feb 2017 11:54:37 -0800
Subject: [PATCH 18/75] fix broken link to hinge loss

---
 docs/tutorial/layers.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/tutorial/layers.md b/docs/tutorial/layers.md
index a903d5ac985..2faacc5836d 100644
--- a/docs/tutorial/layers.md
+++ b/docs/tutorial/layers.md
@@ -128,7 +128,7 @@ Layers:
 * [Infogain Loss](layers/infogainloss.html) - a generalization of MultinomialLogisticLossLayer.
 * [Softmax with Loss](layers/softmaxwithloss.html) - computes the multinomial logistic loss of the softmax of its inputs. It's conceptually identical to a softmax layer followed by a multinomial logistic loss layer, but provides a more numerically stable gradient.
 * [Sum-of-Squares / Euclidean](layers/euclideanloss.html) - computes the sum of squares of differences of its two inputs, $$\frac 1 {2N} \sum_{i=1}^N \| x^1_i - x^2_i \|_2^2$$.
-* [Hinge / Margin](layers/hiddenloss.html) - The hinge loss layer computes a one-vs-all hinge (L1) or squared hinge loss (L2).
+* [Hinge / Margin](layers/hingeloss.html) - The hinge loss layer computes a one-vs-all hinge (L1) or squared hinge loss (L2).
 * [Sigmoid Cross-Entropy Loss](layers/sigmoidcrossentropyloss.html) - computes the cross-entropy (logistic) loss, often used for predicting targets interpreted as probabilities.
 * [Accuracy / Top-k layer](layers/accuracy.html) - scores the output as an accuracy with respect to target -- it is not actually a loss and has no backward step.
 * [Contrastive Loss](layers/contrastiveloss.html)

From fe9e58d6360d99cde0a883a06590631bb11911e0 Mon Sep 17 00:00:00 2001
From: zhuyuanhao <nju.zhuyuanhao@gmail.com>
Date: Wed, 1 Mar 2017 20:42:30 +0800
Subject: [PATCH 19/75] Remove not used variable in base_conv_layer.cpp

---
 src/caffe/layers/base_conv_layer.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/src/caffe/layers/base_conv_layer.cpp b/src/caffe/layers/base_conv_layer.cpp
index 4a4c68e009a..35c90145e31 100644
--- a/src/caffe/layers/base_conv_layer.cpp
+++ b/src/caffe/layers/base_conv_layer.cpp
@@ -19,7 +19,6 @@ void BaseConvolutionLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
   const int num_axes = bottom[0]->num_axes();
   num_spatial_axes_ = num_axes - first_spatial_axis;
   CHECK_GE(num_spatial_axes_, 0);
-  vector<int> bottom_dim_blob_shape(1, num_spatial_axes_ + 1);
   vector<int> spatial_dim_blob_shape(1, std::max(num_spatial_axes_, 1));
   // Setup filter kernel dimensions (kernel_shape_).
   kernel_shape_.Reshape(spatial_dim_blob_shape);

From 4529f12bdcd27d74655473b6665f5a23cd1214b1 Mon Sep 17 00:00:00 2001
From: gineshidalgo99 <gineshidalgo99@gmail.com>
Date: Thu, 9 Mar 2017 19:24:06 -0500
Subject: [PATCH 20/75] =?UTF-8?q?Removed=20some=20'warning:=20extra=20?=
 =?UTF-8?q?=E2=80=98;=E2=80=99=20[-Wpedantic]'?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 include/caffe/util/math_functions.hpp |  6 +++---
 include/caffe/util/mkl_alternate.hpp  | 18 +++++++++---------
 2 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/include/caffe/util/math_functions.hpp b/include/caffe/util/math_functions.hpp
index 51068fe2b80..37abce5eccc 100644
--- a/include/caffe/util/math_functions.hpp
+++ b/include/caffe/util/math_functions.hpp
@@ -128,16 +128,16 @@ inline int8_t caffe_sign(Dtype val) {
   }
 
 // output is 1 for the positives, 0 for zero, and -1 for the negatives
-DEFINE_CAFFE_CPU_UNARY_FUNC(sign, y[i] = caffe_sign<Dtype>(x[i]));
+DEFINE_CAFFE_CPU_UNARY_FUNC(sign, y[i] = caffe_sign<Dtype>(x[i]))
 
 // This returns a nonzero value if the input has its sign bit set.
 // The name sngbit is meant to avoid conflicts with std::signbit in the macro.
 // The extra parens are needed because CUDA < 6.5 defines signbit as a macro,
 // and we don't want that to expand here when CUDA headers are also included.
 DEFINE_CAFFE_CPU_UNARY_FUNC(sgnbit, \
-    y[i] = static_cast<bool>((std::signbit)(x[i])));
+    y[i] = static_cast<bool>((std::signbit)(x[i])))
 
-DEFINE_CAFFE_CPU_UNARY_FUNC(fabs, y[i] = std::fabs(x[i]));
+DEFINE_CAFFE_CPU_UNARY_FUNC(fabs, y[i] = std::fabs(x[i]))
 
 template <typename Dtype>
 void caffe_cpu_scale(const int n, const Dtype alpha, const Dtype *x, Dtype* y);
diff --git a/include/caffe/util/mkl_alternate.hpp b/include/caffe/util/mkl_alternate.hpp
index 95df0f93b5e..79b2c32de94 100644
--- a/include/caffe/util/mkl_alternate.hpp
+++ b/include/caffe/util/mkl_alternate.hpp
@@ -36,10 +36,10 @@ extern "C" {
     v##name<double>(n, a, y); \
   }
 
-DEFINE_VSL_UNARY_FUNC(Sqr, y[i] = a[i] * a[i]);
-DEFINE_VSL_UNARY_FUNC(Exp, y[i] = exp(a[i]));
-DEFINE_VSL_UNARY_FUNC(Ln, y[i] = log(a[i]));
-DEFINE_VSL_UNARY_FUNC(Abs, y[i] = fabs(a[i]));
+DEFINE_VSL_UNARY_FUNC(Sqr, y[i] = a[i] * a[i])
+DEFINE_VSL_UNARY_FUNC(Exp, y[i] = exp(a[i]))
+DEFINE_VSL_UNARY_FUNC(Ln, y[i] = log(a[i]))
+DEFINE_VSL_UNARY_FUNC(Abs, y[i] = fabs(a[i]))
 
 // A simple way to define the vsl unary functions with singular parameter b.
 // The operation should be in the form e.g. y[i] = pow(a[i], b)
@@ -58,7 +58,7 @@ DEFINE_VSL_UNARY_FUNC(Abs, y[i] = fabs(a[i]));
     v##name<double>(n, a, b, y); \
   }
 
-DEFINE_VSL_UNARY_FUNC_WITH_PARAM(Powx, y[i] = pow(a[i], b));
+DEFINE_VSL_UNARY_FUNC_WITH_PARAM(Powx, y[i] = pow(a[i], b))
 
 // A simple way to define the vsl binary functions. The operation should
 // be in the form e.g. y[i] = a[i] + b[i]
@@ -77,10 +77,10 @@ DEFINE_VSL_UNARY_FUNC_WITH_PARAM(Powx, y[i] = pow(a[i], b));
     v##name<double>(n, a, b, y); \
   }
 
-DEFINE_VSL_BINARY_FUNC(Add, y[i] = a[i] + b[i]);
-DEFINE_VSL_BINARY_FUNC(Sub, y[i] = a[i] - b[i]);
-DEFINE_VSL_BINARY_FUNC(Mul, y[i] = a[i] * b[i]);
-DEFINE_VSL_BINARY_FUNC(Div, y[i] = a[i] / b[i]);
+DEFINE_VSL_BINARY_FUNC(Add, y[i] = a[i] + b[i])
+DEFINE_VSL_BINARY_FUNC(Sub, y[i] = a[i] - b[i])
+DEFINE_VSL_BINARY_FUNC(Mul, y[i] = a[i] * b[i])
+DEFINE_VSL_BINARY_FUNC(Div, y[i] = a[i] / b[i])
 
 // In addition, MKL comes with an additional function axpby that is not present
 // in standard blas. We will simply use a two-step (inefficient, of course) way

From 1d3e6e4522a95faf954e775b23a2f907e66caf31 Mon Sep 17 00:00:00 2001
From: folz <joachim.folz@dfki.de>
Date: Mon, 13 Mar 2017 11:04:30 +0100
Subject: [PATCH 21/75] Solver_add_nccl accepts any kind of Solver

---
 python/caffe/_caffe.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/caffe/_caffe.cpp b/python/caffe/_caffe.cpp
index 3589e476f5c..be011699098 100644
--- a/python/caffe/_caffe.cpp
+++ b/python/caffe/_caffe.cpp
@@ -288,7 +288,7 @@ void Solver_add_callback(Solver<Dtype> * solver, bp::object on_start,
 }
 
 // Seems boost cannot call the base method directly
-void Solver_add_nccl(SGDSolver<Dtype>* solver
+void Solver_add_nccl(Solver<Dtype>* solver
 #ifdef USE_NCCL
   , NCCL<Dtype>* nccl
 #endif

From 93993a3c2b25ad683dbf13ef3085b0ea5912911f Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Tue, 14 Mar 2017 15:41:40 -0700
Subject: [PATCH 22/75] Init test net on all GPUs, allows parallel inference

---
 src/caffe/solver.cpp | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/caffe/solver.cpp b/src/caffe/solver.cpp
index fd4c03724ef..044269371ad 100644
--- a/src/caffe/solver.cpp
+++ b/src/caffe/solver.cpp
@@ -51,8 +51,8 @@ void Solver<Dtype>::Init(const SolverParameter& param) {
   }
   // Scaffolding code
   InitTrainNet();
+  InitTestNets();
   if (Caffe::root_solver()) {
-    InitTestNets();
     LOG(INFO) << "Solver scaffolding done.";
   }
   iter_ = 0;
@@ -102,7 +102,6 @@ void Solver<Dtype>::InitTrainNet() {
 
 template <typename Dtype>
 void Solver<Dtype>::InitTestNets() {
-  CHECK(Caffe::root_solver());
   const bool has_net_param = param_.has_net_param();
   const bool has_net_file = param_.has_net();
   const int num_generic_nets = has_net_param + has_net_file;

From 802d90fe81f04e5e9c28c088da0f1b22e1b9fed2 Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Thu, 16 Mar 2017 23:08:20 -0400
Subject: [PATCH 23/75] Added python 3 compatibility to cpp_lint.py

---
 scripts/cpp_lint.py | 39 ++++++++++++++++++++++-----------------
 1 file changed, 22 insertions(+), 17 deletions(-)

diff --git a/scripts/cpp_lint.py b/scripts/cpp_lint.py
index 6ec4fb76e2c..b2016d4b6dd 100755
--- a/scripts/cpp_lint.py
+++ b/scripts/cpp_lint.py
@@ -1,4 +1,4 @@
-#!/usr/bin/python2
+#!/usr/bin/env python
 #
 # Copyright (c) 2009 Google Inc. All rights reserved.
 #
@@ -52,6 +52,10 @@
 import sys
 import unicodedata
 
+import six
+
+from six import iteritems, itervalues
+from six.moves import xrange
 
 _USAGE = """
 Syntax: cpp_lint.py [--verbose=#] [--output=vs7] [--filter=-x,+y,...]
@@ -756,7 +760,7 @@ def IncrementErrorCount(self, category):
 
   def PrintErrorCounts(self):
     """Print a summary of errors by category, and the total."""
-    for category, count in self.errors_by_category.iteritems():
+    for category, count in iteritems(self.errors_by_category):
       sys.stderr.write('Category \'%s\' errors found: %d\n' %
                        (category, count))
     sys.stderr.write('Total errors found: %d\n' % self.error_count)
@@ -3444,16 +3448,16 @@ def GetLineWidth(line):
     The width of the line in column positions, accounting for Unicode
     combining characters and wide characters.
   """
-  if isinstance(line, unicode):
-    width = 0
-    for uc in unicodedata.normalize('NFC', line):
-      if unicodedata.east_asian_width(uc) in ('W', 'F'):
-        width += 2
-      elif not unicodedata.combining(uc):
-        width += 1
-    return width
-  else:
-    return len(line)
+  if six.PY2:
+    if isinstance(line, unicode):
+      width = 0
+      for uc in unicodedata.normalize('NFC', line):
+        if unicodedata.east_asian_width(uc) in ('W', 'F'):
+          width += 2
+        elif not unicodedata.combining(uc):
+          width += 1
+      return width
+  return len(line)
 
 
 def CheckStyle(filename, clean_lines, linenum, file_extension, nesting_state,
@@ -3774,7 +3778,7 @@ def _GetTextInside(text, start_pattern):
 
   # Give opening punctuations to get the matching close-punctuations.
   matching_punctuation = {'(': ')', '{': '}', '[': ']'}
-  closing_punctuation = set(matching_punctuation.itervalues())
+  closing_punctuation = set(itervalues(matching_punctuation))
 
   # Find the position to start extracting text.
   match = re.search(start_pattern, text, re.M)
@@ -4851,10 +4855,11 @@ def main():
 
   # Change stderr to write with replacement characters so we don't die
   # if we try to print something containing non-ASCII characters.
-  sys.stderr = codecs.StreamReaderWriter(sys.stderr,
-                                         codecs.getreader('utf8'),
-                                         codecs.getwriter('utf8'),
-                                         'replace')
+  if six.PY2:
+    sys.stderr = codecs.StreamReaderWriter(sys.stderr,
+                                          codecs.getreader('utf8'),
+                                          codecs.getwriter('utf8'),
+                                          'replace')
 
   _cpplint_state.ResetErrorCounts()
   for filename in filenames:

From accd188d3241c27a6d24b95cd95a4dca4f4078bc Mon Sep 17 00:00:00 2001
From: max argus <argus.max@gmail.com>
Date: Wed, 8 Mar 2017 15:04:29 +0000
Subject: [PATCH 24/75] sane h5df file type check for weights

---
 src/caffe/net.cpp | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/caffe/net.cpp b/src/caffe/net.cpp
index 70d51806d8a..353c2f95b9e 100644
--- a/src/caffe/net.cpp
+++ b/src/caffe/net.cpp
@@ -769,8 +769,7 @@ void Net<Dtype>::CopyTrainedLayersFrom(const NetParameter& param) {
 
 template <typename Dtype>
 void Net<Dtype>::CopyTrainedLayersFrom(const string trained_filename) {
-  if (trained_filename.size() >= 3 &&
-      trained_filename.compare(trained_filename.size() - 3, 3, ".h5") == 0) {
+  if (H5Fis_hdf5(trained_filename.c_str())) {
     CopyTrainedLayersFromHDF5(trained_filename);
   } else {
     CopyTrainedLayersFromBinaryProto(trained_filename);

From 11930f1416efb66795e1fabc5e362a568446d37d Mon Sep 17 00:00:00 2001
From: "Jonathan R. Williford" <jonathan@neural.vision>
Date: Wed, 22 Mar 2017 22:36:14 +0100
Subject: [PATCH 25/75] Clarify batch norm parameter documentation.

---
 src/caffe/proto/caffe.proto | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index a145c541957..02e0ddf57c1 100644
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -502,11 +502,21 @@ message ConcatParameter {
 }
 
 message BatchNormParameter {
-  // If false, accumulate global mean/variance values via a moving average. If
-  // true, use those accumulated values instead of computing mean/variance
-  // across the batch.
+  // If false, normalization is performed over the current mini-batch
+  // and global statistics are accumulated (but not yet used) by a moving
+  // average.
+  // If true, those accumulated mean and variance values are used for the
+  // normalization.
+  // By default, it is set to false when the network is in the training
+  // phase and true when the network is in the testing phase.
   optional bool use_global_stats = 1;
-  // How much does the moving average decay each iteration?
+  // What fraction of the moving average remains each iteration?
+  // Smaller values make the moving average decay faster, giving more
+  // weight to the recent values.
+  // Each iteration updates the moving average @f$S_{t-1}@f$ with the
+  // current mean @f$ Y_t @f$ by
+  // @f$ S_t = (1-\beta)Y_t + \beta \cdot S_{t-1} @f$, where @f$ \beta @f$
+  // is the moving_average_fraction parameter.
   optional float moving_average_fraction = 2 [default = .999];
   // Small value to add to the variance estimate so that we don't divide by
   // zero.

From 5c8e3545c650e9d3924f707334bde7cd67cf4e07 Mon Sep 17 00:00:00 2001
From: max argus <argus.max@gmail.com>
Date: Wed, 22 Mar 2017 23:15:34 +0000
Subject: [PATCH 26/75] [caffe][build] added Atlas lapack Library name
 atllapack

---
 cmake/Modules/FindAtlas.cmake | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/cmake/Modules/FindAtlas.cmake b/cmake/Modules/FindAtlas.cmake
index 9c665a47bd5..7ffa6393bbc 100644
--- a/cmake/Modules/FindAtlas.cmake
+++ b/cmake/Modules/FindAtlas.cmake
@@ -28,7 +28,7 @@ find_path(Atlas_CLAPACK_INCLUDE_DIR NAMES clapack.h PATHS ${Atlas_INCLUDE_SEARCH
 
 find_library(Atlas_CBLAS_LIBRARY NAMES  ptcblas_r ptcblas cblas_r cblas       PATHS ${Atlas_LIB_SEARCH_PATHS})
 find_library(Atlas_BLAS_LIBRARY NAMES   atlas_r   atlas                       PATHS ${Atlas_LIB_SEARCH_PATHS})
-find_library(Atlas_LAPACK_LIBRARY NAMES lapack alapack_r alapack lapack_atlas PATHS ${Atlas_LIB_SEARCH_PATHS})
+find_library(Atlas_LAPACK_LIBRARY NAMES lapack alapack_r alapack lapack_atlas atllapack PATHS ${Atlas_LIB_SEARCH_PATHS})
 
 set(LOOKED_FOR
   Atlas_CBLAS_INCLUDE_DIR
@@ -47,6 +47,6 @@ if(ATLAS_FOUND)
   set(Atlas_LIBRARIES ${Atlas_LAPACK_LIBRARY} ${Atlas_CBLAS_LIBRARY} ${Atlas_BLAS_LIBRARY})
   mark_as_advanced(${LOOKED_FOR})
 
-  message(STATUS "Found Atlas (include: ${Atlas_CBLAS_INCLUDE_DIR}, library: ${Atlas_BLAS_LIBRARY})")
+  message(STATUS "Found Atlas (include: ${Atlas_CBLAS_INCLUDE_DIR} library: ${Atlas_BLAS_LIBRARY} lapack: ${Atlas_LAPACK_LIBRARY}")
 endif(ATLAS_FOUND)
 

From 1e02d622da5aa01fbcf1185bced8e4b0daa0a50b Mon Sep 17 00:00:00 2001
From: max argus <argus.max@gmail.com>
Date: Wed, 22 Mar 2017 23:24:13 +0000
Subject: [PATCH 27/75] [caffe][build] added ABS_TEST_DATA_DIR var.

---
 cmake/Templates/caffe_config.h.in             | 15 ++++-----------
 include/caffe/test/test_caffe_main.hpp        |  3 +--
 src/caffe/test/test_gradient_based_solver.cpp |  2 +-
 src/caffe/test/test_hdf5_output_layer.cpp     |  3 +--
 src/caffe/test/test_hdf5data_layer.cpp        |  3 +--
 5 files changed, 8 insertions(+), 18 deletions(-)

diff --git a/cmake/Templates/caffe_config.h.in b/cmake/Templates/caffe_config.h.in
index 45465b98305..2080c63df36 100644
--- a/cmake/Templates/caffe_config.h.in
+++ b/cmake/Templates/caffe_config.h.in
@@ -4,16 +4,9 @@
 /* Binaries directory */
 #define BINARY_FOLDER "${PROJECT_BINARY_DIR}"
 
+/* This is an absolute path so that we can run test from any build
+ * directory */
+#define ABS_TEST_DATA_DIR "${PROJECT_SOURCE_DIR}/src/caffe/test/test_data/"
+
 /* Test device */
 #define CUDA_TEST_DEVICE ${CUDA_TEST_DEVICE}
-
-/* Temporary (TODO: remove) */
-#if 1
-  #define CMAKE_SOURCE_DIR SOURCE_FOLDER "/src/"
-  #define EXAMPLES_SOURCE_DIR BINARY_FOLDER "/examples/"
-  #define CMAKE_EXT ".gen.cmake"
-#else
-  #define CMAKE_SOURCE_DIR "src/"
-  #define EXAMPLES_SOURCE_DIR "examples/"
-  #define CMAKE_EXT ""
-#endif
diff --git a/include/caffe/test/test_caffe_main.hpp b/include/caffe/test/test_caffe_main.hpp
index fc156091476..294f7e5011a 100644
--- a/include/caffe/test/test_caffe_main.hpp
+++ b/include/caffe/test/test_caffe_main.hpp
@@ -18,9 +18,8 @@ using std::endl;
   #include "caffe_config.h"
 #else
   #define CUDA_TEST_DEVICE -1
-  #define CMAKE_SOURCE_DIR "src/"
   #define EXAMPLES_SOURCE_DIR "examples/"
-  #define CMAKE_EXT ""
+  #define ABS_TEST_DATA_DIR "src/caffe/test/test_data"
 #endif
 
 int main(int argc, char** argv);
diff --git a/src/caffe/test/test_gradient_based_solver.cpp b/src/caffe/test/test_gradient_based_solver.cpp
index 6ad0d8f6544..465140f28a6 100644
--- a/src/caffe/test/test_gradient_based_solver.cpp
+++ b/src/caffe/test/test_gradient_based_solver.cpp
@@ -28,7 +28,7 @@ class GradientBasedSolverTest : public MultiDeviceTest<TypeParam> {
       seed_(1701), num_(4), channels_(3), height_(10), width_(10),
       share_(false) {
         input_file_ = new string(
-        CMAKE_SOURCE_DIR "caffe/test/test_data/solver_data_list.txt" CMAKE_EXT);
+        ABS_TEST_DATA_DIR "/solver_data_list.txt");
       }
   ~GradientBasedSolverTest() {
     delete input_file_;
diff --git a/src/caffe/test/test_hdf5_output_layer.cpp b/src/caffe/test/test_hdf5_output_layer.cpp
index 2bc2de1e647..f94dd57e7de 100644
--- a/src/caffe/test/test_hdf5_output_layer.cpp
+++ b/src/caffe/test/test_hdf5_output_layer.cpp
@@ -20,8 +20,7 @@ class HDF5OutputLayerTest : public MultiDeviceTest<TypeParam> {
 
  protected:
   HDF5OutputLayerTest()
-      : input_file_name_(
-        CMAKE_SOURCE_DIR "caffe/test/test_data/sample_data.h5"),
+      : input_file_name_(ABS_TEST_DATA_DIR "/sample_data.h5"),
         blob_data_(new Blob<Dtype>()),
         blob_label_(new Blob<Dtype>()),
         num_(5),
diff --git a/src/caffe/test/test_hdf5data_layer.cpp b/src/caffe/test/test_hdf5data_layer.cpp
index 487f5176caf..3977c4866c7 100644
--- a/src/caffe/test/test_hdf5data_layer.cpp
+++ b/src/caffe/test/test_hdf5data_layer.cpp
@@ -30,8 +30,7 @@ class HDF5DataLayerTest : public MultiDeviceTest<TypeParam> {
     blob_top_vec_.push_back(blob_top_label2_);
 
     // Check out generate_sample_data.py in the same directory.
-    filename = new string(
-    CMAKE_SOURCE_DIR "caffe/test/test_data/sample_data_list.txt" CMAKE_EXT);
+    filename = new string(ABS_TEST_DATA_DIR "/sample_data_list.txt");
     LOG(INFO)<< "Using sample HDF5 data file " << filename;
   }
 

From 8602a238a712d50ac5a2d7dffadee2f34d755e3f Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Mon, 27 Mar 2017 11:33:06 -0700
Subject: [PATCH 28/75] Expose share_weights to python to allow running test
 nets

---
 python/caffe/_caffe.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/python/caffe/_caffe.cpp b/python/caffe/_caffe.cpp
index be011699098..276f21f85a5 100644
--- a/python/caffe/_caffe.cpp
+++ b/python/caffe/_caffe.cpp
@@ -298,6 +298,10 @@ void Solver_add_nccl(Solver<Dtype>* solver
 #endif
 }
 
+void share_weights(Solver<Dtype>* solver, Net<Dtype>* net) {
+  net->ShareTrainedLayersWith(solver->net().get());
+}
+
 template<typename Dtype>
 class NetCallback: public Net<Dtype>::Callback {
  public:
@@ -459,6 +463,7 @@ BOOST_PYTHON_MODULE(_caffe) {
     .def("step", &Solver<Dtype>::Step)
     .def("restore", &Solver<Dtype>::Restore)
     .def("snapshot", &Solver<Dtype>::Snapshot)
+    .def("share_weights", &share_weights)
     .add_property("param", bp::make_function(&Solver<Dtype>::param,
               bp::return_value_policy<bp::copy_const_reference>()));
   BP_REGISTER_SHARED_PTR_TO_PYTHON(Solver<Dtype>);

From 850ffd8d1cf18cabe36eb269b63d693db2b167ef Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Mon, 27 Mar 2017 13:15:18 -0700
Subject: [PATCH 29/75] Remove missed legacy parallel code

---
 include/caffe/layers/base_data_layer.hpp   | 2 --
 include/caffe/layers/data_layer.hpp        | 2 --
 include/caffe/layers/dummy_data_layer.hpp  | 2 --
 include/caffe/layers/hdf5_data_layer.hpp   | 2 --
 include/caffe/layers/hdf5_output_layer.hpp | 2 --
 include/caffe/layers/input_layer.hpp       | 2 --
 include/caffe/layers/python_layer.hpp      | 4 ----
 src/caffe/proto/caffe.proto                | 4 +---
 8 files changed, 1 insertion(+), 19 deletions(-)

diff --git a/include/caffe/layers/base_data_layer.hpp b/include/caffe/layers/base_data_layer.hpp
index 21d3ada50d0..c8b6998c8f2 100644
--- a/include/caffe/layers/base_data_layer.hpp
+++ b/include/caffe/layers/base_data_layer.hpp
@@ -26,8 +26,6 @@ class BaseDataLayer : public Layer<Dtype> {
   // This method may not be overridden except by the BasePrefetchingDataLayer.
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // Data layers should be shared by multiple solvers in parallel
-  virtual inline bool ShareInParallel() const { return true; }
   virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
   // Data layers have no bottoms, so reshaping is trivial.
diff --git a/include/caffe/layers/data_layer.hpp b/include/caffe/layers/data_layer.hpp
index dec58180976..667a4ae43a5 100644
--- a/include/caffe/layers/data_layer.hpp
+++ b/include/caffe/layers/data_layer.hpp
@@ -20,8 +20,6 @@ class DataLayer : public BasePrefetchingDataLayer<Dtype> {
   virtual ~DataLayer();
   virtual void DataLayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // DataLayer uses DataReader instead for sharing for parallelism
-  virtual inline bool ShareInParallel() const { return false; }
   virtual inline const char* type() const { return "Data"; }
   virtual inline int ExactNumBottomBlobs() const { return 0; }
   virtual inline int MinTopBlobs() const { return 1; }
diff --git a/include/caffe/layers/dummy_data_layer.hpp b/include/caffe/layers/dummy_data_layer.hpp
index 4180f1d01e4..13a63d47ec4 100644
--- a/include/caffe/layers/dummy_data_layer.hpp
+++ b/include/caffe/layers/dummy_data_layer.hpp
@@ -22,8 +22,6 @@ class DummyDataLayer : public Layer<Dtype> {
       : Layer<Dtype>(param) {}
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // Data layers should be shared by multiple solvers in parallel
-  virtual inline bool ShareInParallel() const { return true; }
   // Data layers have no bottoms, so reshaping is trivial.
   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
diff --git a/include/caffe/layers/hdf5_data_layer.hpp b/include/caffe/layers/hdf5_data_layer.hpp
index 650a3fb0c87..601b36c6b89 100644
--- a/include/caffe/layers/hdf5_data_layer.hpp
+++ b/include/caffe/layers/hdf5_data_layer.hpp
@@ -27,8 +27,6 @@ class HDF5DataLayer : public Layer<Dtype> {
   virtual ~HDF5DataLayer();
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // Data layers should be shared by multiple solvers in parallel
-  virtual inline bool ShareInParallel() const { return true; }
   // Data layers have no bottoms, so reshaping is trivial.
   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
diff --git a/include/caffe/layers/hdf5_output_layer.hpp b/include/caffe/layers/hdf5_output_layer.hpp
index 487d08fc06c..061e279d7a0 100644
--- a/include/caffe/layers/hdf5_output_layer.hpp
+++ b/include/caffe/layers/hdf5_output_layer.hpp
@@ -28,8 +28,6 @@ class HDF5OutputLayer : public Layer<Dtype> {
   virtual ~HDF5OutputLayer();
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // Data layers should be shared by multiple solvers in parallel
-  virtual inline bool ShareInParallel() const { return true; }
   // Data layers have no bottoms, so reshaping is trivial.
   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
diff --git a/include/caffe/layers/input_layer.hpp b/include/caffe/layers/input_layer.hpp
index f4472678c69..0ffdc724894 100644
--- a/include/caffe/layers/input_layer.hpp
+++ b/include/caffe/layers/input_layer.hpp
@@ -22,8 +22,6 @@ class InputLayer : public Layer<Dtype> {
       : Layer<Dtype>(param) {}
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
-  // Data layers should be shared by multiple solvers in parallel
-  virtual inline bool ShareInParallel() const { return true; }
   // Data layers have no bottoms, so reshaping is trivial.
   virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
diff --git a/include/caffe/layers/python_layer.hpp b/include/caffe/layers/python_layer.hpp
index 10c4bfd0250..1407d9217aa 100644
--- a/include/caffe/layers/python_layer.hpp
+++ b/include/caffe/layers/python_layer.hpp
@@ -34,10 +34,6 @@ class PythonLayer : public Layer<Dtype> {
     self_.attr("reshape")(bottom, top);
   }
 
-  virtual inline bool ShareInParallel() const {
-    return this->layer_param_.python_param().share_in_parallel();
-  }
-
   virtual inline const char* type() const { return "Python"; }
 
  protected:
diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index 02e0ddf57c1..8e528e8e083 100644
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -937,9 +937,7 @@ message PythonParameter {
   // string, dictionary in Python dict format, JSON, etc. You may parse this
   // string in `setup` method and use it in `forward` and `backward`.
   optional string param_str = 3 [default = ''];
-  // Whether this PythonLayer is shared among worker solvers during data parallelism.
-  // If true, each worker solver sequentially run forward from this layer.
-  // This value should be set true if you are using it as a data layer.
+  // DEPRECATED
   optional bool share_in_parallel = 4 [default = false];
 }
 

From 9bd80b2f12649c6336b64c8ebcc2d1210755d1c7 Mon Sep 17 00:00:00 2001
From: Yuduo Wu <yuduowu@users.noreply.github.com>
Date: Wed, 29 Mar 2017 14:42:36 -0700
Subject: [PATCH 30/75] Fix typo in test_caffe_main.cpp: defice -> device

---
 src/caffe/test/test_caffe_main.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/caffe/test/test_caffe_main.cpp b/src/caffe/test/test_caffe_main.cpp
index 6473b74d0a6..8f333bd7105 100644
--- a/src/caffe/test/test_caffe_main.cpp
+++ b/src/caffe/test/test_caffe_main.cpp
@@ -15,7 +15,7 @@ int main(int argc, char** argv) {
   ::testing::InitGoogleTest(&argc, argv);
   caffe::GlobalInit(&argc, &argv);
 #ifndef CPU_ONLY
-  // Before starting testing, let's first print out a few cuda defice info.
+  // Before starting testing, let's first print out a few cuda device info.
   int device;
   cudaGetDeviceCount(&device);
   cout << "Cuda number of devices: " << device << endl;

From a32114e6b2e098e2fdef47e397542b105eb58b66 Mon Sep 17 00:00:00 2001
From: Will Crichton <wcrichto@stanford.edu>
Date: Fri, 31 Mar 2017 11:22:22 -0400
Subject: [PATCH 31/75] Fixed memory leaks in cudnn conv and relu

---
 src/caffe/layers/cudnn_conv_layer.cpp | 1 +
 src/caffe/layers/cudnn_relu_layer.cpp | 1 +
 2 files changed, 2 insertions(+)

diff --git a/src/caffe/layers/cudnn_conv_layer.cpp b/src/caffe/layers/cudnn_conv_layer.cpp
index 1987fb096b0..efc9e04e8c0 100644
--- a/src/caffe/layers/cudnn_conv_layer.cpp
+++ b/src/caffe/layers/cudnn_conv_layer.cpp
@@ -252,6 +252,7 @@ CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
   }
 
   cudaFree(workspaceData);
+  delete [] workspace;
   delete [] stream_;
   delete [] handle_;
   delete [] fwd_algo_;
diff --git a/src/caffe/layers/cudnn_relu_layer.cpp b/src/caffe/layers/cudnn_relu_layer.cpp
index 795e0a9efb0..687c905763e 100644
--- a/src/caffe/layers/cudnn_relu_layer.cpp
+++ b/src/caffe/layers/cudnn_relu_layer.cpp
@@ -36,6 +36,7 @@ CuDNNReLULayer<Dtype>::~CuDNNReLULayer() {
 
   cudnnDestroyTensorDescriptor(this->bottom_desc_);
   cudnnDestroyTensorDescriptor(this->top_desc_);
+  cudnnDestroyActivationDescriptor(this->activ_desc_);
   cudnnDestroy(this->handle_);
 }
 

From a2601eddf65bab54429244e350899b6d994f4f37 Mon Sep 17 00:00:00 2001
From: Luke Yeager <lukeyeager@users.noreply.github.com>
Date: Fri, 31 Mar 2017 11:01:13 -0700
Subject: [PATCH 32/75] Revert "Fix Python net drawing script"

This reverts commit db6cf0a728cad63c93b345f2203f3ad1f5d5c2f4.
---
 python/caffe/draw.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/python/caffe/draw.py b/python/caffe/draw.py
index e4fd7aacce7..9eecf6d7b46 100644
--- a/python/caffe/draw.py
+++ b/python/caffe/draw.py
@@ -104,11 +104,11 @@ def get_layer_label(layer, rankdir):
                       pooling_types_dict[layer.pooling_param.pool],
                       layer.type,
                       separator,
-                      layer.pooling_param.kernel_size[0] if len(layer.pooling_param.kernel_size._values) else 1,
+                      layer.pooling_param.kernel_size,
                       separator,
-                      layer.pooling_param.stride[0] if len(layer.pooling_param.stride._values) else 1,
+                      layer.pooling_param.stride,
                       separator,
-                      layer.pooling_param.pad[0] if len(layer.pooling_param.pad._values) else 0)
+                      layer.pooling_param.pad)
     else:
         node_label = '"%s%s(%s)"' % (layer.name, separator, layer.type)
     return node_label

From 0096fe3d270a4833479076e18492de8b28564c80 Mon Sep 17 00:00:00 2001
From: Felix Abecassis <fabecassis@nvidia.com>
Date: Fri, 31 Mar 2017 11:18:39 -0700
Subject: [PATCH 33/75] Add support for cuDNN v6

Support for cuDNN v4 and v5 is preserved.
---
 docs/installation.md           |  4 ++--
 include/caffe/util/cudnn.hpp   | 10 ++++++++++
 scripts/travis/install-deps.sh |  2 +-
 3 files changed, 13 insertions(+), 3 deletions(-)

diff --git a/docs/installation.md b/docs/installation.md
index 2e558027678..42f1d0ce09b 100644
--- a/docs/installation.md
+++ b/docs/installation.md
@@ -42,14 +42,14 @@ Optional dependencies:
 
 * [OpenCV](http://opencv.org/) >= 2.4 including 3.0
 * IO libraries: `lmdb`, `leveldb` (note: leveldb requires `snappy`)
-* cuDNN for GPU acceleration (v5)
+* cuDNN for GPU acceleration (v6)
 
 Pycaffe and Matcaffe interfaces have their own natural needs.
 
 * For Python Caffe:  `Python 2.7` or `Python 3.3+`, `numpy (>= 1.7)`, boost-provided `boost.python`
 * For MATLAB Caffe: MATLAB with the `mex` compiler.
 
-**cuDNN Caffe**: for fastest operation Caffe is accelerated by drop-in integration of [NVIDIA cuDNN](https://developer.nvidia.com/cudnn). To speed up your Caffe models, install cuDNN then uncomment the `USE_CUDNN := 1` flag in `Makefile.config` when installing Caffe. Acceleration is automatic. The current version is cuDNN v5; older versions are supported in older Caffe.
+**cuDNN Caffe**: for fastest operation Caffe is accelerated by drop-in integration of [NVIDIA cuDNN](https://developer.nvidia.com/cudnn). To speed up your Caffe models, install cuDNN then uncomment the `USE_CUDNN := 1` flag in `Makefile.config` when installing Caffe. Acceleration is automatic. The current version is cuDNN v6; older versions are supported in older Caffe.
 
 **CPU-only Caffe**: for cold-brewed CPU-only Caffe uncomment the `CPU_ONLY := 1` flag in `Makefile.config` to configure and build Caffe without CUDA. This is helpful for cloud or cluster deployment.
 
diff --git a/include/caffe/util/cudnn.hpp b/include/caffe/util/cudnn.hpp
index a7d8dbbad4c..498cfe385de 100644
--- a/include/caffe/util/cudnn.hpp
+++ b/include/caffe/util/cudnn.hpp
@@ -41,6 +41,10 @@ inline const char* cudnnGetErrorString(cudnnStatus_t status) {
       return "CUDNN_STATUS_NOT_SUPPORTED";
     case CUDNN_STATUS_LICENSE_ERROR:
       return "CUDNN_STATUS_LICENSE_ERROR";
+#if CUDNN_VERSION_MIN(6, 0, 0)
+    case CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING:
+      return "CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING";
+#endif
   }
   return "Unknown cudnn status";
 }
@@ -109,8 +113,14 @@ template <typename Dtype>
 inline void setConvolutionDesc(cudnnConvolutionDescriptor_t* conv,
     cudnnTensorDescriptor_t bottom, cudnnFilterDescriptor_t filter,
     int pad_h, int pad_w, int stride_h, int stride_w) {
+#if CUDNN_VERSION_MIN(6, 0, 0)
   CUDNN_CHECK(cudnnSetConvolution2dDescriptor(*conv,
+      pad_h, pad_w, stride_h, stride_w, 1, 1, CUDNN_CROSS_CORRELATION,
+      dataType<Dtype>::type));
+#else
+    CUDNN_CHECK(cudnnSetConvolution2dDescriptor(*conv,
       pad_h, pad_w, stride_h, stride_w, 1, 1, CUDNN_CROSS_CORRELATION));
+#endif
 }
 
 template <typename Dtype>
diff --git a/scripts/travis/install-deps.sh b/scripts/travis/install-deps.sh
index 1900b16df54..1593ed8b59a 100755
--- a/scripts/travis/install-deps.sh
+++ b/scripts/travis/install-deps.sh
@@ -104,7 +104,7 @@ if $WITH_CUDA ; then
   ln -s /usr/local/cuda-$CUDA_VERSION /usr/local/cuda
 
   if $WITH_CUDNN ; then
-    apt-get install -y --no-install-recommends libcudnn5-dev
+    apt-get install -y --no-install-recommends libcudnn6-dev
   fi
 fi
 

From 179dafdb1a930cf86ff0956618bf8411b8dcd90e Mon Sep 17 00:00:00 2001
From: Luke Yeager <lukeyeager@users.noreply.github.com>
Date: Fri, 31 Mar 2017 11:24:56 -0700
Subject: [PATCH 34/75] Add test for caffe.draw.draw_net()

---
 python/caffe/test/test_draw.py        | 33 +++++++++++++++++++++++++++++++++
 scripts/travis/install-deps.sh        |  2 ++
 scripts/travis/install-python-deps.sh |  1 +
 3 files changed, 36 insertions(+)
 create mode 100644 python/caffe/test/test_draw.py

diff --git a/python/caffe/test/test_draw.py b/python/caffe/test/test_draw.py
new file mode 100644
index 00000000000..1634145ee9d
--- /dev/null
+++ b/python/caffe/test/test_draw.py
@@ -0,0 +1,33 @@
+import os
+import unittest
+
+from google import protobuf
+
+import caffe.draw
+from caffe.proto import caffe_pb2
+
+def getFilenames():
+    """Yields files in the source tree which are Net prototxts."""
+    result = []
+
+    root_dir = os.path.abspath(os.path.join(
+        os.path.dirname(__file__), '..', '..', '..'))
+    assert os.path.exists(root_dir)
+
+    for dirname in ('models', 'examples'):
+        dirname = os.path.join(root_dir, dirname)
+        assert os.path.exists(dirname)
+        for cwd, _, filenames in os.walk(dirname):
+            for filename in filenames:
+                filename = os.path.join(cwd, filename)
+                if filename.endswith('.prototxt') and 'solver' not in filename:
+                    yield os.path.join(dirname, filename)
+
+
+class TestDraw(unittest.TestCase):
+    def test_draw_net(self):
+        for filename in getFilenames():
+            net = caffe_pb2.NetParameter()
+            with open(filename) as infile:
+                protobuf.text_format.Merge(infile.read(), net)
+            caffe.draw.draw_net(net, 'LR')
diff --git a/scripts/travis/install-deps.sh b/scripts/travis/install-deps.sh
index 1900b16df54..59a9163d5fc 100755
--- a/scripts/travis/install-deps.sh
+++ b/scripts/travis/install-deps.sh
@@ -8,6 +8,7 @@ source $BASEDIR/defaults.sh
 apt-get -y update
 apt-get install -y --no-install-recommends \
   build-essential \
+  graphviz \
   libboost-filesystem-dev \
   libboost-python-dev \
   libboost-system-dev \
@@ -31,6 +32,7 @@ if ! $WITH_PYTHON3 ; then
     python-dev \
     python-numpy \
     python-protobuf \
+    python-pydot \
     python-skimage
 else
   # Python3
diff --git a/scripts/travis/install-python-deps.sh b/scripts/travis/install-python-deps.sh
index eeec302791f..910d35a93be 100755
--- a/scripts/travis/install-python-deps.sh
+++ b/scripts/travis/install-python-deps.sh
@@ -11,4 +11,5 @@ if ! $WITH_PYTHON3 ; then
 else
   # Python3
   pip install --pre protobuf==3.0.0b3
+  pip install pydot
 fi

From 41e34c9061e9577c2b1dd56be65fd23ef26457fd Mon Sep 17 00:00:00 2001
From: Nitheesh <nitheeshas91@gmail.com>
Date: Tue, 4 Apr 2017 13:36:20 +0530
Subject: [PATCH 35/75] Minor fix for net drawing script

---
 python/caffe/draw.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/python/caffe/draw.py b/python/caffe/draw.py
index 9eecf6d7b46..8411a41d1d4 100644
--- a/python/caffe/draw.py
+++ b/python/caffe/draw.py
@@ -91,11 +91,11 @@ def get_layer_label(layer, rankdir):
                       separator,
                       layer.type,
                       separator,
-                      layer.convolution_param.kernel_size[0] if len(layer.convolution_param.kernel_size._values) else 1,
+                      layer.convolution_param.kernel_size[0] if len(layer.convolution_param.kernel_size) else 1,
                       separator,
-                      layer.convolution_param.stride[0] if len(layer.convolution_param.stride._values) else 1,
+                      layer.convolution_param.stride[0] if len(layer.convolution_param.stride) else 1,
                       separator,
-                      layer.convolution_param.pad[0] if len(layer.convolution_param.pad._values) else 0)
+                      layer.convolution_param.pad[0] if len(layer.convolution_param.pad) else 0)
     elif layer.type == 'Pooling':
         pooling_types_dict = get_pooling_types_dict()
         node_label = '"%s%s(%s %s)%skernel size: %d%sstride: %d%spad: %d"' %\

From 31bfe8fb498ea2e528da6463c9045b397992e028 Mon Sep 17 00:00:00 2001
From: Nitheesh <nitheeshas91@gmail.com>
Date: Tue, 4 Apr 2017 13:40:31 +0530
Subject: [PATCH 36/75] Add main() for draw_net unittest, fix import errors

---
 python/caffe/test/test_draw.py | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/python/caffe/test/test_draw.py b/python/caffe/test/test_draw.py
index 1634145ee9d..835bb5df010 100644
--- a/python/caffe/test/test_draw.py
+++ b/python/caffe/test/test_draw.py
@@ -1,7 +1,7 @@
 import os
 import unittest
 
-from google import protobuf
+from google.protobuf import text_format
 
 import caffe.draw
 from caffe.proto import caffe_pb2
@@ -29,5 +29,9 @@ def test_draw_net(self):
         for filename in getFilenames():
             net = caffe_pb2.NetParameter()
             with open(filename) as infile:
-                protobuf.text_format.Merge(infile.read(), net)
+                text_format.Merge(infile.read(), net)
             caffe.draw.draw_net(net, 'LR')
+
+
+if __name__ == "__main__":
+    unittest.main()

From 5f1ca848f8c9daa73f61f64413e15ab2cd6602e7 Mon Sep 17 00:00:00 2001
From: "Jonathan R. Williford" <jonathan@neural.vision>
Date: Wed, 5 Apr 2017 10:03:31 +0000
Subject: [PATCH 37/75] Add example and small blurb about sigmoid layer.

---
 docs/tutorial/layers/sigmoid.md | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/docs/tutorial/layers/sigmoid.md b/docs/tutorial/layers/sigmoid.md
index 505318352c9..f18ac4b84ec 100644
--- a/docs/tutorial/layers/sigmoid.md
+++ b/docs/tutorial/layers/sigmoid.md
@@ -9,6 +9,16 @@ title: Sigmoid Layer
 * Header: [`./include/caffe/layers/sigmoid_layer.hpp`](https://github.com/BVLC/caffe/blob/master/include/caffe/layers/sigmoid_layer.hpp)
 * CPU implementation: [`./src/caffe/layers/sigmoid_layer.cpp`](https://github.com/BVLC/caffe/blob/master/src/caffe/layers/sigmoid_layer.cpp)
 * CUDA GPU implementation: [`./src/caffe/layers/sigmoid_layer.cu`](https://github.com/BVLC/caffe/blob/master/src/caffe/layers/sigmoid_layer.cu)
+* Example (from [`./examples/mnist/mnist_autoencoder.prototxt`](https://github.com/BVLC/caffe/blob/master/examples/mnist/mnist_autoencoder.prototxt)):
+
+      layer {
+        name: "encode1neuron"
+        bottom: "encode1"
+        top: "encode1neuron"
+        type: "Sigmoid"
+      }
+
+The `Sigmoid` layer computes `sigmoid(x)` for each element `x` in the bottom blob.
 
 ## Parameters
 

From ce7193c7385298825c8cabebd20f664f3f93f06a Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Sat, 8 Apr 2017 12:59:24 -0400
Subject: [PATCH 38/75] Removed repeated import Layer, get_solver

---
 python/caffe/__init__.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/caffe/__init__.py b/python/caffe/__init__.py
index 43a0c49be63..80f51716d82 100644
--- a/python/caffe/__init__.py
+++ b/python/caffe/__init__.py
@@ -1,5 +1,5 @@
 from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer
-from ._caffe import init_log, log, set_mode_cpu, set_mode_gpu, set_device, Layer, get_solver, layer_type_list, set_random_seed, solver_count, set_solver_count, solver_rank, set_solver_rank, set_multiprocess, Layer, get_solver
+from ._caffe import init_log, log, set_mode_cpu, set_mode_gpu, set_device, Layer, get_solver, layer_type_list, set_random_seed, solver_count, set_solver_count, solver_rank, set_solver_rank, set_multiprocess
 from ._caffe import __version__
 from .proto.caffe_pb2 import TRAIN, TEST
 from .classifier import Classifier

From b2a95fa7fcba2089b981eb30b47d9aeba2b89ce9 Mon Sep 17 00:00:00 2001
From: Bruno Bowden <github@brunobowden.com>
Date: Sat, 8 Apr 2017 15:54:04 -0700
Subject: [PATCH 39/75] Log shape dimensions for eltwise layer shape mismatch

When layer shapes mismatch for the eltwise layer, caffe will fail a
check but doesn't give any information on how the shapes mismatch.
This logging information will make it easier to debug. Additionally
this reorders the variables to CHECK(expected == actual), matching
the JUnit convention.

BEFORE: Check failed: bottom[i]->shape() == bottom[0]->shape()

AFTER:  Check failed: bottom[0]->shape() == bottom[i]->shape()
        bottom[0]: 1 4 (4), bottom[3]: 1 6 (6)

NOTE: This removes use of CHECK_EQ in an earlier version of this PR,
which caused a build warning due to include of glog/stl_logging.h.
---
 src/caffe/layers/eltwise_layer.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/caffe/layers/eltwise_layer.cpp b/src/caffe/layers/eltwise_layer.cpp
index 21256166bfa..3d82b0e1cbf 100644
--- a/src/caffe/layers/eltwise_layer.cpp
+++ b/src/caffe/layers/eltwise_layer.cpp
@@ -31,7 +31,9 @@ template <typename Dtype>
 void EltwiseLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {
   for (int i = 1; i < bottom.size(); ++i) {
-    CHECK(bottom[i]->shape() == bottom[0]->shape());
+    CHECK(bottom[0]->shape() == bottom[i]->shape())
+        << "bottom[0]: " << bottom[0]->shape_string()
+        << ", bottom[" << i << "]: " << bottom[i]->shape_string();
   }
   top[0]->ReshapeLike(*bottom[0]);
   // If max operation, we will initialize the vector index part.

From 51728d1532dbee2853acb89a8a9653e82219953b Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Wed, 12 Apr 2017 01:42:59 -0700
Subject: [PATCH 40/75] Fix log parsing #5422

---
 tools/extra/parse_log.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tools/extra/parse_log.sh b/tools/extra/parse_log.sh
index 9892c897682..122eb9e6eed 100755
--- a/tools/extra/parse_log.sh
+++ b/tools/extra/parse_log.sh
@@ -39,7 +39,7 @@ rm aux.txt aux0.txt aux1.txt aux2.txt aux3.txt aux4.txt
 grep '] Solving ' $1 > aux.txt
 grep ', loss = ' $1 >> aux.txt
 grep 'Iteration ' aux.txt | sed  's/.*Iteration \([[:digit:]]*\).*/\1/g' > aux0.txt
-grep ', loss = ' $1 | awk '{print $9}' > aux1.txt
+grep ', loss = ' $1 | awk -F = '{print $2}' > aux1.txt
 grep ', lr = ' $1 | awk '{print $9}' > aux2.txt
 
 # Extracting elapsed seconds

From bac59bed485dfa195600b5b12031401613fade05 Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Wed, 12 Apr 2017 02:05:34 -0700
Subject: [PATCH 41/75] Allow using env vars for glog init from python

---
 python/caffe/_caffe.cpp | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/python/caffe/_caffe.cpp b/python/caffe/_caffe.cpp
index 276f21f85a5..01b34b84190 100644
--- a/python/caffe/_caffe.cpp
+++ b/python/caffe/_caffe.cpp
@@ -51,14 +51,18 @@ const int NPY_DTYPE = NPY_FLOAT32;
 void set_mode_cpu() { Caffe::set_mode(Caffe::CPU); }
 void set_mode_gpu() { Caffe::set_mode(Caffe::GPU); }
 
-void InitLog(int level) {
-  FLAGS_logtostderr = 1;
-  FLAGS_minloglevel = level;
+void InitLog() {
   ::google::InitGoogleLogging("");
   ::google::InstallFailureSignalHandler();
 }
-void InitLogInfo() {
-  InitLog(google::INFO);
+void InitLogLevel(int level) {
+  FLAGS_minloglevel = level;
+  InitLog();
+}
+void InitLogLevelPipe(int level, bool stderr) {
+  FLAGS_minloglevel = level;
+  FLAGS_logtostderr = stderr;
+  InitLog();
 }
 void Log(const string& s) {
   LOG(INFO) << s;
@@ -353,7 +357,8 @@ BOOST_PYTHON_MODULE(_caffe) {
 
   // Caffe utility functions
   bp::def("init_log", &InitLog);
-  bp::def("init_log", &InitLogInfo);
+  bp::def("init_log", &InitLogLevel);
+  bp::def("init_log", &InitLogLevelPipe);
   bp::def("log", &Log);
   bp::def("set_mode_cpu", &set_mode_cpu);
   bp::def("set_mode_gpu", &set_mode_gpu);

From 35a7b87ad87457291dfc79bf8a7e7cf7ef278cbb Mon Sep 17 00:00:00 2001
From: Noiredd <snowball91b@gmail.com>
Date: Wed, 12 Apr 2017 11:59:06 +0200
Subject: [PATCH 42/75] fixes pycaffe forward() and backward() behavior for
 nets whose layer names do not match respective tops

---
 python/caffe/pycaffe.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/python/caffe/pycaffe.py b/python/caffe/pycaffe.py
index 63606591bb4..4a7b5a24c46 100644
--- a/python/caffe/pycaffe.py
+++ b/python/caffe/pycaffe.py
@@ -113,7 +113,7 @@ def _Net_forward(self, blobs=None, start=None, end=None, **kwargs):
 
     if end is not None:
         end_ind = list(self._layer_names).index(end)
-        outputs = set([end] + blobs)
+        outputs = set(self.top_names[end] + blobs)
     else:
         end_ind = len(self.layers) - 1
         outputs = set(self.outputs + blobs)
@@ -161,7 +161,7 @@ def _Net_backward(self, diffs=None, start=None, end=None, **kwargs):
 
     if end is not None:
         end_ind = list(self._layer_names).index(end)
-        outputs = set([end] + diffs)
+        outputs = set(self.bottom_names[end] + diffs)
     else:
         end_ind = 0
         outputs = set(self.inputs + diffs)

From 3a987960d6a08b179eb6c0c526b27ab761ea2d6e Mon Sep 17 00:00:00 2001
From: Kang Kim <kangk@qti.qualcomm.com>
Date: Thu, 13 Apr 2017 15:23:26 +0900
Subject: [PATCH 43/75] remove redundant check in LSTMUnitLayer

---
 src/caffe/layers/lstm_unit_layer.cpp | 1 -
 1 file changed, 1 deletion(-)

diff --git a/src/caffe/layers/lstm_unit_layer.cpp b/src/caffe/layers/lstm_unit_layer.cpp
index 277c031ad15..d1ab59c4bd1 100644
--- a/src/caffe/layers/lstm_unit_layer.cpp
+++ b/src/caffe/layers/lstm_unit_layer.cpp
@@ -31,7 +31,6 @@ void LSTMUnitLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
     CHECK_EQ(num_instances, bottom[i]->shape(1));
   }
   hidden_dim_ = bottom[0]->shape(2);
-  CHECK_EQ(num_instances, bottom[1]->shape(1));
   CHECK_EQ(4 * hidden_dim_, bottom[1]->shape(2));
   top[0]->ReshapeLike(*bottom[0]);
   top[1]->ReshapeLike(*bottom[0]);

From 96870628698090813d92a9b1f8af9a8311469354 Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Thu, 13 Apr 2017 13:15:24 -0400
Subject: [PATCH 44/75] Bump boost version to 1.55 in CMake build

---
 cmake/Dependencies.cmake | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
index 02c81525bce..4a5bac471b4 100644
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -5,7 +5,7 @@ set(Caffe_DEFINITIONS "")
 set(Caffe_COMPILE_OPTIONS "")
 
 # ---[ Boost
-find_package(Boost 1.46 REQUIRED COMPONENTS system thread filesystem)
+find_package(Boost 1.55 REQUIRED COMPONENTS system thread filesystem)
 list(APPEND Caffe_INCLUDE_DIRS PUBLIC ${Boost_INCLUDE_DIRS})
 list(APPEND Caffe_LINKER_LIBS PUBLIC ${Boost_LIBRARIES})
 

From 0c9cc62379e4061b58b0dfa257d79c2ecaeb2be8 Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Sat, 11 Mar 2017 20:12:40 -0500
Subject: [PATCH 45/75] Added support for python 3 and NCCL

---
 python/caffe/__init__.py       |  2 +-
 python/caffe/_caffe.cpp        | 32 +++++++++++++++++++++++++++++++-
 python/caffe/test/test_nccl.py | 19 +++++++++++++++++++
 3 files changed, 51 insertions(+), 2 deletions(-)
 create mode 100644 python/caffe/test/test_nccl.py

diff --git a/python/caffe/__init__.py b/python/caffe/__init__.py
index 80f51716d82..776945eec88 100644
--- a/python/caffe/__init__.py
+++ b/python/caffe/__init__.py
@@ -1,5 +1,5 @@
 from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver, NCCL, Timer
-from ._caffe import init_log, log, set_mode_cpu, set_mode_gpu, set_device, Layer, get_solver, layer_type_list, set_random_seed, solver_count, set_solver_count, solver_rank, set_solver_rank, set_multiprocess
+from ._caffe import init_log, log, set_mode_cpu, set_mode_gpu, set_device, Layer, get_solver, layer_type_list, set_random_seed, solver_count, set_solver_count, solver_rank, set_solver_rank, set_multiprocess, has_nccl
 from ._caffe import __version__
 from .proto.caffe_pb2 import TRAIN, TEST
 from .classifier import Classifier
diff --git a/python/caffe/_caffe.cpp b/python/caffe/_caffe.cpp
index 01b34b84190..7fc06c08f73 100644
--- a/python/caffe/_caffe.cpp
+++ b/python/caffe/_caffe.cpp
@@ -347,6 +347,35 @@ class NCCL {
 };
 #endif
 
+bool HasNCCL() {
+#ifdef USE_NCCL
+  return true;
+#else
+  return false;
+#endif
+}
+
+#ifdef USE_NCCL
+bp::object NCCL_New_Uid() {
+  std::string uid = NCCL<Dtype>::new_uid();
+#if PY_MAJOR_VERSION >= 3
+  // Convert std::string to bytes so that Python does not
+  // try to decode the string using the current locale.
+
+  // Since boost 1.53 boost.python will convert str and bytes
+  // to std::string but will convert std::string to str. Here we
+  // force a bytes object to be returned. When this object
+  // is passed back to the NCCL constructor boost.python will
+  // correctly convert the bytes to std::string automatically
+  PyObject* py_uid = PyBytes_FromString(uid.c_str());
+  return bp::object(bp::handle<>(py_uid));
+#else
+  // automatic conversion is correct for python 2.
+  return uid;
+#endif
+}
+#endif
+
 BOOST_PYTHON_MEMBER_FUNCTION_OVERLOADS(SolveOverloads, Solve, 0, 1);
 
 BOOST_PYTHON_MODULE(_caffe) {
@@ -360,6 +389,7 @@ BOOST_PYTHON_MODULE(_caffe) {
   bp::def("init_log", &InitLogLevel);
   bp::def("init_log", &InitLogLevelPipe);
   bp::def("log", &Log);
+  bp::def("has_nccl", &HasNCCL);
   bp::def("set_mode_cpu", &set_mode_cpu);
   bp::def("set_mode_gpu", &set_mode_gpu);
   bp::def("set_random_seed", &set_random_seed);
@@ -518,7 +548,7 @@ BOOST_PYTHON_MODULE(_caffe) {
     boost::noncopyable>("NCCL",
                         bp::init<shared_ptr<Solver<Dtype> >, const string&>())
 #ifdef USE_NCCL
-    .def("new_uid", &NCCL<Dtype>::new_uid).staticmethod("new_uid")
+    .def("new_uid", NCCL_New_Uid).staticmethod("new_uid")
     .def("bcast", &NCCL<Dtype>::Broadcast)
 #endif
     /* NOLINT_NEXT_LINE(whitespace/semicolon) */
diff --git a/python/caffe/test/test_nccl.py b/python/caffe/test/test_nccl.py
new file mode 100644
index 00000000000..127a9337040
--- /dev/null
+++ b/python/caffe/test/test_nccl.py
@@ -0,0 +1,19 @@
+import sys
+import unittest
+
+import caffe
+
+
+class TestNCCL(unittest.TestCase):
+
+    def test_newuid(self):
+        """
+        Test that NCCL uids are of the proper type
+        according to python version
+        """
+        if caffe.has_nccl():
+            uid = caffe.NCCL.new_uid()
+            if sys.version_info.major >= 3:
+                self.assertTrue(isinstance(uid, bytes))
+            else:
+                self.assertTrue(isinstance(uid, str))

From e98023af4a570e3105486b661e4c4d1855c0dd79 Mon Sep 17 00:00:00 2001
From: Patrick Follmann <follmann@mvtec.com>
Date: Thu, 29 Dec 2016 14:37:21 +0100
Subject: [PATCH 46/75] Add GPU sqrt functions

---
 include/caffe/util/math_functions.hpp |  3 +++
 src/caffe/util/math_functions.cu      | 21 +++++++++++++++++++++
 2 files changed, 24 insertions(+)

diff --git a/include/caffe/util/math_functions.hpp b/include/caffe/util/math_functions.hpp
index 37abce5eccc..60a8404a044 100644
--- a/include/caffe/util/math_functions.hpp
+++ b/include/caffe/util/math_functions.hpp
@@ -214,6 +214,9 @@ void caffe_gpu_log(const int n, const Dtype* a, Dtype* y);
 template <typename Dtype>
 void caffe_gpu_powx(const int n, const Dtype* a, const Dtype b, Dtype* y);
 
+template <typename Dtype>
+void caffe_gpu_sqrt(const int n, const Dtype* a, Dtype* y);
+
 // caffe_gpu_rng_uniform with two arguments generates integers in the range
 // [0, UINT_MAX].
 void caffe_gpu_rng_uniform(const int n, unsigned int* r);
diff --git a/src/caffe/util/math_functions.cu b/src/caffe/util/math_functions.cu
index 6d001026082..314e6ba0f63 100644
--- a/src/caffe/util/math_functions.cu
+++ b/src/caffe/util/math_functions.cu
@@ -387,6 +387,27 @@ void caffe_gpu_powx<double>(const int N, const double* a,
       N, a, alpha, y);
 }
 
+template <typename Dtype>
+__global__ void sqrt_kernel(const int n, const Dtype* a, Dtype* y) {
+  CUDA_KERNEL_LOOP(index, n) {
+    y[index] = sqrt(a[index]);
+  }
+}
+
+template <>
+void caffe_gpu_sqrt<float>(const int N, const float* a, float* y) {
+  // NOLINT_NEXT_LINE(whitespace/operators)
+  sqrt_kernel<float><<<CAFFE_GET_BLOCKS(N), CAFFE_CUDA_NUM_THREADS>>>(
+      N, a, y);
+}
+
+template <>
+void caffe_gpu_sqrt<double>(const int N, const double* a, double* y) {
+  // NOLINT_NEXT_LINE(whitespace/operators)
+  sqrt_kernel<double><<<CAFFE_GET_BLOCKS(N), CAFFE_CUDA_NUM_THREADS>>>(
+      N, a, y);
+}
+
 DEFINE_AND_INSTANTIATE_GPU_UNARY_FUNC(sign, y[index] = (Dtype(0) < x[index])
                                       - (x[index] < Dtype(0)));
 DEFINE_AND_INSTANTIATE_GPU_UNARY_FUNC(sgnbit, y[index] = signbit(x[index]));

From e93fcd267582888f960ca48d6e0c2e719d4ea09b Mon Sep 17 00:00:00 2001
From: Patrick Follmann <follmann@mvtec.com>
Date: Thu, 29 Dec 2016 14:46:16 +0100
Subject: [PATCH 47/75] GPU BatchNormLayer: replace powx with mul and sqrt

---
 src/caffe/layers/batch_norm_layer.cu | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/caffe/layers/batch_norm_layer.cu b/src/caffe/layers/batch_norm_layer.cu
index c21713c81d9..a35e778e2f1 100644
--- a/src/caffe/layers/batch_norm_layer.cu
+++ b/src/caffe/layers/batch_norm_layer.cu
@@ -48,14 +48,14 @@ void BatchNormLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 
   if (!use_global_stats_) {
     // compute variance using var(X) = E((X-EX)^2)
-    caffe_gpu_powx(top[0]->count(), top_data, Dtype(2),
+    caffe_gpu_mul(top[0]->count(), top[0]->gpu_data(), top[0]->gpu_data(),
         temp_.mutable_gpu_data());  // (X-EX)^2
     caffe_gpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim,
         1. / (num * spatial_dim), temp_.gpu_data(),
         spatial_sum_multiplier_.gpu_data(), 0.,
         num_by_chans_.mutable_gpu_data());
-    caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, 1.,
-        num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), 0.,
+    caffe_gpu_gemv<Dtype>(CblasTrans, num, channels_, Dtype(1.),
+        num_by_chans_.gpu_data(), batch_sum_multiplier_.gpu_data(), Dtype(0.),
         variance_.mutable_gpu_data());  // E((X_EX)^2)
 
     // compute and save moving average
@@ -72,7 +72,7 @@ void BatchNormLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 
   // normalize variance
   caffe_gpu_add_scalar(variance_.count(), eps_, variance_.mutable_gpu_data());
-  caffe_gpu_powx(variance_.count(), variance_.gpu_data(), Dtype(0.5),
+  caffe_gpu_sqrt(variance_.count(), variance_.gpu_data(),
       variance_.mutable_gpu_data());
 
   // replicate variance to input size

From ab3398832964c1ff1bf6b78501e4e43a11f282a1 Mon Sep 17 00:00:00 2001
From: Jeff Donahue <jeff.donahue@gmail.com>
Date: Thu, 13 Apr 2017 13:25:16 -0700
Subject: [PATCH 48/75] Add CPU sqrt functions

---
 include/caffe/util/math_functions.hpp |  3 +++
 src/caffe/util/math_functions.cpp     | 10 ++++++++++
 2 files changed, 13 insertions(+)

diff --git a/include/caffe/util/math_functions.hpp b/include/caffe/util/math_functions.hpp
index 60a8404a044..e549120a933 100644
--- a/include/caffe/util/math_functions.hpp
+++ b/include/caffe/util/math_functions.hpp
@@ -53,6 +53,9 @@ template <typename Dtype>
 void caffe_sqr(const int N, const Dtype* a, Dtype* y);
 
 template <typename Dtype>
+void caffe_sqrt(const int N, const Dtype* a, Dtype* y);
+
+template <typename Dtype>
 void caffe_add(const int N, const Dtype* a, const Dtype* b, Dtype* y);
 
 template <typename Dtype>
diff --git a/src/caffe/util/math_functions.cpp b/src/caffe/util/math_functions.cpp
index 71c02274a75..59625bc05ce 100644
--- a/src/caffe/util/math_functions.cpp
+++ b/src/caffe/util/math_functions.cpp
@@ -197,6 +197,16 @@ void caffe_sqr<double>(const int n, const double* a, double* y) {
 }
 
 template <>
+void caffe_sqrt<float>(const int n, const float* a, float* y) {
+  vsSqrt(n, a, y);
+}
+
+template <>
+void caffe_sqrt<double>(const int n, const double* a, double* y) {
+  vdSqrt(n, a, y);
+}
+
+template <>
 void caffe_exp<float>(const int n, const float* a, float* y) {
   vsExp(n, a, y);
 }

From 1c15d94f7da736945450e6ed321077f3045445b1 Mon Sep 17 00:00:00 2001
From: Jeff Donahue <jeff.donahue@gmail.com>
Date: Thu, 13 Apr 2017 13:26:16 -0700
Subject: [PATCH 49/75] CPU BatchNormLayer: replace powx with sqr and sqrt

---
 src/caffe/layers/batch_norm_layer.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/caffe/layers/batch_norm_layer.cpp b/src/caffe/layers/batch_norm_layer.cpp
index 0a08ed4cb07..c6a1d5b1b2c 100644
--- a/src/caffe/layers/batch_norm_layer.cpp
+++ b/src/caffe/layers/batch_norm_layer.cpp
@@ -124,8 +124,8 @@ void BatchNormLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
 
   if (!use_global_stats_) {
     // compute variance using var(X) = E((X-EX)^2)
-    caffe_powx(top[0]->count(), top_data, Dtype(2),
-        temp_.mutable_cpu_data());  // (X-EX)^2
+    caffe_sqr<Dtype>(top[0]->count(), top_data,
+                     temp_.mutable_cpu_data());  // (X-EX)^2
     caffe_cpu_gemv<Dtype>(CblasNoTrans, channels_ * num, spatial_dim,
         1. / (num * spatial_dim), temp_.cpu_data(),
         spatial_sum_multiplier_.cpu_data(), 0.,
@@ -148,7 +148,7 @@ void BatchNormLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
 
   // normalize variance
   caffe_add_scalar(variance_.count(), eps_, variance_.mutable_cpu_data());
-  caffe_powx(variance_.count(), variance_.cpu_data(), Dtype(0.5),
+  caffe_sqrt(variance_.count(), variance_.cpu_data(),
              variance_.mutable_cpu_data());
 
   // replicate variance to input size

From 3d5bed06a9b6b8a5dfd3db8da33f2fa3bc9a1213 Mon Sep 17 00:00:00 2001
From: Jeff Donahue <jeff.donahue@gmail.com>
Date: Thu, 13 Apr 2017 14:15:16 -0700
Subject: [PATCH 50/75] fix: add non-MKL sqrt (should have been included in
 ab33988)

---
 include/caffe/util/mkl_alternate.hpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/caffe/util/mkl_alternate.hpp b/include/caffe/util/mkl_alternate.hpp
index 79b2c32de94..8c2294c7c86 100644
--- a/include/caffe/util/mkl_alternate.hpp
+++ b/include/caffe/util/mkl_alternate.hpp
@@ -37,6 +37,7 @@ extern "C" {
   }
 
 DEFINE_VSL_UNARY_FUNC(Sqr, y[i] = a[i] * a[i])
+DEFINE_VSL_UNARY_FUNC(Sqrt, y[i] = sqrt(a[i]))
 DEFINE_VSL_UNARY_FUNC(Exp, y[i] = exp(a[i]))
 DEFINE_VSL_UNARY_FUNC(Ln, y[i] = log(a[i]))
 DEFINE_VSL_UNARY_FUNC(Abs, y[i] = fabs(a[i]))

From 2ec19b6177111526d2df362d29d0e08aa5645a22 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Thu, 13 Apr 2017 14:22:30 -0700
Subject: [PATCH 51/75] deprecate WindowData layer type

---
 include/caffe/layers/window_data_layer.hpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/caffe/layers/window_data_layer.hpp b/include/caffe/layers/window_data_layer.hpp
index 35f41b80e63..b9b66b7cf1d 100644
--- a/include/caffe/layers/window_data_layer.hpp
+++ b/include/caffe/layers/window_data_layer.hpp
@@ -16,7 +16,8 @@ namespace caffe {
 
 /**
  * @brief Provides data to the Net from windows of images files, specified
- *        by a window data file.
+ *        by a window data file. This layer is *DEPRECATED* and only kept for
+ *        archival purposes for use by the original R-CNN.
  *
  * TODO(dox): thorough documentation for Forward and proto params.
  */

From e7163f650885b9f7b9cae1c3253aa97d9fe30d86 Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Thu, 13 Apr 2017 20:32:40 -0400
Subject: [PATCH 52/75] Updated Travis boost dependencies

---
 scripts/travis/install-deps.sh | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/scripts/travis/install-deps.sh b/scripts/travis/install-deps.sh
index 2fa2a74a486..dac5d2f9d37 100755
--- a/scripts/travis/install-deps.sh
+++ b/scripts/travis/install-deps.sh
@@ -9,10 +9,10 @@ apt-get -y update
 apt-get install -y --no-install-recommends \
   build-essential \
   graphviz \
-  libboost-filesystem-dev \
-  libboost-python-dev \
-  libboost-system-dev \
-  libboost-thread-dev \
+  libboost-filesystem1.55-dev \
+  libboost-python1.55-dev \
+  libboost-system1.55-dev \
+  libboost-thread1.55-dev \
   libgflags-dev \
   libgoogle-glog-dev \
   libhdf5-serial-dev \

From 8bc82c635914676d51ecd2849cc69f6fb6042496 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Thu, 13 Apr 2017 19:14:57 -0700
Subject: [PATCH 53/75] [examples] switch cifar-10 back to proto instead of h5
 serialization

(it's more common)
---
 examples/cifar10/cifar10_quick_solver.prototxt | 1 -
 examples/cifar10/train_full.sh                 | 4 ++--
 examples/cifar10/train_quick.sh                | 2 +-
 3 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/examples/cifar10/cifar10_quick_solver.prototxt b/examples/cifar10/cifar10_quick_solver.prototxt
index 5de276f722f..14b4401ba16 100644
--- a/examples/cifar10/cifar10_quick_solver.prototxt
+++ b/examples/cifar10/cifar10_quick_solver.prototxt
@@ -20,7 +20,6 @@ display: 100
 max_iter: 4000
 # snapshot intermediate results
 snapshot: 4000
-snapshot_format: HDF5
 snapshot_prefix: "examples/cifar10/cifar10_quick"
 # solver mode: CPU or GPU
 solver_mode: GPU
diff --git a/examples/cifar10/train_full.sh b/examples/cifar10/train_full.sh
index 06ecc2dccb0..fe46e60d795 100755
--- a/examples/cifar10/train_full.sh
+++ b/examples/cifar10/train_full.sh
@@ -9,9 +9,9 @@ $TOOLS/caffe train \
 # reduce learning rate by factor of 10
 $TOOLS/caffe train \
     --solver=examples/cifar10/cifar10_full_solver_lr1.prototxt \
-    --snapshot=examples/cifar10/cifar10_full_iter_60000.solverstate.h5 $@
+    --snapshot=examples/cifar10/cifar10_full_iter_60000.solverstate $@
 
 # reduce learning rate by factor of 10
 $TOOLS/caffe train \
     --solver=examples/cifar10/cifar10_full_solver_lr2.prototxt \
-    --snapshot=examples/cifar10/cifar10_full_iter_65000.solverstate.h5 $@
+    --snapshot=examples/cifar10/cifar10_full_iter_65000.solverstate $@
diff --git a/examples/cifar10/train_quick.sh b/examples/cifar10/train_quick.sh
index d2b875340ee..257479e0d77 100755
--- a/examples/cifar10/train_quick.sh
+++ b/examples/cifar10/train_quick.sh
@@ -9,4 +9,4 @@ $TOOLS/caffe train \
 # reduce learning rate by factor of 10 after 8 epochs
 $TOOLS/caffe train \
   --solver=examples/cifar10/cifar10_quick_solver_lr1.prototxt \
-  --snapshot=examples/cifar10/cifar10_quick_iter_4000.solverstate.h5 $@
+  --snapshot=examples/cifar10/cifar10_quick_iter_4000.solverstate $@

From aa29eba26b781349174cb856b6ea96360ebbb3f2 Mon Sep 17 00:00:00 2001
From: Guillaume Dumont <dumont.guillaume@gmail.com>
Date: Thu, 13 Apr 2017 22:37:13 -0400
Subject: [PATCH 54/75] Explicit std::string to bp::object conversion

---
 python/caffe/_caffe.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/caffe/_caffe.cpp b/python/caffe/_caffe.cpp
index 7fc06c08f73..d7f43fff62d 100644
--- a/python/caffe/_caffe.cpp
+++ b/python/caffe/_caffe.cpp
@@ -371,7 +371,7 @@ bp::object NCCL_New_Uid() {
   return bp::object(bp::handle<>(py_uid));
 #else
   // automatic conversion is correct for python 2.
-  return uid;
+  return bp::object(uid);
 #endif
 }
 #endif

From c19c9602d031274ce77eb6a94ce2a9e8d843d98f Mon Sep 17 00:00:00 2001
From: Carl Doersch <cdoersch@cs.cmu.edu>
Date: Tue, 25 Aug 2015 11:26:14 -0700
Subject: [PATCH 55/75] Test for python forward and backward with start and end
 layer.

---
 python/caffe/test/test_net.py | 45 +++++++++++++++++++++++++++++++++++++++----
 1 file changed, 41 insertions(+), 4 deletions(-)

diff --git a/python/caffe/test/test_net.py b/python/caffe/test/test_net.py
index 24391cc50c4..afd27690981 100644
--- a/python/caffe/test/test_net.py
+++ b/python/caffe/test/test_net.py
@@ -25,11 +25,11 @@ def simple_net_file(num_output):
         bias_filler { type: 'constant' value: 2 } }
         param { decay_mult: 1 } param { decay_mult: 0 }
         }
-    layer { type: 'InnerProduct' name: 'ip' bottom: 'conv' top: 'ip'
+    layer { type: 'InnerProduct' name: 'ip' bottom: 'conv' top: 'ip_blob'
       inner_product_param { num_output: """ + str(num_output) + """
         weight_filler { type: 'gaussian' std: 2.5 }
         bias_filler { type: 'constant' value: -3 } } }
-    layer { type: 'SoftmaxWithLoss' name: 'loss' bottom: 'ip' bottom: 'label'
+    layer { type: 'SoftmaxWithLoss' name: 'loss' bottom: 'ip_blob' bottom: 'label'
       top: 'loss' }""")
     f.close()
     return f.name
@@ -71,6 +71,43 @@ def test_forward_backward(self):
         self.net.forward()
         self.net.backward()
 
+    def test_forward_start_end(self):
+        conv_blob=self.net.blobs['conv'];
+        ip_blob=self.net.blobs['ip_blob'];
+        sample_data=np.random.uniform(size=conv_blob.data.shape);
+        sample_data=sample_data.astype(np.float32);
+        conv_blob.data[:]=sample_data;
+        forward_blob=self.net.forward(start='ip',end='ip');
+        self.assertIn('ip_blob',forward_blob);
+
+        manual_forward=[];
+        for i in range(0,conv_blob.data.shape[0]):
+          dot=np.dot(self.net.params['ip'][0].data,
+                     conv_blob.data[i].reshape(-1));
+          manual_forward.append(dot+self.net.params['ip'][1].data);
+        manual_forward=np.array(manual_forward);
+
+        np.testing.assert_allclose(ip_blob.data,manual_forward,rtol=1e-3);
+
+    def test_backward_start_end(self):
+        conv_blob=self.net.blobs['conv'];
+        ip_blob=self.net.blobs['ip_blob'];
+        sample_data=np.random.uniform(size=ip_blob.data.shape)
+        sample_data=sample_data.astype(np.float32);
+        ip_blob.diff[:]=sample_data;
+        backward_blob=self.net.backward(start='ip',end='ip');
+        self.assertIn('conv',backward_blob);
+
+        manual_backward=[];
+        for i in range(0,conv_blob.data.shape[0]):
+          dot=np.dot(self.net.params['ip'][0].data.transpose(),
+                     sample_data[i].reshape(-1));
+          manual_backward.append(dot);
+        manual_backward=np.array(manual_backward);
+        manual_backward=manual_backward.reshape(conv_blob.data.shape);
+
+        np.testing.assert_allclose(conv_blob.diff,manual_backward,rtol=1e-3);
+
     def test_clear_param_diffs(self):
         # Run a forward/backward step to have non-zero diffs
         self.net.forward()
@@ -90,13 +127,13 @@ def test_top_bottom_names(self):
         self.assertEqual(self.net.top_names,
                          OrderedDict([('data', ['data', 'label']),
                                       ('conv', ['conv']),
-                                      ('ip', ['ip']),
+                                      ('ip', ['ip_blob']),
                                       ('loss', ['loss'])]))
         self.assertEqual(self.net.bottom_names,
                          OrderedDict([('data', []),
                                       ('conv', ['data']),
                                       ('ip', ['conv']),
-                                      ('loss', ['ip', 'label'])]))
+                                      ('loss', ['ip_blob', 'label'])]))
 
     def test_save_and_read(self):
         f = tempfile.NamedTemporaryFile(mode='w+', delete=False)

From 451944333510e1ea9b0bdac11e4ec201e5284714 Mon Sep 17 00:00:00 2001
From: jgyllinsky <jgyllinsky@users.noreply.github.com>
Date: Fri, 14 Apr 2017 03:11:59 -0400
Subject: [PATCH 56/75] [docs] added apt command to install OpenBLAS (#4718)

---
 docs/install_apt.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/install_apt.md b/docs/install_apt.md
index bc1566b0be9..ee2cd287701 100644
--- a/docs/install_apt.md
+++ b/docs/install_apt.md
@@ -14,7 +14,7 @@ The NVIDIA package tends to follow more recent library and driver versions, but
 If installing from packages, install the library and latest driver separately; the driver bundled with the library is usually out-of-date.
 This can be skipped for CPU-only installation.
 
-**BLAS**: install ATLAS by `sudo apt-get install libatlas-base-dev` or install OpenBLAS or MKL for better CPU performance.
+**BLAS**: install ATLAS by `sudo apt-get install libatlas-base-dev` or install OpenBLAS by `sudo apt-get install libopenblas-dev` or MKL for better CPU performance.
 
 **Python** (optional): if you use the default Python you will need to `sudo apt-get install` the `python-dev` package to have the Python headers for building the pycaffe interface.
 

From 80073497045d3101492a28a8a2c87dff65d64ff4 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 10:17:52 -0700
Subject: [PATCH 57/75] fix lint errors that snuck in by #4566

---
 src/caffe/test/test_gradient_based_solver.cpp | 12 ++++++++----
 src/caffe/test/test_neuron_layer.cpp          |  9 ++++++---
 2 files changed, 14 insertions(+), 7 deletions(-)

diff --git a/src/caffe/test/test_gradient_based_solver.cpp b/src/caffe/test/test_gradient_based_solver.cpp
index 05cab909798..f4395f5311c 100644
--- a/src/caffe/test/test_gradient_based_solver.cpp
+++ b/src/caffe/test/test_gradient_based_solver.cpp
@@ -558,9 +558,11 @@ class GradientBasedSolverTest : public MultiDeviceTest<TypeParam> {
     const vector<Blob<Dtype>*>& params = solver_->net()->learnable_params();
     for (int i = 0; i < params.size(); ++i) {
       for (int j = 0; j < params[i]->count(); ++j) {
-        EXPECT_FLOAT_EQ(param_copies[i]->cpu_data()[j], params[i]->cpu_data()[j])
+        EXPECT_FLOAT_EQ(param_copies[i]->cpu_data()[j],
+            params[i]->cpu_data()[j])
             << "param " << i << " data differed at dim " << j;
-        EXPECT_FLOAT_EQ(param_copies[i]->cpu_diff()[j], params[i]->cpu_diff()[j])
+        EXPECT_FLOAT_EQ(param_copies[i]->cpu_diff()[j],
+            params[i]->cpu_diff()[j])
             << "param " << i << " diff differed at dim " << j;
       }
     }
@@ -569,9 +571,11 @@ class GradientBasedSolverTest : public MultiDeviceTest<TypeParam> {
     const vector<shared_ptr<Blob<Dtype> > >& history = solver_->history();
     for (int i = 0; i < history.size(); ++i) {
       for (int j = 0; j < history[i]->count(); ++j) {
-        EXPECT_FLOAT_EQ(history_copies[i]->cpu_data()[j], history[i]->cpu_data()[j])
+        EXPECT_FLOAT_EQ(history_copies[i]->cpu_data()[j],
+            history[i]->cpu_data()[j])
             << "history blob " << i << " data differed at dim " << j;
-        EXPECT_FLOAT_EQ(history_copies[i]->cpu_diff()[j], history[i]->cpu_diff()[j])
+        EXPECT_FLOAT_EQ(history_copies[i]->cpu_diff()[j],
+            history[i]->cpu_diff()[j])
             << "history blob " << i << " diff differed at dim " << j;
       }
     }
diff --git a/src/caffe/test/test_neuron_layer.cpp b/src/caffe/test/test_neuron_layer.cpp
index 57bd47b3a2e..180871a29ee 100644
--- a/src/caffe/test/test_neuron_layer.cpp
+++ b/src/caffe/test/test_neuron_layer.cpp
@@ -791,13 +791,16 @@ TYPED_TEST(NeuronLayerTest, TestPReLUInPlace) {
   ip2.Backward(blob_middle_vec_2, propagate_down, blob_bottom_vec_2);
   // Check numbers
   for (int s = 0; s < blob_bottom_2->count(); ++s) {
-    EXPECT_FLOAT_EQ(this->blob_bottom_->cpu_diff()[s], blob_bottom_2->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(this->blob_bottom_->cpu_diff()[s],
+        blob_bottom_2->cpu_diff()[s]);
   }
   for (int s = 0; s < ip.blobs()[0]->count(); ++s) {
-    EXPECT_FLOAT_EQ(ip.blobs()[0]->cpu_diff()[s], ip2.blobs()[0]->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(ip.blobs()[0]->cpu_diff()[s],
+        ip2.blobs()[0]->cpu_diff()[s]);
   }
   for (int s = 0; s < ip.blobs()[1]->count(); ++s) {
-    EXPECT_FLOAT_EQ(ip.blobs()[1]->cpu_diff()[s], ip2.blobs()[1]->cpu_diff()[s]);
+    EXPECT_FLOAT_EQ(ip.blobs()[1]->cpu_diff()[s],
+        ip2.blobs()[1]->cpu_diff()[s]);
   }
   for (int s = 0; s < prelu.blobs()[0]->count(); ++s) {
     EXPECT_FLOAT_EQ(prelu.blobs()[0]->cpu_diff()[s],

From 4db619aec9cd384b11a1c55fac257d14b704bb15 Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Fri, 14 Apr 2017 12:30:50 -0700
Subject: [PATCH 58/75] Docker update to cuDNN 6

---
 docker/cpu/Dockerfile | 3 ++-
 docker/gpu/Dockerfile | 5 +++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/docker/cpu/Dockerfile b/docker/cpu/Dockerfile
index af6c03c6589..67e2e61bd57 100644
--- a/docker/cpu/Dockerfile
+++ b/docker/cpu/Dockerfile
@@ -28,7 +28,8 @@ ENV CAFFE_ROOT=/opt/caffe
 WORKDIR $CAFFE_ROOT
 
 # FIXME: use ARG instead of ENV once DockerHub supports this
-ENV CLONE_TAG=rc4
+# https://github.com/docker/hub-feedback/issues/460
+ENV CLONE_TAG=1.0
 
 RUN git clone -b ${CLONE_TAG} --depth 1 https://github.com/BVLC/caffe.git . && \
     pip install --upgrade pip && \
diff --git a/docker/gpu/Dockerfile b/docker/gpu/Dockerfile
index 0785b10f1e7..dcdbdf326fb 100644
--- a/docker/gpu/Dockerfile
+++ b/docker/gpu/Dockerfile
@@ -1,4 +1,4 @@
-FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04
+FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04
 LABEL maintainer caffe-maint@googlegroups.com
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -28,7 +28,8 @@ ENV CAFFE_ROOT=/opt/caffe
 WORKDIR $CAFFE_ROOT
 
 # FIXME: use ARG instead of ENV once DockerHub supports this
-ENV CLONE_TAG=rc4
+# https://github.com/docker/hub-feedback/issues/460
+ENV CLONE_TAG=1.0
 
 RUN git clone -b ${CLONE_TAG} --depth 1 https://github.com/BVLC/caffe.git . && \
     pip install --upgrade pip && \

From 44da39f662a24de746fa83b92bd670fe41b3a7da Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 12:36:41 -0700
Subject: [PATCH 59/75] BVLC -> BAIR

Berkeley AI Research (BAIR) is the the successor to
the Berkeley Vision and Learning Center (BVLC).
---
 CONTRIBUTORS.md                               |  2 +-
 README.md                                     |  6 +++---
 docs/_layouts/default.html                    |  2 +-
 docs/development.md                           |  4 ++--
 docs/index.md                                 | 10 +++++-----
 docs/model_zoo.md                             | 18 +++++++++---------
 docs/multigpu.md                              |  4 ++--
 docs/performance_hardware.md                  |  2 +-
 docs/tutorial/interfaces.md                   |  4 ++--
 examples/finetune_flickr_style/readme.md      |  2 +-
 models/bvlc_alexnet/readme.md                 |  2 +-
 models/bvlc_googlenet/readme.md               |  2 +-
 models/bvlc_reference_caffenet/readme.md      |  2 +-
 models/bvlc_reference_rcnn_ilsvrc13/readme.md |  2 +-
 14 files changed, 31 insertions(+), 31 deletions(-)

diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 8db66ea82c6..3fd767812e9 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -1,6 +1,6 @@
 # Contributors
 
-Caffe is developed by a core set of BVLC members and the open-source community.
+Caffe is developed by a core set of BAIR members and the open-source community.
 
 We thank all of our [contributors](https://github.com/BVLC/caffe/graphs/contributors)!
 
diff --git a/README.md b/README.md
index 44b9e62c157..0ae3616b4a6 100644
--- a/README.md
+++ b/README.md
@@ -4,13 +4,13 @@
 [![License](https://img.shields.io/badge/license-BSD-blue.svg)](LICENSE)
 
 Caffe is a deep learning framework made with expression, speed, and modularity in mind.
-It is developed by the Berkeley Vision and Learning Center ([BVLC](http://bvlc.eecs.berkeley.edu)) and community contributors.
+It is developed by Berkeley AI Research ([BAIR](http://bair.berkeley.edu))/The Berkeley Vision and Learning Center (BVLC) and community contributors.
 
 Check out the [project site](http://caffe.berkeleyvision.org) for all the details like
 
 - [DIY Deep Learning for Vision with Caffe](https://docs.google.com/presentation/d/1UeKXVgRvvxg9OUdh_UiC5G71UMscNPlvArsWER41PsU/edit#slide=id.p)
 - [Tutorial Documentation](http://caffe.berkeleyvision.org/tutorial/)
-- [BVLC reference models](http://caffe.berkeleyvision.org/model_zoo.html) and the [community model zoo](https://github.com/BVLC/caffe/wiki/Model-Zoo)
+- [BAIR reference models](http://caffe.berkeleyvision.org/model_zoo.html) and the [community model zoo](https://github.com/BVLC/caffe/wiki/Model-Zoo)
 - [Installation instructions](http://caffe.berkeleyvision.org/installation.html)
 
 and step-by-step examples.
@@ -25,7 +25,7 @@ Happy brewing!
 ## License and Citation
 
 Caffe is released under the [BSD 2-Clause license](https://github.com/BVLC/caffe/blob/master/LICENSE).
-The BVLC reference models are released for unrestricted use.
+The BAIR/BVLC reference models are released for unrestricted use.
 
 Please cite Caffe in your publications if it helps your research:
 
diff --git a/docs/_layouts/default.html b/docs/_layouts/default.html
index b8efe60bc3b..3799e95afde 100644
--- a/docs/_layouts/default.html
+++ b/docs/_layouts/default.html
@@ -36,7 +36,7 @@
       <header>
         <h1 class="header"><a href="/">Caffe</a></h1>
         <p class="header">
-          Deep learning framework by the <a class="header name" href="http://bvlc.eecs.berkeley.edu/">BVLC</a>
+          Deep learning framework by <a class="header name" href="http://bair.berkeley.edu/">BAIR</a>
         </p>
         <p class="header">
           Created by
diff --git a/docs/development.md b/docs/development.md
index 107c2c3b281..ec05bbee102 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -4,7 +4,7 @@ title: Developing and Contributing
 # Development and Contributing
 
 Caffe is developed with active participation of the community.<br>
-The [BVLC](http://bvlc.eecs.berkeley.edu/) brewers welcome all contributions!
+The [BAIR](http://bair.berkeley.edu/)/BVLC brewers welcome all contributions!
 
 The exact details of contributions are recorded by versioning and cited in our [acknowledgements](http://caffe.berkeleyvision.org/#acknowledgements).
 This method is impartial and always up-to-date.
@@ -37,7 +37,7 @@ We absolutely appreciate any contribution to this effort!
 
 The `master` branch receives all new development including community contributions.
 We try to keep it in a reliable state, but it is the bleeding edge, and things do get broken every now and then.
-BVLC maintainers will periodically make releases by marking stable checkpoints as tags and maintenance branches. [Past releases](https://github.com/BVLC/caffe/releases) are catalogued online.
+BAIR maintainers will periodically make releases by marking stable checkpoints as tags and maintenance branches. [Past releases](https://github.com/BVLC/caffe/releases) are catalogued online.
 
 #### Issues & Pull Request Protocol
 
diff --git a/docs/index.md b/docs/index.md
index 932b3b58d1d..302a7d56f88 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -5,7 +5,7 @@ title: Deep Learning Framework
 # Caffe
 
 Caffe is a deep learning framework made with expression, speed, and modularity in mind.
-It is developed by the Berkeley Vision and Learning Center ([BVLC](http://bvlc.eecs.berkeley.edu)) and by community contributors.
+It is developed by Berkeley AI Research ([BAIR](http://bair.berkeley.edu)) and by community contributors.
 [Yangqing Jia](http://daggerfs.com) created the project during his PhD at UC Berkeley.
 Caffe is released under the [BSD 2-Clause license](https://github.com/BVLC/caffe/blob/master/LICENSE).
 
@@ -45,7 +45,7 @@ A 4-page report for the ACM Multimedia Open Source competition (arXiv:1408.5093v
 - [Installation instructions](/installation.html)<br>
 Tested on Ubuntu, Red Hat, OS X.
 * [Model Zoo](/model_zoo.html)<br>
-BVLC suggests a standard distribution format for Caffe models, and provides trained models.
+BAIR suggests a standard distribution format for Caffe models, and provides trained models.
 * [Developing & Contributing](/development.html)<br>
 Guidelines for development and contributing to Caffe.
 * [API Documentation](/doxygen/annotated.html)<br>
@@ -92,9 +92,9 @@ The core Caffe developers offer [consulting services](mailto:caffe-coldpress@goo
 
 ## Acknowledgements
 
-The BVLC Caffe developers would like to thank NVIDIA for GPU donation, A9 and Amazon Web Services for a research grant in support of Caffe development and reproducible research in deep learning, and BVLC PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for guidance.
+The BAIR Caffe developers would like to thank NVIDIA for GPU donation, A9 and Amazon Web Services for a research grant in support of Caffe development and reproducible research in deep learning, and BAIR PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for guidance.
 
-The BVLC members who have contributed to Caffe are (alphabetical by first name):
+The BAIR members who have contributed to Caffe are (alphabetical by first name):
 [Eric Tzeng](https://github.com/erictzeng), [Evan Shelhamer](http://imaginarynumber.net/), [Jeff Donahue](http://jeffdonahue.com/), [Jon Long](https://github.com/longjon), [Ross Girshick](http://www.cs.berkeley.edu/~rbg/), [Sergey Karayev](http://sergeykarayev.com/), [Sergio Guadarrama](http://www.eecs.berkeley.edu/~sguada/), and [Yangqing Jia](http://daggerfs.com/).
 
 The open-source community plays an important and growing role in Caffe's development.
@@ -103,4 +103,4 @@ Check out the Github [project pulse](https://github.com/BVLC/caffe/pulse) for re
 We sincerely appreciate your interest and contributions!
 If you'd like to contribute, please read the [developing & contributing](development.html) guide.
 
-Yangqing would like to give a personal thanks to the NVIDIA Academic program for providing GPUs, [Oriol Vinyals](http://www1.icsi.berkeley.edu/~vinyals/) for discussions along the journey, and BVLC PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for advice.
+Yangqing would like to give a personal thanks to the NVIDIA Academic program for providing GPUs, [Oriol Vinyals](http://www1.icsi.berkeley.edu/~vinyals/) for discussions along the journey, and BAIR PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for advice.
diff --git a/docs/model_zoo.md b/docs/model_zoo.md
index 06dc0a49ec7..f9078718a8b 100644
--- a/docs/model_zoo.md
+++ b/docs/model_zoo.md
@@ -14,15 +14,15 @@ To help share these models, we introduce the model zoo framework:
 
 ## Where to get trained models
 
-First of all, we bundle BVLC-trained models for unrestricted, out of the box use.
+First of all, we bundle BAIR-trained models for unrestricted, out of the box use.
 <br>
-See the [BVLC model license](#bvlc-model-license) for details.
+See the [BAIR model license](#bair-model-license) for details.
 Each one of these can be downloaded by running `scripts/download_model_binary.py <dirname>` where `<dirname>` is specified below:
 
-- **BVLC Reference CaffeNet** in `models/bvlc_reference_caffenet`: AlexNet trained on ILSVRC 2012, with a minor variation from the version as described in [ImageNet classification with deep convolutional neural networks](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks) by Krizhevsky et al. in NIPS 2012. (Trained by Jeff Donahue @jeffdonahue)
-- **BVLC AlexNet** in `models/bvlc_alexnet`: AlexNet trained on ILSVRC 2012, almost exactly as described in [ImageNet classification with deep convolutional neural networks](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks) by Krizhevsky et al. in NIPS 2012. (Trained by Evan Shelhamer @shelhamer)
-- **BVLC Reference R-CNN ILSVRC-2013** in `models/bvlc_reference_rcnn_ilsvrc13`: pure Caffe implementation of [R-CNN](https://github.com/rbgirshick/rcnn) as described by Girshick et al. in CVPR 2014. (Trained by Ross Girshick @rbgirshick)
-- **BVLC GoogLeNet** in `models/bvlc_googlenet`: GoogLeNet trained on ILSVRC 2012, almost exactly as described in [Going Deeper with Convolutions](http://arxiv.org/abs/1409.4842) by Szegedy et al. in ILSVRC 2014. (Trained by Sergio Guadarrama @sguada)
+- **BAIR Reference CaffeNet** in `models/bvlc_reference_caffenet`: AlexNet trained on ILSVRC 2012, with a minor variation from the version as described in [ImageNet classification with deep convolutional neural networks](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks) by Krizhevsky et al. in NIPS 2012. (Trained by Jeff Donahue @jeffdonahue)
+- **BAIR AlexNet** in `models/bvlc_alexnet`: AlexNet trained on ILSVRC 2012, almost exactly as described in [ImageNet classification with deep convolutional neural networks](http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks) by Krizhevsky et al. in NIPS 2012. (Trained by Evan Shelhamer @shelhamer)
+- **BAIR Reference R-CNN ILSVRC-2013** in `models/bvlc_reference_rcnn_ilsvrc13`: pure Caffe implementation of [R-CNN](https://github.com/rbgirshick/rcnn) as described by Girshick et al. in CVPR 2014. (Trained by Ross Girshick @rbgirshick)
+- **BAIR GoogLeNet** in `models/bvlc_googlenet`: GoogLeNet trained on ILSVRC 2012, almost exactly as described in [Going Deeper with Convolutions](http://arxiv.org/abs/1409.4842) by Szegedy et al. in ILSVRC 2014. (Trained by Sergio Guadarrama @sguada)
 
 **Community models** made by Caffe users are posted to a publicly editable [wiki page](https://github.com/BVLC/caffe/wiki/Model-Zoo).
 These models are subject to conditions of their respective authors such as citation and license.
@@ -55,14 +55,14 @@ Downloading model info is done just as easily with `scripts/download_model_from_
 ### Hosting trained models
 
 It is up to the user where to host the `.caffemodel` file.
-We host our BVLC-provided models on our own server.
+We host our BAIR-provided models on our own server.
 Dropbox also works fine (tip: make sure that `?dl=1` is appended to the end of the URL).
 
 `scripts/download_model_binary.py <dirname>` downloads the `.caffemodel` from the URL specified in the `<dirname>/readme.md` frontmatter and confirms SHA1.
 
-## BVLC model license
+## BAIR model license
 
-The Caffe models bundled by the BVLC are released for unrestricted use.
+The Caffe models bundled by the BAIR are released for unrestricted use.
 
 These models are trained on data from the [ImageNet project](http://www.image-net.org/) and training data includes internet photos that may be subject to copyright.
 
diff --git a/docs/multigpu.md b/docs/multigpu.md
index d91acef980d..e04ebb0b7c8 100644
--- a/docs/multigpu.md
+++ b/docs/multigpu.md
@@ -13,7 +13,7 @@ The GPUs to be used for training can be set with the "-gpu" flag on the command
 # Hardware Configuration Assumptions
 
 The current implementation uses a tree reduction strategy.  e.g. if there are 4 GPUs in the system, 0:1, 2:3 will exchange gradients, then 0:2 (top of the tree) will exchange gradients, 0 will calculate
-updated model, 0\-\>2, and then 0\-\>1, 2\-\>3. 
+updated model, 0\-\>2, and then 0\-\>1, 2\-\>3.
 
 For best performance, P2P DMA access between devices is needed. Without P2P access, for example crossing PCIe root complex, data is copied through host and effective exchange bandwidth is greatly reduced.
 
@@ -23,4 +23,4 @@ Current implementation has a "soft" assumption that the devices being used are h
 
 # Scaling Performance
 
-Performance is **heavily** dependent on the PCIe topology of the system, the configuration of the neural network you are training, and the speed of each of the layers.  Systems like the DIGITS DevBox have an optimized PCIe topology (X99-E WS chipset).  In general, scaling on 2 GPUs tends to be ~1.8X on average for networks like AlexNet, CaffeNet, VGG, GoogleNet.  4 GPUs begins to have falloff in scaling.  Generally with "weak scaling" where the batchsize increases with the number of GPUs you will see 3.5x scaling or so.  With "strong scaling", the system can become communication bound, especially with layer performance optimizations like those in [cuDNNv3](http://nvidia.com/cudnn), and you will likely see closer to mid 2.x scaling in performance.  Networks that have heavy computation compared to the number of parameters tend to have the best scaling performance.
\ No newline at end of file
+Performance is **heavily** dependent on the PCIe topology of the system, the configuration of the neural network you are training, and the speed of each of the layers.  Systems like the DIGITS DevBox have an optimized PCIe topology (X99-E WS chipset).  In general, scaling on 2 GPUs tends to be ~1.8X on average for networks like AlexNet, CaffeNet, VGG, GoogleNet.  4 GPUs begins to have falloff in scaling.  Generally with "weak scaling" where the batchsize increases with the number of GPUs you will see 3.5x scaling or so.  With "strong scaling", the system can become communication bound, especially with layer performance optimizations like those in [cuDNNv3](http://nvidia.com/cudnn), and you will likely see closer to mid 2.x scaling in performance.  Networks that have heavy computation compared to the number of parameters tend to have the best scaling performance.
diff --git a/docs/performance_hardware.md b/docs/performance_hardware.md
index cdd4b361dea..fbf256842f1 100644
--- a/docs/performance_hardware.md
+++ b/docs/performance_hardware.md
@@ -8,7 +8,7 @@ To measure performance on different NVIDIA GPUs we use CaffeNet, the Caffe refer
 
 For training, each time point is 20 iterations/minibatches of 256 images for 5,120 images total. For testing, a 50,000 image validation set is classified.
 
-**Acknowledgements**: BVLC members are very grateful to NVIDIA for providing several GPUs to conduct this research.
+**Acknowledgements**: BAIR members are very grateful to NVIDIA for providing several GPUs to conduct this research.
 
 ## NVIDIA K40
 
diff --git a/docs/tutorial/interfaces.md b/docs/tutorial/interfaces.md
index d7ff378239d..b5a4f1ad069 100644
--- a/docs/tutorial/interfaces.md
+++ b/docs/tutorial/interfaces.md
@@ -91,7 +91,7 @@ In MatCaffe, you can
 * Run for a certain number of iterations and give back control to Matlab
 * Intermingle arbitrary Matlab code with gradient steps
 
-An ILSVRC image classification demo is in caffe/matlab/demo/classification_demo.m (you need to download BVLC CaffeNet from [Model Zoo](http://caffe.berkeleyvision.org/model_zoo.html) to run it).
+An ILSVRC image classification demo is in caffe/matlab/demo/classification_demo.m (you need to download BAIR CaffeNet from [Model Zoo](http://caffe.berkeleyvision.org/model_zoo.html) to run it).
 
 ### Build MatCaffe
 
@@ -114,7 +114,7 @@ You can save your Matlab search PATH by running `savepath` so that you don't hav
 
 MatCaffe is very similar to PyCaffe in usage.
 
-Examples below shows detailed usages and assumes you have downloaded BVLC CaffeNet from [Model Zoo](http://caffe.berkeleyvision.org/model_zoo.html) and started `matlab` from caffe root folder.
+Examples below shows detailed usages and assumes you have downloaded BAIR CaffeNet from [Model Zoo](http://caffe.berkeleyvision.org/model_zoo.html) and started `matlab` from caffe root folder.
 
     model = './models/bvlc_reference_caffenet/deploy.prototxt';
     weights = './models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel';
diff --git a/examples/finetune_flickr_style/readme.md b/examples/finetune_flickr_style/readme.md
index 188dedf1b9a..dacfd01c8e1 100644
--- a/examples/finetune_flickr_style/readme.md
+++ b/examples/finetune_flickr_style/readme.md
@@ -9,7 +9,7 @@ priority: 5
 # Fine-tuning CaffeNet for Style Recognition on "Flickr Style" Data
 
 Fine-tuning takes an already learned model, adapts the architecture, and resumes training from the already learned model weights.
-Let's fine-tune the BVLC-distributed CaffeNet model on a different dataset, [Flickr Style](http://sergeykarayev.com/files/1311.3715v3.pdf), to predict image style instead of object category.
+Let's fine-tune the BAIR-distributed CaffeNet model on a different dataset, [Flickr Style](http://sergeykarayev.com/files/1311.3715v3.pdf), to predict image style instead of object category.
 
 ## Explanation
 
diff --git a/models/bvlc_alexnet/readme.md b/models/bvlc_alexnet/readme.md
index 008d690f7f4..a83e3d4e27c 100644
--- a/models/bvlc_alexnet/readme.md
+++ b/models/bvlc_alexnet/readme.md
@@ -1,5 +1,5 @@
 ---
-name: BVLC AlexNet Model
+name: BAIR/BVLC AlexNet Model
 caffemodel: bvlc_alexnet.caffemodel
 caffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel
 license: unrestricted
diff --git a/models/bvlc_googlenet/readme.md b/models/bvlc_googlenet/readme.md
index 061b6d74530..ef04db62ab2 100644
--- a/models/bvlc_googlenet/readme.md
+++ b/models/bvlc_googlenet/readme.md
@@ -1,5 +1,5 @@
 ---
-name: BVLC GoogleNet Model
+name: BAIR/BVLC GoogleNet Model
 caffemodel: bvlc_googlenet.caffemodel
 caffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel
 license: unrestricted
diff --git a/models/bvlc_reference_caffenet/readme.md b/models/bvlc_reference_caffenet/readme.md
index 671e47a5056..5352e536a07 100644
--- a/models/bvlc_reference_caffenet/readme.md
+++ b/models/bvlc_reference_caffenet/readme.md
@@ -1,5 +1,5 @@
 ---
-name: BVLC CaffeNet Model
+name: BAIR/BVLC CaffeNet Model
 caffemodel: bvlc_reference_caffenet.caffemodel
 caffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_reference_caffenet.caffemodel
 license: unrestricted
diff --git a/models/bvlc_reference_rcnn_ilsvrc13/readme.md b/models/bvlc_reference_rcnn_ilsvrc13/readme.md
index 9a11a24d8f8..12543b2bd2c 100644
--- a/models/bvlc_reference_rcnn_ilsvrc13/readme.md
+++ b/models/bvlc_reference_rcnn_ilsvrc13/readme.md
@@ -1,5 +1,5 @@
 ---
-name: BVLC Reference RCNN ILSVRC13 Model
+name: BAIR/BVLC Reference RCNN ILSVRC13 Model
 caffemodel: bvlc_reference_rcnn_ilsvrc13.caffemodel
 caffemodel_url: http://dl.caffe.berkeleyvision.org/bvlc_reference_rcnn_ilsvrc13.caffemodel
 license: unrestricted

From 3562698afb4b1f12f51eca752740e279f85714c4 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 12:45:21 -0700
Subject: [PATCH 60/75] drop performance + hardware page and switch to sheet

simpler to read and update
---
 docs/index.md                |  9 +++---
 docs/performance_hardware.md | 73 --------------------------------------------
 2 files changed, 5 insertions(+), 77 deletions(-)
 delete mode 100644 docs/performance_hardware.md

diff --git a/docs/index.md b/docs/index.md
index 302a7d56f88..bbfd91fc7b9 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -23,15 +23,14 @@ Thanks to these contributors the framework tracks the state-of-the-art in both c
 
 **Speed** makes Caffe perfect for research experiments and industry deployment.
 Caffe can process **over 60M images per day** with a single NVIDIA K40 GPU\*.
-That's 1 ms/image for inference and 4 ms/image for learning.
-We believe that Caffe is the fastest convnet implementation available.
+That's 1 ms/image for inference and 4 ms/image for learning and more recent library versions and hardware are faster still.
+We believe that Caffe is among the fastest convnet implementations available.
 
 **Community**: Caffe already powers academic research projects, startup prototypes, and even large-scale industrial applications in vision, speech, and multimedia.
 Join our community of brewers on the [caffe-users group](https://groups.google.com/forum/#!forum/caffe-users) and [Github](https://github.com/BVLC/caffe/).
 
 <p class="footnote" markdown="1">
-\* With the ILSVRC2012-winning [SuperVision](http://www.image-net.org/challenges/LSVRC/2012/supervision.pdf) model and caching IO.
-Consult performance [details](/performance_hardware.html).
+\* With the ILSVRC2012-winning [SuperVision](http://www.image-net.org/challenges/LSVRC/2012/supervision.pdf) model and prefetching IO.
 </p>
 
 ## Documentation
@@ -50,6 +49,8 @@ BAIR suggests a standard distribution format for Caffe models, and provides trai
 Guidelines for development and contributing to Caffe.
 * [API Documentation](/doxygen/annotated.html)<br>
 Developer documentation automagically generated from code comments.
+* [Benchmarking](https://docs.google.com/spreadsheets/d/1Yp4rqHpT7mKxOPbpzYeUfEFLnELDAgxSSBQKp5uKDGQ/edit#gid=0)<br>
+Comparison of inference and learning for different networks and GPUs.
 
 ### Examples
 
diff --git a/docs/performance_hardware.md b/docs/performance_hardware.md
deleted file mode 100644
index fbf256842f1..00000000000
--- a/docs/performance_hardware.md
+++ /dev/null
@@ -1,73 +0,0 @@
----
-title: Performance and Hardware Configuration
----
-
-# Performance and Hardware Configuration
-
-To measure performance on different NVIDIA GPUs we use CaffeNet, the Caffe reference ImageNet model.
-
-For training, each time point is 20 iterations/minibatches of 256 images for 5,120 images total. For testing, a 50,000 image validation set is classified.
-
-**Acknowledgements**: BAIR members are very grateful to NVIDIA for providing several GPUs to conduct this research.
-
-## NVIDIA K40
-
-Performance is best with ECC off and boost clock enabled. While ECC makes a negligible difference in speed, disabling it frees ~1 GB of GPU memory.
-
-Best settings with ECC off and maximum clock speed in standard Caffe:
-
-* Training is 26.5 secs / 20 iterations (5,120 images)
-* Testing is 100 secs / validation set (50,000 images)
-
-Best settings with Caffe + [cuDNN acceleration](http://nvidia.com/cudnn):
-
-* Training is 19.2 secs / 20 iterations (5,120 images)
-* Testing is 60.7 secs / validation set (50,000 images)
-
-Other settings:
-
-* ECC on, max speed: training 26.7 secs / 20 iterations, test 101 secs / validation set
-* ECC on, default speed: training 31 secs / 20 iterations, test 117 secs / validation set
-* ECC off, default speed: training 31 secs / 20 iterations, test 118 secs / validation set
-
-### K40 configuration tips
-
-For maximum K40 performance, turn off ECC and boost the clock speed (at your own risk).
-
-To turn off ECC, do
-
-    sudo nvidia-smi -i 0 --ecc-config=0    # repeat with -i x for each GPU ID
-
-then reboot.
-
-Set the "persistence" mode of the GPU settings by
-
-    sudo nvidia-smi -pm 1
-
-and then set the clock speed with
-
-    sudo nvidia-smi -i 0 -ac 3004,875    # repeat with -i x for each GPU ID
-
-but note that this configuration resets across driver reloading / rebooting. Include these commands in a boot script to initialize these settings. For a simple fix, add these commands to `/etc/rc.local` (on Ubuntu).
-
-## NVIDIA Titan
-
-Training: 26.26 secs / 20 iterations (5,120 images).
-Testing: 100 secs / validation set (50,000 images).
-
-cuDNN Training: 20.25 secs / 20 iterations (5,120 images).
-cuDNN Testing: 66.3 secs / validation set (50,000 images).
-
-
-## NVIDIA K20
-
-Training: 36.0 secs / 20 iterations (5,120 images).
-Testing: 133 secs / validation set (50,000 images).
-
-## NVIDIA GTX 770
-
-Training: 33.0 secs / 20 iterations (5,120 images).
-Testing: 129 secs / validation set (50,000 images).
-
-cuDNN Training: 24.3 secs / 20 iterations (5,120 images).
-cuDNN Testing: 104 secs / validation set (50,000 images).

From 0f5bfc34e0b37b9ab3437d6755eb04a8dc9e8656 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 12:46:56 -0700
Subject: [PATCH 61/75] favor notebook examples as more clear and popular

---
 docs/index.md | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/docs/index.md b/docs/index.md
index bbfd91fc7b9..82eb059e325 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -52,13 +52,6 @@ Developer documentation automagically generated from code comments.
 * [Benchmarking](https://docs.google.com/spreadsheets/d/1Yp4rqHpT7mKxOPbpzYeUfEFLnELDAgxSSBQKp5uKDGQ/edit#gid=0)<br>
 Comparison of inference and learning for different networks and GPUs.
 
-### Examples
-
-{% assign examples = site.pages | where:'category','example' | sort: 'priority' %}
-{% for page in examples %}
-- <div><a href="{{page.url}}">{{page.title}}</a><br>{{page.description}}</div>
-{% endfor %}
-
 ### Notebook Examples
 
 {% assign notebooks = site.pages | where:'category','notebook' | sort: 'priority' %}
@@ -66,6 +59,13 @@ Comparison of inference and learning for different networks and GPUs.
 - <div><a href="http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/{{page.original_path}}">{{page.title}}</a><br>{{page.description}}</div>
 {% endfor %}
 
+### Command Line Examples
+
+{% assign examples = site.pages | where:'category','example' | sort: 'priority' %}
+{% for page in examples %}
+- <div><a href="{{page.url}}">{{page.title}}</a><br>{{page.description}}</div>
+{% endfor %}
+
 ## Citing Caffe
 
 Please cite Caffe in your publications if it helps your research:

From 2158bbb2151049dec2486b720c0a351164a0eb6b Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 12:50:19 -0700
Subject: [PATCH 62/75] model zoo: point out wiki link immediately, explain
 manual editing

---
 docs/model_zoo.md | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/docs/model_zoo.md b/docs/model_zoo.md
index f9078718a8b..3f77e82572c 100644
--- a/docs/model_zoo.md
+++ b/docs/model_zoo.md
@@ -3,7 +3,7 @@ title: Model Zoo
 ---
 # Caffe Model Zoo
 
-Lots of researchers and engineers have made Caffe models for different tasks with all kinds of architectures and data.
+Lots of researchers and engineers have made Caffe models for different tasks with all kinds of architectures and data: check out the [model zoo](https://github.com/BVLC/caffe/wiki/Model-Zoo)!
 These models are learned and applied for problems ranging from simple regression, to large-scale visual classification, to Siamese networks for image similarity, to speech and robotics applications.
 
 To help share these models, we introduce the model zoo framework:
@@ -24,7 +24,7 @@ Each one of these can be downloaded by running `scripts/download_model_binary.py
 - **BAIR Reference R-CNN ILSVRC-2013** in `models/bvlc_reference_rcnn_ilsvrc13`: pure Caffe implementation of [R-CNN](https://github.com/rbgirshick/rcnn) as described by Girshick et al. in CVPR 2014. (Trained by Ross Girshick @rbgirshick)
 - **BAIR GoogLeNet** in `models/bvlc_googlenet`: GoogLeNet trained on ILSVRC 2012, almost exactly as described in [Going Deeper with Convolutions](http://arxiv.org/abs/1409.4842) by Szegedy et al. in ILSVRC 2014. (Trained by Sergio Guadarrama @sguada)
 
-**Community models** made by Caffe users are posted to a publicly editable [wiki page](https://github.com/BVLC/caffe/wiki/Model-Zoo).
+**Community models** made by Caffe users are posted to a publicly editable [model zoo wiki page](https://github.com/BVLC/caffe/wiki/Model-Zoo).
 These models are subject to conditions of their respective authors such as citation and license.
 Thank you for sharing your models!
 
@@ -42,6 +42,8 @@ A caffe model is distributed as a directory containing:
     - License information.
 - [optional] Other helpful scripts.
 
+This simple format can be handled through bundled scripts or manually if need be.
+
 ### Hosting model info
 
 Github Gist is a good format for model info distribution because it can contain multiple files, is versionable, and has in-browser syntax highlighting and markdown rendering.

From 414b74c06038c17924745b68954ef10827fe1edd Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 13:19:53 -0700
Subject: [PATCH 63/75] add missing names to BAIR roster

---
 docs/index.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/index.md b/docs/index.md
index 82eb059e325..db8eaffbe34 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -96,7 +96,7 @@ The core Caffe developers offer [consulting services](mailto:caffe-coldpress@goo
 The BAIR Caffe developers would like to thank NVIDIA for GPU donation, A9 and Amazon Web Services for a research grant in support of Caffe development and reproducible research in deep learning, and BAIR PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for guidance.
 
 The BAIR members who have contributed to Caffe are (alphabetical by first name):
-[Eric Tzeng](https://github.com/erictzeng), [Evan Shelhamer](http://imaginarynumber.net/), [Jeff Donahue](http://jeffdonahue.com/), [Jon Long](https://github.com/longjon), [Ross Girshick](http://www.cs.berkeley.edu/~rbg/), [Sergey Karayev](http://sergeykarayev.com/), [Sergio Guadarrama](http://www.eecs.berkeley.edu/~sguada/), and [Yangqing Jia](http://daggerfs.com/).
+[Carl Doersch](http://www.carldoersch.com/), [Eric Tzeng](https://github.com/erictzeng), [Evan Shelhamer](http://imaginarynumber.net/), [Jeff Donahue](http://jeffdonahue.com/), [Jon Long](https://github.com/longjon), [Philipp Krähenbühl](http://www.philkr.net/), [Ronghang Hu](http://ronghanghu.com/), [Ross Girshick](http://www.cs.berkeley.edu/~rbg/), [Sergey Karayev](http://sergeykarayev.com/), [Sergio Guadarrama](http://www.eecs.berkeley.edu/~sguada/), [Takuya Narihira](https://github.com/tnarihi), and [Yangqing Jia](http://daggerfs.com/).
 
 The open-source community plays an important and growing role in Caffe's development.
 Check out the Github [project pulse](https://github.com/BVLC/caffe/pulse) for recent activity and the [contributors](https://github.com/BVLC/caffe/graphs/contributors) for the full list.

From e90a6a6ca29423afb15f39adb1157bff9e6f8655 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 13:24:30 -0700
Subject: [PATCH 64/75] retire caffe-dev and caffe-coldpress

dev has diffused into the community from the original Caffe core
---
 docs/index.md | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/docs/index.md b/docs/index.md
index db8eaffbe34..0e21ae821b0 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -86,11 +86,6 @@ Join the [caffe-users group](https://groups.google.com/forum/#!forum/caffe-users
 
 Framework development discussions and thorough bug reports are collected on [Issues](https://github.com/BVLC/caffe/issues).
 
-Contact [caffe-dev](mailto:caffe-dev@googlegroups.com) if you have a confidential proposal for the framework *and the ability to act on it*.
-Requests for features, explanations, or personal help will be ignored; post to [caffe-users](https://groups.google.com/forum/#!forum/caffe-users) instead.
-
-The core Caffe developers offer [consulting services](mailto:caffe-coldpress@googlegroups.com) for appropriate projects.
-
 ## Acknowledgements
 
 The BAIR Caffe developers would like to thank NVIDIA for GPU donation, A9 and Amazon Web Services for a research grant in support of Caffe development and reproducible research in deep learning, and BAIR PI [Trevor Darrell](http://www.eecs.berkeley.edu/~trevor/) for guidance.

From 8985818e4fbb5fc207e4f383c63c28d80fd286f2 Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 13:28:24 -0700
Subject: [PATCH 65/75] track publications by google scholar and not the wiki

---
 docs/index.md | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/docs/index.md b/docs/index.md
index 0e21ae821b0..3385747c565 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -77,8 +77,7 @@ Please cite Caffe in your publications if it helps your research:
       Year = {2014}
     }
 
-If you do publish a paper where Caffe helped your research, we encourage you to update the [publications wiki](https://github.com/BVLC/caffe/wiki/Publications).
-Citations are also tracked automatically by [Google Scholar](http://scholar.google.com/scholar?oi=bibs&hl=en&cites=17333247995453974016).
+If you do publish a paper where Caffe helped your research, we encourage you to cite the framework for tracking by [Google Scholar](https://scholar.google.com/citations?view_op=view_citation&hl=en&citation_for_view=-ltRSM0AAAAJ:u5HHmVD_uO8C).
 
 ## Contacting Us
 

From 8b8f2dd40ba87543f066cb157c6d65dd8187253f Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 15:26:30 -0700
Subject: [PATCH 66/75] link to new full-day crash course

---
 docs/index.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/index.md b/docs/index.md
index 3385747c565..b633f7cfddc 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -35,8 +35,8 @@ Join our community of brewers on the [caffe-users group](https://groups.google.c
 
 ## Documentation
 
-- [DIY Deep Learning for Vision with Caffe](https://docs.google.com/presentation/d/1UeKXVgRvvxg9OUdh_UiC5G71UMscNPlvArsWER41PsU/edit#slide=id.p)<br>
-Tutorial presentation.
+- [DIY Deep Learning for Vision with Caffe](https://docs.google.com/presentation/d/1UeKXVgRvvxg9OUdh_UiC5G71UMscNPlvArsWER41PsU/edit#slide=id.p) and [Caffe in a Day](https://docs.google.com/presentation/d/1HxGdeq8MPktHaPb-rlmYYQ723iWzq9ur6Gjo71YiG0Y/edit#slide=id.gc2fcdcce7_216_0)<br>
+Tutorial presentation of the framework and a full-day crash course.
 - [Tutorial Documentation](/tutorial)<br>
 Practical guide and framework reference.
 - [arXiv / ACM MM '14 paper](http://arxiv.org/abs/1408.5093)<br>

From 49761d34d18b7063af995b13ecca0fee1bdaf02c Mon Sep 17 00:00:00 2001
From: Evan Shelhamer <shelhamer@imaginarynumber.net>
Date: Fri, 14 Apr 2017 15:32:50 -0700
Subject: [PATCH 67/75] Caffe 1.0

---
 CMakeLists.txt | 4 ++--
 Makefile       | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c52ff466471..08f56a33a59 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -10,8 +10,8 @@ endif()
 project(Caffe C CXX)
 
 # ---[ Caffe version
-set(CAFFE_TARGET_VERSION "1.0.0-rc5" CACHE STRING "Caffe logical version")
-set(CAFFE_TARGET_SOVERSION "1.0.0-rc5" CACHE STRING "Caffe soname version")
+set(CAFFE_TARGET_VERSION "1.0.0" CACHE STRING "Caffe logical version")
+set(CAFFE_TARGET_SOVERSION "1.0.0" CACHE STRING "Caffe soname version")
 add_definitions(-DCAFFE_VERSION=${CAFFE_TARGET_VERSION})
 
 # ---[ Using cmake scripts and modules
diff --git a/Makefile b/Makefile
index 77900b69b97..4d324160c08 100644
--- a/Makefile
+++ b/Makefile
@@ -34,7 +34,7 @@ LIB_BUILD_DIR := $(BUILD_DIR)/lib
 STATIC_NAME := $(LIB_BUILD_DIR)/lib$(LIBRARY_NAME).a
 DYNAMIC_VERSION_MAJOR 		:= 1
 DYNAMIC_VERSION_MINOR 		:= 0
-DYNAMIC_VERSION_REVISION 	:= 0-rc5
+DYNAMIC_VERSION_REVISION 	:= 0
 DYNAMIC_NAME_SHORT := lib$(LIBRARY_NAME).so
 #DYNAMIC_SONAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR)
 DYNAMIC_VERSIONED_NAME_SHORT := $(DYNAMIC_NAME_SHORT).$(DYNAMIC_VERSION_MAJOR).$(DYNAMIC_VERSION_MINOR).$(DYNAMIC_VERSION_REVISION)

From 33f86122970392fcda19ef80ed5cd349279b896d Mon Sep 17 00:00:00 2001
From: Eric Tzeng <etzeng@pinterest.com>
Date: Tue, 18 Apr 2017 18:22:38 -0700
Subject: [PATCH 68/75] Rewrite crop cuda kernel

---
 include/caffe/layers/crop_layer.hpp |   6 +-
 src/caffe/layers/crop_layer.cpp     |  21 +++++--
 src/caffe/layers/crop_layer.cu      | 122 +++++++++++++++---------------------
 3 files changed, 69 insertions(+), 80 deletions(-)

diff --git a/include/caffe/layers/crop_layer.hpp b/include/caffe/layers/crop_layer.hpp
index c4fda1220c3..5219fa5cb5f 100644
--- a/include/caffe/layers/crop_layer.hpp
+++ b/include/caffe/layers/crop_layer.hpp
@@ -41,13 +41,15 @@ class CropLayer : public Layer<Dtype> {
   virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
       const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
 
-  vector<int> offsets;
+  Blob<int> offsets;
+  Blob<int> src_strides_;
+  Blob<int> dest_strides_;
 
  private:
   // Recursive copy function.
   void crop_copy(const vector<Blob<Dtype>*>& bottom,
                const vector<Blob<Dtype>*>& top,
-               const vector<int>& offsets,
+               const int* offsets,
                vector<int> indices,
                int cur_dim,
                const Dtype* src_data,
diff --git a/src/caffe/layers/crop_layer.cpp b/src/caffe/layers/crop_layer.cpp
index ef8c177c4dd..65ea8f8b7d0 100644
--- a/src/caffe/layers/crop_layer.cpp
+++ b/src/caffe/layers/crop_layer.cpp
@@ -40,8 +40,10 @@ void CropLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
   const int start_axis = bottom[0]->CanonicalAxisIndex(param.axis());
 
   // Initialize offsets to 0 and the new shape to the current shape of the data.
-  offsets = vector<int>(input_dim, 0);
   vector<int> new_shape(bottom[0]->shape());
+  vector<int> offsets_shape(1, input_dim);
+  offsets.Reshape(offsets_shape);
+  int* offset_data = offsets.mutable_cpu_data();
 
   // Determine crop offsets and the new shape post-crop.
   for (int i = 0; i < input_dim; ++i) {
@@ -63,15 +65,22 @@ void CropLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
           << "size " << bottom[1]->shape(i) << " and offset " << crop_offset;
     }
     new_shape[i] = new_size;
-    offsets[i] = crop_offset;
+    offset_data[i] = crop_offset;
   }
   top[0]->Reshape(new_shape);
+  // Compute strides
+  src_strides_.Reshape(offsets_shape);
+  dest_strides_.Reshape(offsets_shape);
+  for (int i = 0; i < input_dim; ++i) {
+    src_strides_.mutable_cpu_data()[i] = bottom[0]->count(i + 1, input_dim);
+    dest_strides_.mutable_cpu_data()[i] = top[0]->count(i + 1, input_dim);
+  }
 }
 
 template <typename Dtype>
 void CropLayer<Dtype>::crop_copy(const vector<Blob<Dtype>*>& bottom,
              const vector<Blob<Dtype>*>& top,
-             const vector<int>& offsets,
+             const int* offsets,
              vector<int> indices,
              int cur_dim,
              const Dtype* src_data,
@@ -115,7 +124,8 @@ void CropLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   std::vector<int> indices(top[0]->num_axes(), 0);
   const Dtype* bottom_data = bottom[0]->cpu_data();
   Dtype* top_data = top[0]->mutable_cpu_data();
-  crop_copy(bottom, top, offsets, indices, 0, bottom_data, top_data, true);
+  crop_copy(bottom, top, offsets.cpu_data(), indices, 0, bottom_data, top_data,
+      true);
 }
 
 template <typename Dtype>
@@ -127,7 +137,8 @@ void CropLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
   if (propagate_down[0]) {
     caffe_set(bottom[0]->count(), static_cast<Dtype>(0), bottom_diff);
     std::vector<int> indices(top[0]->num_axes(), 0);
-    crop_copy(bottom, top, offsets, indices, 0, top_diff, bottom_diff, false);
+    crop_copy(bottom, top, offsets.cpu_data(), indices, 0, top_diff,
+        bottom_diff, false);
   }
 }
 
diff --git a/src/caffe/layers/crop_layer.cu b/src/caffe/layers/crop_layer.cu
index 677077cdd8b..a400f333e14 100644
--- a/src/caffe/layers/crop_layer.cu
+++ b/src/caffe/layers/crop_layer.cu
@@ -4,90 +4,62 @@
 
 namespace caffe {
 
-// Copy (one line per thread) from one array to another, with arbitrary
-// strides in the last two dimensions.
+__device__ int compute_uncropped_index(
+    int index,
+    const int ndims,
+    const int* src_strides,
+    const int* dest_strides,
+    const int* offsets) {
+  int dest_index = index;
+  int src_index = 0;
+  for (int i = 0; i < ndims; ++i) {
+      int coord = dest_index / dest_strides[i];
+      dest_index -= coord * dest_strides[i];
+      src_index += src_strides[i] * (coord + offsets[i]);
+  }
+  return src_index;
+}
+
 template <typename Dtype>
-__global__ void copy_kernel(const int n, const int height, const int width,
-    const int src_inner_stride,
-    const int dest_inner_stride,
+__global__ void crop_kernel_forward(const int nthreads,
+    const int ndims,
+    const int* src_strides,
+    const int* dest_strides,
+    const int* offsets,
     const Dtype* src, Dtype* dest) {
-  CUDA_KERNEL_LOOP(index, n) {
-    int src_start = index * src_inner_stride;
-    int dest_start = index * dest_inner_stride;
-    for (int i = 0; i < width; ++i) {
-      dest[dest_start + i] = src[src_start + i];
-    }
+  CUDA_KERNEL_LOOP(index, nthreads) {
+    int src_index = compute_uncropped_index(
+        index, ndims, src_strides, dest_strides, offsets);
+    dest[index] = src[src_index];
   }
 }
 
 template <typename Dtype>
-void CropLayer<Dtype>::crop_copy_gpu(const vector<Blob<Dtype>*>& bottom,
-             const vector<Blob<Dtype>*>& top,
-             const vector<int>& offsets,
-             vector<int> indices,
-             int cur_dim,
-             const Dtype* src_data,
-             Dtype* dest_data,
-             bool is_forward) {
-  if (cur_dim + 2 < top[0]->num_axes()) {
-    // We are not yet at the final dimension, call copy recursivley
-    for (int i = 0; i < top[0]->shape(cur_dim); ++i) {
-      indices[cur_dim] = i;
-      crop_copy_gpu(bottom, top, offsets, indices, cur_dim+1,
-                src_data, dest_data, is_forward);
-    }
-  } else {
-    // We are at the last two dimensions, which are stored continuously in
-    // memory. With (N,C,H,W)
-    //              (0,1,2,3) cur_dim   -> H
-    //                        cur_dim+1 -> W
-    const int lines = top[0]->shape(cur_dim);
-    const int height = top[0]->shape(cur_dim);
-    const int width = top[0]->shape(cur_dim+1);
-    std::vector<int> ind_off(cur_dim+2, 0);
-    for (int j = 0; j < cur_dim; ++j) {
-        ind_off[j] = indices[j] + offsets[j];
-    }
-    ind_off[cur_dim] = offsets[cur_dim];
-    ind_off[cur_dim+1] = offsets[cur_dim+1];
-    // Compute copy strides
-    const int src_inner_stride = bottom[0]->shape(cur_dim+1);
-    const int dest_inner_stride = top[0]->shape(cur_dim+1);
-
-    if (is_forward) {
-      const Dtype* bottom_data = bottom[0]->gpu_data() +
-          bottom[0]->offset(ind_off);
-      Dtype* top_data = top[0]->mutable_gpu_data() +
-          top[0]->offset(indices);
-      // NOLINT_NEXT_LINE(whitespace/operators)
-      copy_kernel<<<CAFFE_GET_BLOCKS(lines), CAFFE_CUDA_NUM_THREADS>>>(
-          lines, height, width,
-          src_inner_stride,
-          dest_inner_stride,
-          bottom_data, top_data);
-
-    } else {
-      const Dtype* top_diff = top[0]->gpu_diff() +
-          top[0]->offset(indices);
-      Dtype* bottom_diff = bottom[0]->mutable_gpu_diff() +
-          bottom[0]->offset(ind_off);
-      // NOLINT_NEXT_LINE(whitespace/operators)
-      copy_kernel<<<CAFFE_GET_BLOCKS(lines), CAFFE_CUDA_NUM_THREADS>>>(
-          lines, height, width,
-          dest_inner_stride,
-          src_inner_stride,
-          top_diff, bottom_diff);
-    }
+__global__ void crop_kernel_backward(const int nthreads,
+    const int ndims,
+    const int* src_strides,
+    const int* dest_strides,
+    const int* offsets,
+    Dtype* src, const Dtype* dest) {
+  CUDA_KERNEL_LOOP(index, nthreads) {
+    int src_index = compute_uncropped_index(
+        index, ndims, src_strides, dest_strides, offsets);
+    src[src_index] = dest[index];
   }
 }
 
 template <typename Dtype>
 void CropLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
-  std::vector<int> indices(top[0]->num_axes(), 0);
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* top_data = top[0]->mutable_gpu_data();
-  crop_copy_gpu(bottom, top, offsets, indices, 0, bottom_data, top_data, true);
+  int n = top[0]->count();
+  crop_kernel_forward<<<CAFFE_GET_BLOCKS(n), CAFFE_CUDA_NUM_THREADS>>>(n,
+      bottom[0]->num_axes(),
+      src_strides_.gpu_data(),
+      dest_strides_.gpu_data(),
+      offsets.gpu_data(),
+      bottom_data, top_data);
 }
 
 template <typename Dtype>
@@ -95,12 +67,16 @@ void CropLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
     const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
   const Dtype* top_diff = top[0]->gpu_diff();
   Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
+  int n = top[0]->count();
 
   if (propagate_down[0]) {
     caffe_gpu_set(bottom[0]->count(), static_cast<Dtype>(0), bottom_diff);
-    std::vector<int> indices(top[0]->num_axes(), 0);
-    crop_copy_gpu(bottom, top, offsets, indices, 0, top_diff, bottom_diff,
-                  false);
+    crop_kernel_backward<<<CAFFE_GET_BLOCKS(n), CAFFE_CUDA_NUM_THREADS>>>(n,
+        bottom[0]->num_axes(),
+        src_strides_.gpu_data(),
+        dest_strides_.gpu_data(),
+        offsets.gpu_data(),
+        bottom_diff, top_diff);
   }
 }
 

From cd1696d00b995a1d8567cb6f3ad7f65ec4df4176 Mon Sep 17 00:00:00 2001
From: Eric Tzeng <etzeng@pinterest.com>
Date: Tue, 18 Apr 2017 18:48:26 -0700
Subject: [PATCH 69/75] Fix crop layer lint errors

---
 src/caffe/layers/crop_layer.cu | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/src/caffe/layers/crop_layer.cu b/src/caffe/layers/crop_layer.cu
index a400f333e14..4ece9cd1761 100644
--- a/src/caffe/layers/crop_layer.cu
+++ b/src/caffe/layers/crop_layer.cu
@@ -54,6 +54,7 @@ void CropLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* top_data = top[0]->mutable_gpu_data();
   int n = top[0]->count();
+  // NOLINT_NEXT_LINE(whitespace/operators)
   crop_kernel_forward<<<CAFFE_GET_BLOCKS(n), CAFFE_CUDA_NUM_THREADS>>>(n,
       bottom[0]->num_axes(),
       src_strides_.gpu_data(),
@@ -71,6 +72,7 @@ void CropLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
 
   if (propagate_down[0]) {
     caffe_gpu_set(bottom[0]->count(), static_cast<Dtype>(0), bottom_diff);
+    // NOLINT_NEXT_LINE(whitespace/operators)
     crop_kernel_backward<<<CAFFE_GET_BLOCKS(n), CAFFE_CUDA_NUM_THREADS>>>(n,
         bottom[0]->num_axes(),
         src_strides_.gpu_data(),

From ec35395e131a0d5e7c55cbd74dadbd46a49a645c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Malte=20St=C3=A6r=20Nissen?= <nissen@di.ku.dk>
Date: Thu, 4 May 2017 14:33:40 +0200
Subject: [PATCH 70/75] Handling destruction of empty Net objects

---
 matlab/+caffe/Net.m | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/matlab/+caffe/Net.m b/matlab/+caffe/Net.m
index 349e060eb22..bb99ec89049 100644
--- a/matlab/+caffe/Net.m
+++ b/matlab/+caffe/Net.m
@@ -69,7 +69,9 @@
       self.blob_names = self.attributes.blob_names;
     end
     function delete (self)
-      caffe_('delete_net', self.hNet_self);
+      if ~isempty(self.hNet_self)
+        caffe_('delete_net', self.hNet_self);
+      end
     end
     function layer = layers(self, layer_name)
       CHECK(ischar(layer_name), 'layer_name must be a string');

From b7e2b99c7f0aeeb8e24046f8cbf5212065b9ccdf Mon Sep 17 00:00:00 2001
From: Luke Yeager <lukeyeager@users.noreply.github.com>
Date: Fri, 12 May 2017 10:06:51 -0700
Subject: [PATCH 71/75] Downgrade boost requirement from 1.55 to 1.54

---
 cmake/Dependencies.cmake       | 2 +-
 scripts/travis/install-deps.sh | 8 ++++----
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
index 4a5bac471b4..c48255c89f2 100644
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -5,7 +5,7 @@ set(Caffe_DEFINITIONS "")
 set(Caffe_COMPILE_OPTIONS "")
 
 # ---[ Boost
-find_package(Boost 1.55 REQUIRED COMPONENTS system thread filesystem)
+find_package(Boost 1.54 REQUIRED COMPONENTS system thread filesystem)
 list(APPEND Caffe_INCLUDE_DIRS PUBLIC ${Boost_INCLUDE_DIRS})
 list(APPEND Caffe_LINKER_LIBS PUBLIC ${Boost_LIBRARIES})
 
diff --git a/scripts/travis/install-deps.sh b/scripts/travis/install-deps.sh
index dac5d2f9d37..2fa2a74a486 100755
--- a/scripts/travis/install-deps.sh
+++ b/scripts/travis/install-deps.sh
@@ -9,10 +9,10 @@ apt-get -y update
 apt-get install -y --no-install-recommends \
   build-essential \
   graphviz \
-  libboost-filesystem1.55-dev \
-  libboost-python1.55-dev \
-  libboost-system1.55-dev \
-  libboost-thread1.55-dev \
+  libboost-filesystem-dev \
+  libboost-python-dev \
+  libboost-system-dev \
+  libboost-thread-dev \
   libgflags-dev \
   libgoogle-glog-dev \
   libhdf5-serial-dev \

From 30a2ab7e50430911f37ddf981e67e4f36f662f14 Mon Sep 17 00:00:00 2001
From: Zhou Mo <cdluminate@gmail.com>
Date: Mon, 15 May 2017 02:16:19 +0000
Subject: [PATCH 72/75] cmake: rename libproto.a -> libcaffeproto.a

---
 cmake/ConfigGen.cmake    |  2 +-
 src/caffe/CMakeLists.txt | 12 ++++++------
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/cmake/ConfigGen.cmake b/cmake/ConfigGen.cmake
index ad91f542104..09bb09b4ff2 100644
--- a/cmake/ConfigGen.cmake
+++ b/cmake/ConfigGen.cmake
@@ -33,7 +33,7 @@ function(caffe_generate_export_configs)
   configure_file("cmake/Templates/CaffeConfig.cmake.in" "${PROJECT_BINARY_DIR}/CaffeConfig.cmake" @ONLY)
 
   # Add targets to the build-tree export set
-  export(TARGETS caffe proto FILE "${PROJECT_BINARY_DIR}/CaffeTargets.cmake")
+  export(TARGETS caffe caffeproto FILE "${PROJECT_BINARY_DIR}/CaffeTargets.cmake")
   export(PACKAGE Caffe)
 
   # ---[ Configure install-tree CaffeConfig.cmake file ]---
diff --git a/src/caffe/CMakeLists.txt b/src/caffe/CMakeLists.txt
index b9152e9216f..4a805568566 100644
--- a/src/caffe/CMakeLists.txt
+++ b/src/caffe/CMakeLists.txt
@@ -3,12 +3,12 @@ file(GLOB proto_files proto/*.proto)
 caffe_protobuf_generate_cpp_py(${proto_gen_folder} proto_srcs proto_hdrs proto_python ${proto_files})
 
 # include python files either to force generation
-add_library(proto STATIC ${proto_hdrs} ${proto_srcs} ${proto_python})
-caffe_default_properties(proto)
-target_link_libraries(proto PUBLIC ${PROTOBUF_LIBRARIES})
-target_include_directories(proto PUBLIC ${PROTOBUF_INCLUDE_DIR})
+add_library(caffeproto STATIC ${proto_hdrs} ${proto_srcs} ${proto_python})
+caffe_default_properties(caffeproto)
+target_link_libraries(caffeproto PUBLIC ${PROTOBUF_LIBRARIES})
+target_include_directories(caffeproto PUBLIC ${PROTOBUF_INCLUDE_DIR})
 
-list(INSERT Caffe_LINKER_LIBS 0 PUBLIC proto) # note, crucial to prepend!
+list(INSERT Caffe_LINKER_LIBS 0 PUBLIC caffeproto) # note, crucial to prepend!
 
 # --[ Caffe library
 
@@ -42,7 +42,7 @@ set_target_properties(caffe PROPERTIES
 # ---[ Install
 install(DIRECTORY ${Caffe_INCLUDE_DIR}/caffe DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 install(FILES ${proto_hdrs} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/caffe/proto)
-install(TARGETS caffe proto EXPORT CaffeTargets DESTINATION ${CMAKE_INSTALL_LIBDIR})
+install(TARGETS caffe caffeproto EXPORT CaffeTargets DESTINATION ${CMAKE_INSTALL_LIBDIR})
 
 file(WRITE ${PROJECT_BINARY_DIR}/__init__.py)
 list(APPEND proto_python ${PROJECT_BINARY_DIR}/__init__.py)

From 83814da36d5a44039ddc35f58f9b341e9d1bd935 Mon Sep 17 00:00:00 2001
From: Zhou Mo <cdluminate@gmail.com>
Date: Mon, 15 May 2017 03:04:47 +0000
Subject: [PATCH 73/75] docs/debian guide: update compiler combination table

---
 docs/install_apt_debian.md | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)

diff --git a/docs/install_apt_debian.md b/docs/install_apt_debian.md
index 65fe70924e1..bd91124a898 100644
--- a/docs/install_apt_debian.md
+++ b/docs/install_apt_debian.md
@@ -96,18 +96,22 @@ Note, this requires a `deb-src` entry in your `/etc/apt/sources.list`.
 Some users may find their favorate compiler doesn't work with CUDA.
 
 ```
-CXX compiler |  CUDA 7.5  |  CUDA 8.0  |
--------------+------------+------------+-
-GCC-7        |     ?      |     ?      |
-GCC-6        |     ✘      |     ✘      |
-GCC-5        |     ✔ [1]  |     ✔      |
-CLANG-4.0    |     ?      |     ?      |
-CLANG-3.9    |     ✘      |     ✘      |
-CLANG-3.8    |     ?      |     ✔      |
+CXX compiler |  CUDA 7.5  |  CUDA 8.0  |  CUDA 9.0  |
+-------------+------------+------------+------------+
+GCC-8        |     ?      |     ?      |     ?      |
+GCC-7        |     ?      |     ?      |     ?      |
+GCC-6        |     ✘      |     ✘      |     ✔      |
+GCC-5        |     ✔ [1]  |     ✔      |     ✔      |
+-------------+------------+------------+------------+
+CLANG-4.0    |     ?      |     ?      |     ?      |
+CLANG-3.9    |     ✘      |     ✘      |     ✔      |
+CLANG-3.8    |     ?      |     ✔      |     ✔      |
 ```
 
 `[1]` CUDA 7.5 's `host_config.h` must be patched before working with GCC-5.
 
+`[2]` CUDA 9.0: https://devblogs.nvidia.com/parallelforall/cuda-9-features-revealed/
+
 BTW, please forget the GCC-4.X series, since its `libstdc++` ABI is not compatible with GCC-5's.
 You may encounter failure linking GCC-4.X object files against GCC-5 libraries.
 (See https://wiki.debian.org/GCC5 )

From 264cf199e4e8bc44bb97762b1018137704157c2c Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Tue, 13 Jun 2017 11:59:26 -0700
Subject: [PATCH 74/75] List branches in readme

---
 README.md | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/README.md b/README.md
index 0ae3616b4a6..c40aee65c3c 100644
--- a/README.md
+++ b/README.md
@@ -15,6 +15,14 @@ Check out the [project site](http://caffe.berkeleyvision.org) for all the detail
 
 and step-by-step examples.
 
+## Custom distributions
+
+- [Intel optimized branch](https://github.com/BVLC/caffe/tree/intel) for CPU, in particular Xeon processors (HSW, BDW, Xeon Phi).
+- [OpenCL Caffe](https://github.com/BVLC/caffe/tree/opencl) e.g. for AMD or Intel devices.
+- [Windows Caffe](https://github.com/BVLC/caffe/tree/windows)
+
+## Community
+
 [![Join the chat at https://gitter.im/BVLC/caffe](https://badges.gitter.im/Join%20Chat.svg)](https://gitter.im/BVLC/caffe?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge&utm_content=badge)
 
 Please join the [caffe-users group](https://groups.google.com/forum/#!forum/caffe-users) or [gitter chat](https://gitter.im/BVLC/caffe) to ask questions and talk about methods and models.

From 4efdf7ee49cffefdd7ea099c00dc5ea327640f04 Mon Sep 17 00:00:00 2001
From: Cyprien Noel <cyprien.noel@gmail.com>
Date: Tue, 20 Jun 2017 14:20:42 -0700
Subject: [PATCH 75/75] Update README.md

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index c40aee65c3c..5148c69d310 100644
--- a/README.md
+++ b/README.md
@@ -17,7 +17,7 @@ and step-by-step examples.
 
 ## Custom distributions
 
-- [Intel optimized branch](https://github.com/BVLC/caffe/tree/intel) for CPU, in particular Xeon processors (HSW, BDW, Xeon Phi).
+ - [Intel Caffe](https://github.com/BVLC/caffe/tree/intel) (Optimized for CPU and support for multi-node), in particular Xeon processors (HSW, BDW, Xeon Phi).
 - [OpenCL Caffe](https://github.com/BVLC/caffe/tree/opencl) e.g. for AMD or Intel devices.
 - [Windows Caffe](https://github.com/BVLC/caffe/tree/windows)