Updated Caffe

3rdparty/Versions.txt

@@ -1,6 +1,6 @@
 Unix:
     - Caffe:
-        - Version 1.0.0, extracted from GitHub on 09/30/2017 from the current master branch.
+        - Version 1.0.0, extracted from GitHub on 11/03/2017 from the current master branch.
         - Link: https://github.com/BVLC/caffe
 
 Windows:

3rdparty/caffe/Makefile

@@ -699,6 +699,6 @@ $(DISTRIBUTE_DIR): all py | $(DISTRIBUTE_SUBDIRS)
 	install -m 644 $(DYNAMIC_NAME) $(DISTRIBUTE_DIR)/lib
 	cd $(DISTRIBUTE_DIR)/lib; rm -f $(DYNAMIC_NAME_SHORT);   ln -s $(DYNAMIC_VERSIONED_NAME_SHORT) $(DYNAMIC_NAME_SHORT)
 	# add python - it's not the standard way, indeed...
-	cp -r python $(DISTRIBUTE_DIR)/python
+	cp -r python $(DISTRIBUTE_DIR)/
 
 -include $(DEPS)

3rdparty/caffe/docs/development.md

@@ -116,5 +116,5 @@ To get a list of all options `googletest` provides, simply pass the `--help` fla
 
 - **Run `make lint` to check C++ code.**
 - Wrap lines at 80 chars.
-- Follow [Google C++ style](http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml) and [Google python style](http://google-styleguide.googlecode.com/svn/trunk/pyguide.html) + [PEP 8](http://legacy.python.org/dev/peps/pep-0008/).
+- Follow [Google C++ style](https://google.github.io/styleguide/cppguide.html) and [Google python style](https://google.github.io/styleguide/pyguide.html) + [PEP 8](http://legacy.python.org/dev/peps/pep-0008/).
 - Remember that “a foolish consistency is the hobgoblin of little minds,” so use your best judgement to write the clearest code for your particular case.

3rdparty/caffe/docs/install_apt.md

@@ -40,6 +40,7 @@ Continue with [compilation](installation.html#compilation).
 
     sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
     sudo apt-get install --no-install-recommends libboost-all-dev
+    sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
 
 **CUDA**: Install by `apt-get` or the NVIDIA `.run` package.
 The NVIDIA package tends to follow more recent library and driver versions, but the installation is more manual.
@@ -54,12 +55,6 @@ This can be skipped for CPU-only installation.
 
 CUDA 8 is required on Ubuntu 16.04.
 
-**Remaining dependencies, 14.04**
-
-Everything is packaged in 14.04.
-
-    sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
-
 **Remaining dependencies, 12.04**
 
 These dependencies need manual installation in 12.04.

3rdparty/caffe/docs/tutorial/layers.md

@@ -87,7 +87,7 @@ Layers:
 * [ELU](layers/elu.html) - exponential linear rectification.
 * [Sigmoid](layers/sigmoid.html)
 * [TanH](layers/tanh.html)
-* [Absolute Value](layers/abs.html)
+* [Absolute Value](layers/absval.html)
 * [Power](layers/power.html) - f(x) = (shift + scale * x) ^ power.
 * [Exp](layers/exp.html) - f(x) = base ^ (shift + scale * x).
 * [Log](layers/log.html) - f(x) = log(x).

3rdparty/caffe/examples/web_demo/readme.md

@@ -11,7 +11,7 @@ priority: 10
 ## Requirements
 
 The demo server requires Python with some dependencies.
-To make sure you have the dependencies, please run `pip install -r examples/web_demo/requirements.txt`, and also make sure that you've compiled the Python Caffe interface and that it is on your `PYTHONPATH` (see [installation instructions](/installation.html)).
+To make sure you have the dependencies, please run `pip install -r examples/web_demo/requirements.txt`, and also make sure that you've compiled the Python Caffe interface and that it is on your `PYTHONPATH` (see [installation instructions](http://caffe.berkeleyvision.org/installation.html)).
 
 Make sure that you have obtained the Reference CaffeNet Model and the ImageNet Auxiliary Data:
 

3rdparty/caffe/include/caffe/filler.hpp

@@ -250,10 +250,10 @@ class BilinearFiller : public Filler<Dtype> {
     CHECK_EQ(blob->width(), blob->height()) << "Filter must be square";
     Dtype* data = blob->mutable_cpu_data();
     int f = ceil(blob->width() / 2.);
-    float c = (2 * f - 1 - f % 2) / (2. * f);
+    Dtype c = (blob->width() - 1) / (2. * f);
     for (int i = 0; i < blob->count(); ++i) {
-      float x = i % blob->width();
-      float y = (i / blob->width()) % blob->height();
+      Dtype x = i % blob->width();
+      Dtype y = (i / blob->width()) % blob->height();
       data[i] = (1 - fabs(x / f - c)) * (1 - fabs(y / f - c));
     }
     CHECK_EQ(this->filler_param_.sparse(), -1)

3rdparty/caffe/include/caffe/layers/accuracy_layer.hpp

@@ -68,6 +68,8 @@ class AccuracyLayer : public Layer<Dtype> {
    */
   virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top);
+  virtual void Forward_gpu(const vector<Blob<Dtype>*>& bottom,
+      const vector<Blob<Dtype>*>& top);
 
 
   /// @brief Not implemented -- AccuracyLayer cannot be used as a loss.
@@ -77,6 +79,8 @@ class AccuracyLayer : public Layer<Dtype> {
       if (propagate_down[i]) { NOT_IMPLEMENTED; }
     }
   }
+  virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
+      const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);
 
   int label_axis_, outer_num_, inner_num_;
 

3rdparty/caffe/include/caffe/layers/infogain_loss_layer.hpp

@@ -13,20 +13,21 @@
 namespace caffe {
 
 /**
- * @brief A generalization of MultinomialLogisticLossLayer that takes an
+ * @brief A generalization of SoftmaxWithLossLayer that takes an
  *        "information gain" (infogain) matrix specifying the "value" of all label
  *        pairs.
  *
- * Equivalent to the MultinomialLogisticLossLayer if the infogain matrix is the
+ * Equivalent to the SoftmaxWithLossLayer if the infogain matrix is the
  * identity.
  *
  * @param bottom input Blob vector (length 2-3)
  *   -# @f$ (N \times C \times H \times W) @f$
- *      the predictions @f$ \hat{p} @f$, a Blob with values in
- *      @f$ [0, 1] @f$ indicating the predicted probability of each of the
- *      @f$ K = CHW @f$ classes.  Each prediction vector @f$ \hat{p}_n @f$
- *      should sum to 1 as in a probability distribution: @f$
- *      \forall n \sum\limits_{k=1}^K \hat{p}_{nk} = 1 @f$.
+ *      the predictions @f$ x @f$, a Blob with values in
+ *      @f$ [-\infty, +\infty] @f$ indicating the predicted score for each of
+ *      the @f$ K = CHW @f$ classes. This layer maps these scores to a
+ *      probability distribution over classes using the softmax function
+ *      @f$ \hat{p}_{nk} = \exp(x_{nk}) /
+ *      \left[\sum_{k'} \exp(x_{nk'})\right] @f$ (see SoftmaxLayer).
  *   -# @f$ (N \times 1 \times 1 \times 1) @f$
  *      the labels @f$ l @f$, an integer-valued Blob with values
  *      @f$ l_n \in [0, 1, 2, ..., K - 1] @f$
@@ -35,7 +36,7 @@ namespace caffe {
  *      (\b optional) the infogain matrix @f$ H @f$.  This must be provided as
  *      the third bottom blob input if not provided as the infogain_mat in the
  *      InfogainLossParameter. If @f$ H = I @f$, this layer is equivalent to the
- *      MultinomialLogisticLossLayer.
+ *      SoftmaxWithLossLayer.
  * @param top output Blob vector (length 1)
  *   -# @f$ (1 \times 1 \times 1 \times 1) @f$
  *      the computed infogain multinomial logistic loss: @f$ E =
@@ -98,8 +99,8 @@ class InfogainLossLayer : public LossLayer<Dtype> {
    *      infogain matrix, if provided as bottom[2])
    * @param bottom input Blob vector (length 2-3)
    *   -# @f$ (N \times C \times H \times W) @f$
-   *      the predictions @f$ \hat{p} @f$; Backward computes diff
-   *      @f$ \frac{\partial E}{\partial \hat{p}} @f$
+   *      the predictions @f$ x @f$; Backward computes diff
+   *      @f$ \frac{\partial E}{\partial x} @f$
    *   -# @f$ (N \times 1 \times 1 \times 1) @f$
    *      the labels -- ignored as we can't compute their error gradients
    *   -# @f$ (1 \times 1 \times K \times K) @f$

3rdparty/caffe/python/caffe/io.py

@@ -323,7 +323,7 @@ def resize_image(im, new_dims, interp_order=1):
             # skimage is fast but only understands {1,3} channel images
             # in [0, 1].
             im_std = (im - im_min) / (im_max - im_min)
-            resized_std = resize(im_std, new_dims, order=interp_order)
+            resized_std = resize(im_std, new_dims, order=interp_order, mode='constant')
             resized_im = resized_std * (im_max - im_min) + im_min
         else:
             # the image is a constant -- avoid divide by 0

3rdparty/caffe/python/caffe/test/test_net.py

@@ -72,41 +72,41 @@ def test_forward_backward(self):
         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=[];
+        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);
+                     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);
+        np.testing.assert_allclose(ip_blob.data,manual_forward,rtol=1e-3,atol=1e-5)
 
     def test_backward_start_end(self):
-        conv_blob=self.net.blobs['conv'];
-        ip_blob=self.net.blobs['ip_blob'];
+        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);
+        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=[];
+        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);
+                     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);
+        np.testing.assert_allclose(conv_blob.diff,manual_backward,rtol=1e-3,atol=1e-5)
 
     def test_clear_param_diffs(self):
         # Run a forward/backward step to have non-zero diffs

3rdparty/caffe/src/caffe/layers/accuracy_layer.cpp

@@ -52,8 +52,6 @@ void AccuracyLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   const Dtype* bottom_label = bottom[1]->cpu_data();
   const int dim = bottom[0]->count() / outer_num_;
   const int num_labels = bottom[0]->shape(label_axis_);
-  vector<Dtype> maxval(top_k_+1);
-  vector<int> max_id(top_k_+1);
   if (top.size() > 1) {
     caffe_set(nums_buffer_.count(), Dtype(0), nums_buffer_.mutable_cpu_data());
     caffe_set(top[1]->count(), Dtype(0), top[1]->mutable_cpu_data());
@@ -66,25 +64,22 @@ void AccuracyLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
       if (has_ignore_label_ && label_value == ignore_label_) {
         continue;
       }
-      if (top.size() > 1) ++nums_buffer_.mutable_cpu_data()[label_value];
       DCHECK_GE(label_value, 0);
       DCHECK_LT(label_value, num_labels);
+      if (top.size() > 1) ++nums_buffer_.mutable_cpu_data()[label_value];
+      const Dtype prob_of_true_class = bottom_data[i * dim
+                                                   + label_value * inner_num_
+                                                   + j];
+      int num_better_predictions = -1;  // true_class also counts as "better"
       // Top-k accuracy
-      std::vector<std::pair<Dtype, int> > bottom_data_vector;
-      for (int k = 0; k < num_labels; ++k) {
-        bottom_data_vector.push_back(std::make_pair(
-            bottom_data[i * dim + k * inner_num_ + j], k));
+      for (int k = 0; k < num_labels && num_better_predictions < top_k_; ++k) {
+        num_better_predictions +=
+          (bottom_data[i * dim + k * inner_num_ + j] >= prob_of_true_class);
       }
-      std::partial_sort(
-          bottom_data_vector.begin(), bottom_data_vector.begin() + top_k_,
-          bottom_data_vector.end(), std::greater<std::pair<Dtype, int> >());
-      // check if true label is in top k predictions
-      for (int k = 0; k < top_k_; k++) {
-        if (bottom_data_vector[k].second == label_value) {
-          ++accuracy;
-          if (top.size() > 1) ++top[1]->mutable_cpu_data()[label_value];
-          break;
-        }
+      // check if there are less than top_k_ predictions
+      if (num_better_predictions < top_k_) {
+        ++accuracy;
+        if (top.size() > 1) ++top[1]->mutable_cpu_data()[label_value];
       }
       ++count;
     }
@@ -102,6 +97,10 @@ void AccuracyLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,
   // Accuracy layer should not be used as a loss function.
 }
 
+#ifdef CPU_ONLY
+STUB_GPU(AccuracyLayer);
+#endif
+
 INSTANTIATE_CLASS(AccuracyLayer);
 REGISTER_LAYER_CLASS(Accuracy);