Merge pull request #5184 from shaibagon/fix_batch_norm_param_upgrade

[platform/upstream/caffeonacl.git] / include / caffe / filler.hpp

diff --git a/include/caffe/filler.hpp b/include/caffe/filler.hpp
index 0125b30..dad9ad4 100644 (file)
--- a/include/caffe/filler.hpp
+++ b/include/caffe/filler.hpp
@@ -8,7 +8,6 @@
 #include <string>
 
 #include "caffe/blob.hpp"
-#include "caffe/common.hpp"
 #include "caffe/proto/caffe.pb.h"
 #include "caffe/syncedmem.hpp"
 #include "caffe/util/math_functions.hpp"
@@ -79,9 +78,8 @@ class GaussianFiller : public Filler<Dtype> {
       // These have num == channels == 1; width is number of inputs; height is
       // number of outputs.  The 'sparse' variable specifies the mean number
       // of non-zero input weights for a given output.
-      CHECK_EQ(blob->num(), 1);
-      CHECK_EQ(blob->channels(), 1);
-      int num_outputs = blob->height();
+      CHECK_GE(blob->num_axes(), 1);
+      const int num_outputs = blob->shape(0);
       Dtype non_zero_probability = Dtype(sparse) / Dtype(num_outputs);
       rand_vec_.reset(new SyncedMemory(blob->count() * sizeof(int)));
       int* mask = reinterpret_cast<int*>(rand_vec_->mutable_cpu_data());
@@ -175,6 +173,9 @@ class XavierFiller : public Filler<Dtype> {
  * A Filler based on the paper [He, Zhang, Ren and Sun 2015]: Specifically
  * accounts for ReLU nonlinearities.
  *
+ * Aside: for another perspective on the scaling factor, see the derivation of
+ * [Saxe, McClelland, and Ganguli 2013 (v3)].
+ *
  * It fills the incoming matrix by randomly sampling Gaussian data with std =
  * sqrt(2 / n) where n is the fan_in, fan_out, or their average, depending on
  * the variance_norm option. You should make sure the input blob has shape (num,
@@ -206,6 +207,60 @@ class MSRAFiller : public Filler<Dtype> {
   }
 };
 
+/*!
+@brief Fills a Blob with coefficients for bilinear interpolation.
+
+A common use case is with the DeconvolutionLayer acting as upsampling.
+You can upsample a feature map with shape of (B, C, H, W) by any integer factor
+using the following proto.
+\code
+layer {
+  name: "upsample", type: "Deconvolution"
+  bottom: "{{bottom_name}}" top: "{{top_name}}"
+  convolution_param {
+    kernel_size: {{2 * factor - factor % 2}} stride: {{factor}}
+    num_output: {{C}} group: {{C}}
+    pad: {{ceil((factor - 1) / 2.)}}
+    weight_filler: { type: "bilinear" } bias_term: false
+  }
+  param { lr_mult: 0 decay_mult: 0 }
+}
+\endcode
+Please use this by replacing `{{}}` with your values. By specifying
+`num_output: {{C}} group: {{C}}`, it behaves as
+channel-wise convolution. The filter shape of this deconvolution layer will be
+(C, 1, K, K) where K is `kernel_size`, and this filler will set a (K, K)
+interpolation kernel for every channel of the filter identically. The resulting
+shape of the top feature map will be (B, C, factor * H, factor * W).
+Note that the learning rate and the
+weight decay are set to 0 in order to keep coefficient values of bilinear
+interpolation unchanged during training. If you apply this to an image, this
+operation is equivalent to the following call in Python with Scikit.Image.
+\code{.py}
+out = skimage.transform.rescale(img, factor, mode='constant', cval=0)
+\endcode
+ */
+template <typename Dtype>
+class BilinearFiller : public Filler<Dtype> {
+ public:
+  explicit BilinearFiller(const FillerParameter& param)
+      : Filler<Dtype>(param) {}
+  virtual void Fill(Blob<Dtype>* blob) {
+    CHECK_EQ(blob->num_axes(), 4) << "Blob must be 4 dim.";
+    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);
+    for (int i = 0; i < blob->count(); ++i) {
+      float x = i % blob->width();
+      float 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)
+         << "Sparsity not supported by this Filler.";
+  }
+};
+
 /**
  * @brief Get a specific filler from the specification given in FillerParameter.
  *
@@ -227,6 +282,8 @@ Filler<Dtype>* GetFiller(const FillerParameter& param) {
     return new XavierFiller<Dtype>(param);
   } else if (type == "msra") {
     return new MSRAFiller<Dtype>(param);
+  } else if (type == "bilinear") {
+    return new BilinearFiller<Dtype>(param);
   } else {
     CHECK(false) << "Unknown filler name: " << param.type();
   }