1 // Fillers are random number generators that fills a blob using the specified
2 // algorithm. The expectation is that they are only going to be used during
3 // initialization time and will not involve any GPUs.
5 #ifndef CAFFE_FILLER_HPP
6 #define CAFFE_FILLER_HPP
10 #include "caffe/blob.hpp"
11 #include "caffe/common.hpp"
12 #include "caffe/proto/caffe.pb.h"
13 #include "caffe/syncedmem.hpp"
14 #include "caffe/util/math_functions.hpp"
18 /// @brief Fills a Blob with constant or randomly-generated data.
19 template <typename Dtype>
22 explicit Filler(const FillerParameter& param) : filler_param_(param) {}
24 virtual void Fill(Blob<Dtype>* blob) = 0;
26 FillerParameter filler_param_;
30 /// @brief Fills a Blob with constant values @f$ x = 0 @f$.
31 template <typename Dtype>
32 class ConstantFiller : public Filler<Dtype> {
34 explicit ConstantFiller(const FillerParameter& param)
35 : Filler<Dtype>(param) {}
36 virtual void Fill(Blob<Dtype>* blob) {
37 Dtype* data = blob->mutable_cpu_data();
38 const int count = blob->count();
39 const Dtype value = this->filler_param_.value();
41 for (int i = 0; i < count; ++i) {
44 CHECK_EQ(this->filler_param_.sparse(), -1)
45 << "Sparsity not supported by this Filler.";
49 /// @brief Fills a Blob with uniformly distributed values @f$ x\sim U(a, b) @f$.
50 template <typename Dtype>
51 class UniformFiller : public Filler<Dtype> {
53 explicit UniformFiller(const FillerParameter& param)
54 : Filler<Dtype>(param) {}
55 virtual void Fill(Blob<Dtype>* blob) {
57 caffe_rng_uniform<Dtype>(blob->count(), Dtype(this->filler_param_.min()),
58 Dtype(this->filler_param_.max()), blob->mutable_cpu_data());
59 CHECK_EQ(this->filler_param_.sparse(), -1)
60 << "Sparsity not supported by this Filler.";
64 /// @brief Fills a Blob with Gaussian-distributed values @f$ x = a @f$.
65 template <typename Dtype>
66 class GaussianFiller : public Filler<Dtype> {
68 explicit GaussianFiller(const FillerParameter& param)
69 : Filler<Dtype>(param) {}
70 virtual void Fill(Blob<Dtype>* blob) {
71 Dtype* data = blob->mutable_cpu_data();
73 caffe_rng_gaussian<Dtype>(blob->count(), Dtype(this->filler_param_.mean()),
74 Dtype(this->filler_param_.std()), blob->mutable_cpu_data());
75 int sparse = this->filler_param_.sparse();
78 // Sparse initialization is implemented for "weight" blobs; i.e. matrices.
79 // These have num == channels == 1; width is number of inputs; height is
80 // number of outputs. The 'sparse' variable specifies the mean number
81 // of non-zero input weights for a given output.
82 CHECK_GE(blob->num_axes(), 1);
83 const int num_outputs = blob->shape(0);
84 Dtype non_zero_probability = Dtype(sparse) / Dtype(num_outputs);
85 rand_vec_.reset(new SyncedMemory(blob->count() * sizeof(int)));
86 int* mask = reinterpret_cast<int*>(rand_vec_->mutable_cpu_data());
87 caffe_rng_bernoulli(blob->count(), non_zero_probability, mask);
88 for (int i = 0; i < blob->count(); ++i) {
95 shared_ptr<SyncedMemory> rand_vec_;
98 /** @brief Fills a Blob with values @f$ x \in [0, 1] @f$
99 * such that @f$ \forall i \sum_j x_{ij} = 1 @f$.
101 template <typename Dtype>
102 class PositiveUnitballFiller : public Filler<Dtype> {
104 explicit PositiveUnitballFiller(const FillerParameter& param)
105 : Filler<Dtype>(param) {}
106 virtual void Fill(Blob<Dtype>* blob) {
107 Dtype* data = blob->mutable_cpu_data();
108 DCHECK(blob->count());
109 caffe_rng_uniform<Dtype>(blob->count(), 0, 1, blob->mutable_cpu_data());
110 // We expect the filler to not be called very frequently, so we will
111 // just use a simple implementation
112 int dim = blob->count() / blob->num();
114 for (int i = 0; i < blob->num(); ++i) {
116 for (int j = 0; j < dim; ++j) {
117 sum += data[i * dim + j];
119 for (int j = 0; j < dim; ++j) {
120 data[i * dim + j] /= sum;
123 CHECK_EQ(this->filler_param_.sparse(), -1)
124 << "Sparsity not supported by this Filler.";
129 * @brief Fills a Blob with values @f$ x \sim U(-a, +a) @f$ where @f$ a @f$ is
130 * set inversely proportional to number of incoming nodes, outgoing
131 * nodes, or their average.
133 * A Filler based on the paper [Bengio and Glorot 2010]: Understanding
134 * the difficulty of training deep feedforward neuralnetworks.
136 * It fills the incoming matrix by randomly sampling uniform data from [-scale,
137 * scale] where scale = sqrt(3 / n) where n is the fan_in, fan_out, or their
138 * average, depending on the variance_norm option. You should make sure the
139 * input blob has shape (num, a, b, c) where a * b * c = fan_in and num * b * c
140 * = fan_out. Note that this is currently not the case for inner product layers.
142 * TODO(dox): make notation in above comment consistent with rest & use LaTeX.
144 template <typename Dtype>
145 class XavierFiller : public Filler<Dtype> {
147 explicit XavierFiller(const FillerParameter& param)
148 : Filler<Dtype>(param) {}
149 virtual void Fill(Blob<Dtype>* blob) {
150 CHECK(blob->count());
151 int fan_in = blob->count() / blob->num();
152 int fan_out = blob->count() / blob->channels();
153 Dtype n = fan_in; // default to fan_in
154 if (this->filler_param_.variance_norm() ==
155 FillerParameter_VarianceNorm_AVERAGE) {
156 n = (fan_in + fan_out) / Dtype(2);
157 } else if (this->filler_param_.variance_norm() ==
158 FillerParameter_VarianceNorm_FAN_OUT) {
161 Dtype scale = sqrt(Dtype(3) / n);
162 caffe_rng_uniform<Dtype>(blob->count(), -scale, scale,
163 blob->mutable_cpu_data());
164 CHECK_EQ(this->filler_param_.sparse(), -1)
165 << "Sparsity not supported by this Filler.";
170 * @brief Fills a Blob with values @f$ x \sim N(0, \sigma^2) @f$ where
171 * @f$ \sigma^2 @f$ is set inversely proportional to number of incoming
172 * nodes, outgoing nodes, or their average.
174 * A Filler based on the paper [He, Zhang, Ren and Sun 2015]: Specifically
175 * accounts for ReLU nonlinearities.
177 * Aside: for another perspective on the scaling factor, see the derivation of
178 * [Saxe, McClelland, and Ganguli 2013 (v3)].
180 * It fills the incoming matrix by randomly sampling Gaussian data with std =
181 * sqrt(2 / n) where n is the fan_in, fan_out, or their average, depending on
182 * the variance_norm option. You should make sure the input blob has shape (num,
183 * a, b, c) where a * b * c = fan_in and num * b * c = fan_out. Note that this
184 * is currently not the case for inner product layers.
186 template <typename Dtype>
187 class MSRAFiller : public Filler<Dtype> {
189 explicit MSRAFiller(const FillerParameter& param)
190 : Filler<Dtype>(param) {}
191 virtual void Fill(Blob<Dtype>* blob) {
192 CHECK(blob->count());
193 int fan_in = blob->count() / blob->num();
194 int fan_out = blob->count() / blob->channels();
195 Dtype n = fan_in; // default to fan_in
196 if (this->filler_param_.variance_norm() ==
197 FillerParameter_VarianceNorm_AVERAGE) {
198 n = (fan_in + fan_out) / Dtype(2);
199 } else if (this->filler_param_.variance_norm() ==
200 FillerParameter_VarianceNorm_FAN_OUT) {
203 Dtype std = sqrt(Dtype(2) / n);
204 caffe_rng_gaussian<Dtype>(blob->count(), Dtype(0), std,
205 blob->mutable_cpu_data());
206 CHECK_EQ(this->filler_param_.sparse(), -1)
207 << "Sparsity not supported by this Filler.";
212 * @brief Get a specific filler from the specification given in FillerParameter.
214 * Ideally this would be replaced by a factory pattern, but we will leave it
217 template <typename Dtype>
218 Filler<Dtype>* GetFiller(const FillerParameter& param) {
219 const std::string& type = param.type();
220 if (type == "constant") {
221 return new ConstantFiller<Dtype>(param);
222 } else if (type == "gaussian") {
223 return new GaussianFiller<Dtype>(param);
224 } else if (type == "positive_unitball") {
225 return new PositiveUnitballFiller<Dtype>(param);
226 } else if (type == "uniform") {
227 return new UniformFiller<Dtype>(param);
228 } else if (type == "xavier") {
229 return new XavierFiller<Dtype>(param);
230 } else if (type == "msra") {
231 return new MSRAFiller<Dtype>(param);
233 CHECK(false) << "Unknown filler name: " << param.type();
235 return (Filler<Dtype>*)(NULL);
240 #endif // CAFFE_FILLER_HPP_