--- /dev/null
+// Copyright 2014 BVLC and contributors.
+
+#include <algorithm>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "google/protobuf/text_format.h"
+
+#include "gtest/gtest.h"
+#include "caffe/common.hpp"
+#include "caffe/proto/caffe.pb.h"
+#include "caffe/solver.hpp"
+
+#include "caffe/test/test_caffe_main.hpp"
+
+using std::ostringstream;
+
+namespace caffe {
+
+template <typename TypeParam>
+class SGDSolverTest : public MultiDeviceTest<TypeParam> {
+ typedef typename TypeParam::Dtype Dtype;
+
+ protected:
+ SGDSolverTest() :
+ seed_(1701), num_(5), channels_(3), height_(10), width_(10) {}
+
+ // MockSGDSolver: an SGDSolver with public history.
+ class MockSGDSolver : public SGDSolver<Dtype> {
+ public:
+ explicit MockSGDSolver(const SolverParameter& param) :
+ SGDSolver<Dtype>(param) {}
+ vector<shared_ptr<Blob<Dtype> > >& history() { return this->history_; }
+ };
+
+ shared_ptr<MockSGDSolver> solver_;
+ int seed_;
+ int num_, channels_, height_, width_;
+
+ virtual void InitSolverFromProtoString(const string& proto) {
+ SolverParameter param;
+ CHECK(google::protobuf::TextFormat::ParseFromString(proto, ¶m));
+ // Disable saving a final snapshot so the tests don't pollute the user's
+ // working directory with useless snapshots.
+ param.set_snapshot_after_train(false);
+ // Set the solver_mode according to current Caffe::mode.
+ switch (Caffe::mode()) {
+ case Caffe::CPU:
+ param.set_solver_mode(SolverParameter_SolverMode_CPU);
+ break;
+ case Caffe::GPU:
+ param.set_solver_mode(SolverParameter_SolverMode_GPU);
+ break;
+ default:
+ LOG(FATAL) << "Unknown Caffe mode: " << Caffe::mode();
+ }
+ solver_.reset(new MockSGDSolver(param));
+ }
+
+ void RunLeastSquaresSolver(const Dtype learning_rate,
+ const Dtype weight_decay, const Dtype momentum, const int num_iters) {
+ ostringstream proto;
+ proto <<
+ "max_iter: " << num_iters << " "
+ "base_lr: " << learning_rate << " "
+ "lr_policy: 'fixed' "
+ "net_param { "
+ " name: 'TestNetwork' "
+ " layers: { "
+ " name: 'data' "
+ " type: DUMMY_DATA "
+ " dummy_data_param { "
+ " num: " << num_ << " "
+ " channels: " << channels_ << " "
+ " height: " << height_ << " "
+ " width: " << width_ << " "
+ " channels: 1 "
+ " height: 1 "
+ " width: 1 "
+ " data_filler { "
+ " type: 'gaussian' "
+ " std: 1.0 "
+ " } "
+ " } "
+ " top: 'data' "
+ " top: 'targets' "
+ " } "
+ " layers: { "
+ " name: 'innerprod' "
+ " type: INNER_PRODUCT "
+ " inner_product_param { "
+ " num_output: 1 "
+ " weight_filler { "
+ " type: 'gaussian' "
+ " std: 1.0 "
+ " } "
+ " bias_filler { "
+ " type: 'gaussian' "
+ " std: 1.0 "
+ " } "
+ " } "
+ " bottom: 'data' "
+ " top: 'innerprod' "
+ " } "
+ " layers: { "
+ " name: 'loss' "
+ " type: EUCLIDEAN_LOSS "
+ " bottom: 'innerprod' "
+ " bottom: 'targets' "
+ " } "
+ "} ";
+ if (weight_decay != 0) {
+ proto << "weight_decay: " << weight_decay << " ";
+ }
+ if (momentum != 0) {
+ proto << "momentum: " << momentum << " ";
+ }
+ Caffe::set_random_seed(this->seed_);
+ this->InitSolverFromProtoString(proto.str());
+ this->solver_->Solve();
+ }
+
+ // Compute an update value given the current state of the train net,
+ // using the analytical formula for the least squares gradient.
+ // updated_params will store the updated weight and bias results,
+ // using the blobs' diffs to hold the update values themselves.
+ void ComputeLeastSquaresUpdate(const Dtype learning_rate,
+ const Dtype weight_decay, const Dtype momentum,
+ vector<shared_ptr<Blob<Dtype> > >* updated_params) {
+ const int N = num_;
+ const int D = channels_ * height_ * width_;
+
+ // Run a forward pass, and manually compute the update values from the
+ // result.
+ Net<Dtype>& net = *this->solver_->net();
+ vector<Blob<Dtype>*> empty_bottom_vec;
+ net.Forward(empty_bottom_vec);
+ ASSERT_TRUE(net.has_blob("data"));
+ const Blob<Dtype>& data = *net.blob_by_name("data");
+ ASSERT_TRUE(net.has_blob("targets"));
+ const Blob<Dtype>& targets = *net.blob_by_name("targets");
+ ASSERT_TRUE(net.has_layer("innerprod"));
+ const vector<shared_ptr<Blob<Dtype> > >& param_blobs =
+ net.layer_by_name("innerprod")->blobs();
+ const int num_param_blobs = 2;
+ ASSERT_EQ(num_param_blobs, param_blobs.size());
+ const Blob<Dtype>& weights = *param_blobs[0];
+ const Blob<Dtype>& bias = *param_blobs[1];
+ ASSERT_EQ(D * N, data.count());
+ ASSERT_EQ(N, targets.count());
+ ASSERT_EQ(D, weights.count());
+ ASSERT_EQ(1, bias.count());
+
+ updated_params->clear();
+ updated_params->resize(num_param_blobs);
+ for (int i = 0; i < num_param_blobs; ++i) {
+ (*updated_params)[i].reset(new Blob<Dtype>());
+ }
+ Blob<Dtype>& updated_weights = *(*updated_params)[0];
+ updated_weights.ReshapeLike(weights);
+ Blob<Dtype>& updated_bias = *(*updated_params)[1];
+ updated_bias.ReshapeLike(bias);
+
+ for (int i = 0; i <= D; ++i) {
+ // Compute the derivative with respect to the ith weight (i.e., the ith
+ // element of the gradient).
+ Dtype grad = 0;
+ for (int j = 0; j <= D; ++j) {
+ // Compute element (i, j) of X^T * X.
+ Dtype element = 0;
+ for (int k = 0; k < N; ++k) {
+ // (i, k) in X^T (== (k, i) in X) times (k, j) in X.
+ const Dtype element_i = (i == D) ? 1 : data.cpu_data()[k * D + i];
+ const Dtype element_j = (j == D) ? 1 : data.cpu_data()[k * D + j];
+ element += element_i * element_j;
+ }
+ if (j == D) {
+ grad += element * bias.cpu_data()[0];
+ } else {
+ grad += element * weights.cpu_data()[j];
+ }
+ }
+ for (int k = 0; k < N; ++k) {
+ const Dtype element_i = (i == D) ? 1 : data.cpu_data()[k * D + i];
+ grad -= element_i * targets.cpu_data()[k];
+ }
+ // Scale the gradient over the N samples.
+ grad /= N;
+ // Add the weight decay to the gradient.
+ grad += weight_decay *
+ ((i == D) ? bias.cpu_data()[0] : weights.cpu_data()[i]);
+ // Finally, add any momentum.
+ const vector<shared_ptr<Blob<Dtype> > >& history = solver_->history();
+ ASSERT_EQ(2, history.size()); // 1 blob for weights, 1 for bias
+ Dtype update_value = learning_rate * grad;
+ if (i == D) {
+ update_value += momentum * history[1]->cpu_data()[0];
+ updated_bias.mutable_cpu_diff()[0] = update_value;
+ updated_bias.mutable_cpu_data()[0] = bias.cpu_data()[0] - update_value;
+ } else {
+ update_value += momentum * history[0]->cpu_data()[i];
+ updated_weights.mutable_cpu_diff()[i] = update_value;
+ updated_weights.mutable_cpu_data()[i] =
+ weights.cpu_data()[i] - update_value;
+ }
+ }
+ }
+
+ void CheckLeastSquaresUpdate(
+ const vector<shared_ptr<Blob<Dtype> > >& updated_params) {
+ const int D = channels_ * height_ * width_;
+
+ const Blob<Dtype>& updated_weights = *updated_params[0];
+ const Blob<Dtype>& updated_bias = *updated_params[1];
+
+ Net<Dtype>& net = *this->solver_->net();
+ ASSERT_TRUE(net.has_layer("innerprod"));
+ const vector<shared_ptr<Blob<Dtype> > >& param_blobs =
+ net.layer_by_name("innerprod")->blobs();
+ ASSERT_EQ(2, param_blobs.size());
+ const Blob<Dtype>& solver_updated_weights = *param_blobs[0];
+ ASSERT_EQ(D, solver_updated_weights.count());
+ const double kPrecision = 1e-3;
+ const double kMinPrecision = 1e-7;
+ for (int i = 0; i < D; ++i) {
+ const Dtype expected_updated_weight = updated_weights.cpu_data()[i];
+ const Dtype solver_updated_weight = solver_updated_weights.cpu_data()[i];
+ const Dtype error_margin = std::max(kMinPrecision, kPrecision *
+ std::min(fabs(expected_updated_weight), fabs(solver_updated_weight)));
+ EXPECT_NEAR(expected_updated_weight, solver_updated_weight, error_margin);
+ }
+ const Blob<Dtype>& solver_updated_bias_blob = *param_blobs[1];
+ ASSERT_EQ(1, solver_updated_bias_blob.count());
+ const Dtype expected_updated_bias = updated_bias.cpu_data()[0];
+ const Dtype solver_updated_bias = solver_updated_bias_blob.cpu_data()[0];
+ const Dtype error_margin = std::max(kMinPrecision, kPrecision *
+ std::min(fabs(expected_updated_bias), fabs(solver_updated_bias)));
+ EXPECT_NEAR(expected_updated_bias, solver_updated_bias, error_margin);
+
+ // Check the solver's history -- should contain the previous update value.
+ vector<shared_ptr<Blob<Dtype> > >& history = this->solver_->history();
+ ASSERT_EQ(2, history.size());
+ for (int i = 0; i < D; ++i) {
+ const Dtype expected_history = updated_weights.cpu_diff()[i];
+ const Dtype solver_history = history[0]->cpu_data()[i];
+ const Dtype error_margin_hist = std::max(kMinPrecision, kPrecision *
+ std::min(fabs(expected_history), fabs(solver_history)));
+ EXPECT_NEAR(expected_history, solver_history, error_margin_hist);
+ }
+ const Dtype expected_history = updated_bias.cpu_diff()[0];
+ const Dtype solver_history = history[1]->cpu_data()[0];
+ const Dtype error_margin_hist = std::max(kMinPrecision, kPrecision *
+ std::min(fabs(expected_history), fabs(solver_history)));
+ EXPECT_NEAR(expected_history, solver_history, error_margin_hist);
+ }
+
+ // Test that the correct update is computed for a regularized least squares
+ // problem:
+ //
+ // E = (1/(2n)) || X w - y ||^2 + (lambda / 2) || w ||^2
+ // \nabla_w E = (1/n) (X^T X w - X^T y) + lambda * w
+ //
+ // X \in R^{n x (d+1)} (each example is a row, (d+1)th element is always 1)
+ // w \in R^{(d+1) x 1} ((d+1)th element is the bias)
+ // y \in R^{n x 1}
+ // lambda is weight_decay
+ //
+ // TestLeastSquaresUpdate works "inductively", assuming that the solver
+ // correctly updates the net K (= iter_to_check) times, then given the history
+ // from the Kth update, we compute the (K+1)th update and check that it
+ // matches the solver's (K+1)th update.
+ void TestLeastSquaresUpdate(const Dtype learning_rate = 1.0,
+ const Dtype weight_decay = 0.0, const Dtype momentum = 0.0,
+ const int iter_to_check = 0) {
+ // Initialize the solver and run K (= iter_to_check) solver iterations.
+ RunLeastSquaresSolver(learning_rate, weight_decay, momentum, iter_to_check);
+
+ // Compute the (K+1)th update using the analytic least squares gradient.
+ vector<shared_ptr<Blob<Dtype> > > updated_params;
+ ComputeLeastSquaresUpdate(learning_rate, weight_decay, momentum,
+ &updated_params);
+
+ // Reinitialize the solver and run K+1 solver iterations.
+ RunLeastSquaresSolver(learning_rate, weight_decay, momentum,
+ iter_to_check + 1);
+
+ // Check that the solver's solution matches ours.
+ CheckLeastSquaresUpdate(updated_params);
+ }
+};
+
+TYPED_TEST_CASE(SGDSolverTest, TestDtypesAndDevices);
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdate) {
+ typedef typename TypeParam::Dtype Dtype;
+ this->TestLeastSquaresUpdate();
+}
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdateLROneTenth) {
+ typedef typename TypeParam::Dtype Dtype;
+ const Dtype kLearningRate = 0.1;
+ this->TestLeastSquaresUpdate(kLearningRate);
+}
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdateWithWeightDecay) {
+ typedef typename TypeParam::Dtype Dtype;
+ const Dtype kLearningRate = 1.0;
+ const Dtype kWeightDecay = 0.5;
+ this->TestLeastSquaresUpdate(kLearningRate, kWeightDecay);
+}
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdateWithMomentum) {
+ typedef typename TypeParam::Dtype Dtype;
+ const Dtype kLearningRate = 1.0;
+ const Dtype kWeightDecay = 0.0;
+ const Dtype kMomentum = 0.5;
+ const int kNumIters = 1;
+ for (int i = 0; i <= kNumIters; ++i) {
+ this->TestLeastSquaresUpdate(kLearningRate, kWeightDecay, kMomentum, i);
+ }
+}
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdateWithMomentumMultiIter) {
+ typedef typename TypeParam::Dtype Dtype;
+ const Dtype kLearningRate = 1.0;
+ const Dtype kWeightDecay = 0.0;
+ const Dtype kMomentum = 0.5;
+ const int kNumIters = 5;
+ for (int i = 0; i <= kNumIters; ++i) {
+ this->TestLeastSquaresUpdate(kLearningRate, kWeightDecay, kMomentum, i);
+ }
+}
+
+TYPED_TEST(SGDSolverTest, TestLeastSquaresUpdateWithEverything) {
+ typedef typename TypeParam::Dtype Dtype;
+ const Dtype kLearningRate = 0.01;
+ const Dtype kWeightDecay = 0.1;
+ const Dtype kMomentum = 0.9;
+ const int kNumIters = 5;
+ for (int i = 0; i <= kNumIters; ++i) {
+ this->TestLeastSquaresUpdate(kLearningRate, kWeightDecay, kMomentum, i);
+ }
+}
+
+} // namespace caffe