Display averaged loss over the last several iterations

diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index 493bfa7..ba4c41c 100644 (file)
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -63,7 +63,7 @@ message NetParameter {
 // NOTE
 // Update the next available ID when you add a new SolverParameter field.
 //
-// SolverParameter next available ID: 33 (last added: test_initialization)
+// SolverParameter next available ID: 34 (last added: average_loss)
 message SolverParameter {
   //////////////////////////////////////////////////////////////////////////////
   // Specifying the train and test networks
@@ -113,6 +113,8 @@ message SolverParameter {
   // the number of iterations between displaying info. If display = 0, no info
   // will be displayed.
   optional int32 display = 6;
+  // Display the cost averaged over the last average_cost iterations
+  optional int32 average_loss = 33 [default = 1];
   optional int32 max_iter = 7; // the maximum number of iterations
   optional string lr_policy = 8; // The learning rate decay policy.
   optional float gamma = 9; // The parameter to compute the learning rate.

diff --git a/src/caffe/solver.cpp b/src/caffe/solver.cpp
index 0ea4edc..0810f48 100644 (file)
--- a/src/caffe/solver.cpp
+++ b/src/caffe/solver.cpp
@@ -169,6 +169,13 @@ void Solver<Dtype>::Solve(const char* resume_file) {
   // resume_file above.
   const int start_iter = iter_;
 
+  int average_loss = this->param_.average_loss();
+
+  CHECK_GE(average_loss, 1) << "average_cost should be non-negative.";
+
+  vector<Dtype> losses;
+  Dtype smoothed_loss;
+
   // For a network that is trained by the solver, no bottom or top vecs
   // should be given, and we will just provide dummy vecs.
   vector<Blob<Dtype>*> bottom_vec;
@@ -187,8 +194,17 @@ void Solver<Dtype>::Solve(const char* resume_file) {
     const bool display = param_.display() && iter_ % param_.display() == 0;
     net_->set_debug_info(display && param_.debug_info());
     Dtype loss = net_->ForwardBackward(bottom_vec);
+    if (losses.size() < average_loss) {
+      losses.push_back(loss);
+      int size = losses.size();
+      smoothed_loss = (smoothed_loss * (size - 1) + loss) / size;
+    } else {
+      int idx = (iter_ - start_iter) % average_loss;
+      smoothed_loss += (loss - losses[idx]) / average_loss;
+      losses[idx] = loss;
+    }
     if (display) {
-      LOG(INFO) << "Iteration " << iter_ << ", loss = " << loss;
+      LOG(INFO) << "Iteration " << iter_ << ", loss = " << smoothed_loss;
       const vector<Blob<Dtype>*>& result = net_->output_blobs();
       int score_index = 0;
       for (int j = 0; j < result.size(); ++j) {