distributed solver - not checked thoroughly

diff --git a/include/caffe/distributed_solver.hpp b/include/caffe/distributed_solver.hpp
new file mode 100644 (file)
index 0000000..4e08405
--- /dev/null
+++ b/include/caffe/distributed_solver.hpp
@@ -0,0 +1,57 @@
+// Copyright Yangqing Jia 2013
+// This implements the distributed solver in two classes: the parameter server
+// that holds the parameters and the actual solver that does computation.
+
+// Copyright Yangqing Jia 2013
+
+#ifndef CAFFE_OPTIMIZATION_DISTRIBUTED_SOLVER_HPP_
+#define CAFFE_OPTIMIZATION_DISTRIBUTED_SOLVER_HPP_
+
+#include <vector>
+
+#include "caffe/solver.hpp"
+
+namespace caffe {
+
+template <typename Dtype>
+class DistributedSolverParamServer : public Solver<Dtype> {
+ public:
+  DistributedSolverParamServer(const SolverParameter& param)
+      : Solver<Dtype>(param) {}
+  virtual void Solve(const char* resume_file = NULL);
+
+ protected:
+  // The distributed solver does not do solving itself.
+  virtual void ComputeUpdateValue() {}
+  // The distributed solver has nothing to snapshot.
+  virtual void SnapshotSolverState(SolverState* state) {}
+  virtual void RestoreSolverState(const SolverState& state) {}
+  // The function that implements the actual communication.
+  void ReceiveAndSend();
+
+  int next_snapshot_;
+};
+
+
+template <typename Dtype>
+class DistributedSolverParamClient : public SGDSolver<Dtype> {
+ public:
+  DistributedSolverParamClient(const SolverParameter& param)
+      : SGDSolver<Dtype>(param) {
+    CHECK_GT(param.communication_interval(), 0);
+    next_send_iter_ = param.communication_interval();
+  }
+  virtual void Solve(const char* resume_file = NULL);
+ protected:
+  virtual void PreSolve() { SGDSolver<Dtype>::PreSolve(); }
+  virtual void ComputeUpdateValue();
+  void SendAndReceive(bool just_receive = false);
+
+  int next_send_iter_;
+  int next_display_;
+};
+
+
+}  // namespace caffe
+
+#endif  // CAFFE_OPTIMIZATION_DISTRIBUTED_SOLVER_HPP_

diff --git a/src/caffe/distributed_solver.cpp b/src/caffe/distributed_solver.cpp
new file mode 100644 (file)
index 0000000..5e13c54
--- /dev/null
+++ b/src/caffe/distributed_solver.cpp
@@ -0,0 +1,199 @@
+// Copyright Yangqing Jia 2013
+
+#include <cstdio>
+
+#include <algorithm>
+#include <string>
+#include <vector>
+#include <boost/asio.hpp>
+
+#include "caffe/net.hpp"
+#include "caffe/proto/caffe.pb.h"
+#include "caffe/distributed_solver.hpp"
+#include "caffe/util/io.hpp"
+#include "caffe/util/math_functions.hpp"
+
+using boost::asio::ip::tcp;
+
+
+namespace caffe {
+
+template <typename Dtype>
+void DistributedSolverParamServer<Dtype>::Solve(const char* resume_file) {
+  Caffe::set_phase(Caffe::TRAIN);
+  LOG(INFO) << "Solving " << this->net_->name();
+
+  this->iter_ = 0;
+  if (resume_file) {
+    LOG(INFO) << "Restoring previous solver status from " << resume_file;
+    Solver<Dtype>::Restore(resume_file);
+  }
+
+  // the main loop.
+  while (this->iter_ < this->param_.max_iter()) {
+    ReceiveAndSend();
+    // Check if we need to do snapshot
+    if (this->param_.snapshot() && this->iter_ > next_snapshot_) {
+      Solver<Dtype>::Snapshot();
+      next_snapshot_ += this->param_.snapshot();
+    }
+    // TODO: test
+  }
+  LOG(INFO) << "Optimization Done.";
+}
+
+
+// Receive and send: what this function does is to get the accumulated gradient
+// values from the client, stores it to the diff field of the network, and then
+// updates the network. It then sends back the updated network value to the
+// client.
+template <typename Dtype>
+void DistributedSolverParamServer<Dtype>::ReceiveAndSend() {
+  bool send_only;
+  int incoming_iter;
+
+  boost::asio::io_service io_s;
+  tcp::acceptor data_acceptor(
+      io_s, tcp::endpoint(tcp::v4(), atoi(this->param_.tcp_port().c_str())));
+  tcp::iostream data_stream;
+  data_acceptor.accept(*(data_stream.rdbuf()));
+  data_stream >> send_only;
+  vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
+  if (!send_only) {
+    // Receive data
+    data_stream >> incoming_iter;
+    for (int param_id = 0; param_id < net_params.size(); ++param_id) {
+      Dtype* param_diff = net_params[param_id]->mutable_cpu_diff();
+      int count = net_params[param_id]->count();
+      for (int i = 0; i < count; ++i) {
+        data_stream >> param_diff[i];
+      }
+    }
+    // Check Error
+    if (!data_stream) {
+      LOG(ERROR) << "Error in receiving.";
+    } else {
+      // If the read is successful, update the network.
+      this->iter_ += incoming_iter;
+      this->net_->Update();
+    }
+  }
+  // Send data
+  data_stream << this->iter_;
+  for (int param_id = 0; param_id < net_params.size(); ++param_id) {
+    const Dtype* param_data = net_params[param_id]->cpu_data();
+    int count = net_params[param_id]->count();
+    for (int i = 0; i < count; ++i) {
+      data_stream << param_data[i];
+    }
+  }
+  data_stream.flush();
+  if (!data_stream) {
+    LOG(ERROR) << "Error in sending.";
+  }
+  data_stream.close();
+}
+
+
+template <typename Dtype>
+void DistributedSolverParamClient<Dtype>::Solve(const char* resume_file) {
+  // Although we have resume_file, the client never does the actual resuming.
+  // Instead, it will simply request the weights from the server.
+  Caffe::set_phase(Caffe::TRAIN);
+  LOG(INFO) << "Solving " << this->net_->name();
+  PreSolve();
+
+  // Send and receive once to get the current iteration and the parameters
+  SendAndReceive(true);
+
+  // 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;
+  while (this->iter_++ < this->param_.max_iter()) {
+    Dtype loss = this->net_->ForwardBackward(bottom_vec);
+    ComputeUpdateValue();
+    this->net_->Update();
+
+    if (this->param_.display() && this->iter_ > next_display_) {
+      LOG(INFO) << "Iteration " << this->iter_ << ", loss = " << loss;
+      next_display_ += this->param_.display();
+    }
+  }
+  LOG(INFO) << "Optimization Done.";
+}
+
+
+template <typename Dtype>
+void DistributedSolverParamClient<Dtype>::SendAndReceive(bool receive_only) {
+  tcp::iostream data_stream(this->param_.tcp_server(), this->param_.tcp_port());
+  CHECK(data_stream) << "Error in connection.";
+  data_stream << receive_only;
+  if (!receive_only) {
+    data_stream << this->iter_;
+    // TODO: send the accumulated gradient stored at history_, and set it to
+    // zero for future accumulation
+    for (int param_id = 0; param_id < this->history_.size(); ++param_id) {
+      Dtype* accum_history_data = this->history_[param_id]->mutable_cpu_diff();
+      int count = this->history_[param_id]->count();
+      for (int i = 0; i < count; ++i) {
+        data_stream << accum_history_data[i];
+        accum_history_data[i] = 0;
+      }
+    }
+  }
+  data_stream.flush();
+  // Receive parameters
+  data_stream >> this->iter_;
+  vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
+  for (int param_id = 0; param_id < net_params.size(); ++param_id) {
+    Dtype* param_data = net_params[param_id]->mutable_cpu_data();
+    int count = net_params[param_id]->count();
+    for (int i = 0; i < count; ++i) {
+      data_stream >> param_data[i];
+    }
+    // Also, let's set the param_diff to be zero so that this update does not
+    // change the parameter value, since it has already been updated.
+    memset(net_params[param_id]->mutable_cpu_diff(), 0,
+        net_params[param_id]->count() * sizeof(Dtype));
+  }
+}
+
+
+template <typename Dtype>
+void DistributedSolverParamClient<Dtype>::ComputeUpdateValue() {
+  // First, carry out the normal update
+  SGDSolver<Dtype>::ComputeUpdateValue();
+  vector<shared_ptr<Blob<Dtype> > >& net_params = this->net_->params();
+  // Accumulate the gradient history
+  switch (Caffe::mode()) {
+  case Caffe::CPU:
+    for (int param_id = 0; param_id < net_params.size(); ++param_id) {
+      caffe_axpy(net_params[param_id]->count(), Dtype(1.),
+          net_params[param_id]->cpu_diff(),
+          this->history_[param_id]->mutable_cpu_diff());
+    }
+    break;
+  case Caffe::GPU:
+    for (int param_id = 0; param_id < net_params.size(); ++param_id) {
+      caffe_gpu_axpy(net_params[param_id]->count(), Dtype(1.),
+          net_params[param_id]->gpu_diff(),
+          this->history_[param_id]->mutable_gpu_diff());
+    }
+    break;
+  default:
+    LOG(FATAL) << "Unknown caffe mode.";
+  }
+  // See if we need to do communication.
+  if (this->iter_ > next_send_iter_) {
+    DLOG(INFO) << "Send and receive parameters.";
+    SendAndReceive();
+    next_send_iter_ += this->param_.communication_interval();
+  }
+}
+
+
+INSTANTIATE_CLASS(DistributedSolverParamServer);
+INSTANTIATE_CLASS(DistributedSolverParamClient);
+
+
+}  // namespace caffe

diff --git a/src/caffe/proto/caffe.proto b/src/caffe/proto/caffe.proto
index c314070..c2d301f 100644 (file)
--- a/src/caffe/proto/caffe.proto
+++ b/src/caffe/proto/caffe.proto
@@ -116,6 +116,11 @@ message SolverParameter {
   // whether to snapshot diff in the results or not. Snapshotting diff will help
   // debugging but the final protocol buffer size will be much larger.
   optional bool snapshot_diff = 16 [ default = false];
+
+  // the asynchronous solver stuff
+  optional string tcp_port = 20 [ default = "20901"];
+  optional string tcp_server = 21;
+  optional int32 communication_interval = 22;
 }
 
 // A message that stores the solver snapshots