resume training process

[kilmarnock]

Hi!
Is it (currently / in the future / generally spoken) possible to resume the training process of fast-dqn-caffe?

I assumed that giving a -model model/dqn_iter_1000000.caffemodel would resume training? I see a dqn_iter_1000000.solverstate but no option in the ./src/fast_dqn_main.cpp file. -solver has to point to a .prototxt file.

Thank you for the implementation.

[watts4speed]

Hi Kilmarnock,

It's possible but not implemented. That's a good feature to have. :-)

-Peter

On Sat, Jan 23, 2016 at 3:14 AM, kilmarnock [email protected]
wrote:

Hi!
Is it (currently / in the future / generally spoken) possible to resume
the training process of fast-dqn-caffe?

I assumed that giving a -model model/dqn_iter_1000000caffemodel would
resume training? I see a dqn_iter_1000000solverstate but no option in the
/src/fast_dqn_maincpp file -solver has to point to a prototxt file

Thank you for the implementation

—
Reply to this email directly or view it on GitHub
#4.

[kilmarnock]

I have got it resuming using a piece of code like this:

void Fast_DQN::Initialize() {

  // Initialize dummy input data with 0
  std::fill(dummy_input_data_.begin(), dummy_input_data_.end(), 0.0);

  // Initialize net and solver
  caffe::SolverParameter solver_param;
  caffe::ReadProtoFromTextFileOrDie(solver_param_, &solver_param);

  solver_.reset(caffe::GetSolver<float>(solver_param));

  std::string fileName;
  fileName  = "./model/dqn_iter_1000000.solverstate";
  const char* resume_file;
  resume_file = fileName.c_str();
  solver_->Restore(resume_file);

  //http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1Solver.html

I would write a "take the highest iteration count in the model directory", or "parse the solverstate from command line" but cpp is not mine. And I have never been here.

[watts4speed]

another thing to consider is snapshotting the replay_memory_. The network
trains correctly because of the experience built up in the replay memory.
To correctly resume restoring that state is also important. I think
github.com/mhauskn/recurrent-caffe has code to save and restore the replay
memory.

On Sat, Jan 23, 2016 at 1:31 PM, kilmarnock [email protected]
wrote:

I got it resuming using a piece of code like this:

void Fast_DQN::Initialize() {

// Initialize dummy input data with 0
std::fill(dummy_input_data_.begin(), dummy_input_data_.end(), 0.0);

// Initialize net and solver
caffe::SolverParameter solver_param;
caffe::ReadProtoFromTextFileOrDie(solver_param_, &solver_param);

solver_.reset(caffe::GetSolver(solver_param));

std::string fileName;
fileName = "./model/dqn_iter_1000000.solverstate";
const char* resume_file;
resume_file = fileName.c_str();
solver_->Restore(resume_file);

//http://caffe.berkeleyvision.org/doxygen/classcaffe_1_1Solver.html

I would write a "take the highest iteration count in the model directory",
or "parse the solverstate from command line" but cpp is not mine. And I
have never been here.

—
Reply to this email directly or view it on GitHub
#4 (comment)
.

[kilmarnock]

I had a look at Mr. Hausknechts repository https://github.com/mhauskn/dqn in the recurrent branch, loading and storing replay memory is implemented there in full beauty.