From 132fb500a0f7c99b8a2301b575c3f6416faf488e Mon Sep 17 00:00:00 2001
From: Marcin Kirsz <mkirsz@ed.ac.uk>
Date: Tue, 22 Oct 2024 11:02:50 +0100
Subject: [PATCH] Update
---
bin/tadah_cli.cpp | 145 ++++++++++----------------------------
examples/ex_0/targets_new | 34 ++++-----
2 files changed, 54 insertions(+), 125 deletions(-)
diff --git a/bin/tadah_cli.cpp b/bin/tadah_cli.cpp
index 8a42413..f88d560 100644
--- a/bin/tadah_cli.cpp
+++ b/bin/tadah_cli.cpp
@@ -15,6 +15,7 @@
#include "../MLIP/trainer.h"
#ifdef TADAH_ENABLE_HPO
#include "../HPO/hpo.h"
+#include "../HPO/hpo_worker.h"
#endif
#include <iostream>
@@ -68,43 +69,13 @@ void TadahCLI::subcommand_train() {
CLI::Timer timer_tot{"Training", CLI::Timer::Big};
// Initialize host MPI trainer
- MPI_Trainer_HOST HOST(config, rank, ncpu);
- HOST.prep_wpckgs();
-
- // Process requests from workers
- while (true) {
- // Exit loop if there are no more packages
- if (!HOST.has_packages()) break;
-
- // Probe for incoming requests from any worker
- HOST.probe();
-
- // Handle requests based on their tags
- if (HOST.tag == MPI_Trainer::WORK_TAG) {
- HOST.work_tag();
- } else if (HOST.tag == MPI_Trainer::DATA_TAG) {
- HOST.data_tag();
- } else {
- throw std::runtime_error("HOST: Unexpected request from "
- + std::to_string(HOST.worker));
- }
- }
-
- // Collect remaining data and release all workers
- int count = 1;
- while (true) {
- HOST.probe();
- if (HOST.tag == MPI_Trainer::DATA_TAG) {
- HOST.data_tag();
- } else {
- HOST.release_tag(count);
- }
- if (count == ncpu) break; // Exit when all workers are released
- }
+ TrainerHost host(config, rank, ncpu);
+ host.run();
// Perform the final computation and save parameters
- HOST.solve();
- Config param_file = HOST.model->get_param_file();
+ host.solve();
+
+ Config param_file = host.model->get_param_file();
param_file.check_pot_file();
std::ofstream outfile("pot.tadah");
if (outfile.is_open()) {
@@ -113,29 +84,12 @@ void TadahCLI::subcommand_train() {
std::cerr << "Error: Unable to open file pot.tadah" << std::endl;
}
if (is_verbose()) std::cout << timer_tot.to_string() << std::endl;
- } else { // Worker code
- // Initialize worker MPI trainer
- MPI_Trainer_WORKER WORKER(config, rank, ncpu);
- while (true) {
- // Send a request for more work to the host
- MPI_Send(&WORKER.rows_available, 1, MPI_INT, 0, MPI_Trainer::WORK_TAG, MPI_COMM_WORLD);
-
- // Probe and respond to host or worker requests
- WORKER.probe();
-
- // Handle requests based on their tags
- if (WORKER.tag == MPI_Trainer::RELEASE_TAG) {
- WORKER.release_tag();
- break;
- } else if (WORKER.tag == MPI_Trainer::WAIT_TAG) {
- WORKER.wait_tag();
- } else if (WORKER.tag == MPI_Trainer::DATA_TAG) {
- WORKER.data_tag();
- } else if (WORKER.tag == MPI_Trainer::WORK_TAG) {
- WORKER.work_tag();
- }
- }
- WORKER.solve(); // Perform worker computations
+ }
+ else { // Worker code
+ // Initialize worker MPI trainer
+ TrainerWorker worker(config, rank, ncpu);
+ worker.run();
+ worker.solve(); // Perform worker computations
}
#else // Non-MPI version
// Start training with a timer in the non-MPI mode
@@ -244,11 +198,9 @@ void TadahCLI::subcommand_predict() {
stpred = modelp->predict(pot_config,stdb,dc,predicted_error);
}
else {
-
stpred = modelp->predict(pot_config,stdb,dc);
}
-
if (is_verbose()) std::cout << "Done!" << std::endl;
if (is_verbose()) std::cout << "Dumping output..." << std::flush;
@@ -289,7 +241,6 @@ void TadahCLI::subcommand_hpo(
[[maybe_unused]]int argc,
[[maybe_unused]]char**argv) {
-
#ifdef TADAH_ENABLE_HPO
if(hpo->count("--Verbose"))
@@ -310,34 +261,25 @@ void TadahCLI::subcommand_hpo(
}
#ifdef TADAH_BUILD_MPI
- int rank=0;
- int ncpu=1;
+
+ int rank;
+ int ncpu;
MPI_Comm_size(MPI_COMM_WORLD, &ncpu);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
- // Create a communicator for LAMMPS
+ MPI_Status status;
MPI_Group world_group;
MPI_Comm_group(MPI_COMM_WORLD, &world_group);
- std::vector<int> ranks;
- ranks.push_back(0); // for now we use root to run LAMMPS
- // long term we want to distribute
- // work among available workers
-
- MPI_Group lmp_group;
- MPI_Group_incl(world_group, ranks.size(), ranks.data(), &lmp_group);
-
- MPI_Comm lmp_comm;
- MPI_Comm_create(MPI_COMM_WORLD, lmp_group, &lmp_comm);
-
if (rank==0) {
if (ncpu<2) {
- std::cout << "Minimum number of cpus for an mpi version is 2" << std::endl;
+ std::cout << "Minimum number of cpus for an MPI version is 2" << std::endl;
return;
}
if (is_verbose()) std::cout << "Optimising HPs..." << std::endl;
CLI::Timer timer_tot {"HPO", CLI::Timer::Big};
- hpo_run(config, target_file, validation_file, lmp_comm);
+ hpo_run(config, target_file, validation_file);
+ if (is_verbose()) std::cout << "Done" << std::endl;
// shut all workers
int cont=0;
@@ -345,54 +287,38 @@ void TadahCLI::subcommand_hpo(
if (is_verbose()) std::cout << timer_tot.to_string() << std::endl;
}
else { // WORKER
+ HPO_Worker hpo_worker(rank, ncpu, world_group);
+
while (true) {
int cs_size;
MPI_Bcast(&cs_size, 1, MPI_INT, 0, MPI_COMM_WORLD);
if (!cs_size)
break;
+ // Obtain updated config file from the HOST
std::vector<char> serialized(cs_size);
MPI_Bcast(serialized.data(), cs_size, MPI_CHAR, 0, MPI_COMM_WORLD);
Config config;
config.deserialize(serialized);
config.postprocess();
- MPI_Trainer_WORKER WORKER(config, rank, ncpu);
- while (true) {
- // ask for more work...
- MPI_Send (&WORKER.rows_available, 1, MPI_INT, 0, MPI_Trainer::WORK_TAG, MPI_COMM_WORLD);
-
- // request from root or from other workers
- WORKER.probe();
-
- // release a worker
- if (WORKER.tag == MPI_Trainer::RELEASE_TAG) {
- WORKER.release_tag();
- break;
- }
- else if (WORKER.tag == MPI_Trainer::WAIT_TAG) {
- // do nothing here; ask for more work in the next cycle
- WORKER.wait_tag();
- }
- else if (WORKER.tag == MPI_Trainer::DATA_TAG) {
- WORKER.data_tag();
- }
- else if (WORKER.tag == MPI_Trainer::WORK_TAG) {
- WORKER.work_tag();
- }
- }
- WORKER.solve();
- }
+ TrainerWorker worker(config, rank, ncpu);
+ worker.run();
+ worker.solve();
+
+ // Run LAMMPS
+ if (hpo_worker.use_lammps)
+ hpo_worker.run();
+ } // end of WORKER external while loop
}
// Free the communicator and groups when done
//MPI_Comm_free(&lmp_comm);
MPI_Group_free(&world_group);
- MPI_Group_free(&lmp_group);
#else
CLI::Timer timer_tot {"HPO", CLI::Timer::Big};
if (is_verbose()) std::cout << "Optimising HPs..." << std::endl;
- int dummy_comm=0;
- hpo_run(config, target_file, validation_file, dummy_comm);
+ //int dummy_comm=0;
+ hpo_run(config, target_file, validation_file);
if (is_verbose()) std::cout << timer_tot.to_string() << std::endl;
#endif
@@ -510,7 +436,9 @@ TadahCLI::TadahCLI():
/* Hyperparameter Optimizer */
/*---------------------------------------------------------------------------*/
ss.str(std::string());
-#ifdef TADAH_ENABLE_HPO
+#ifndef TADAH_ENABLE_HPO
+ ss << "(UNAVAILABLE)" << std::endl;
+#endif
ss << "Optimize the model's architecture and determine\n";
ss << "the best hyperparameters within predefined constraints.\n";
ss << "This option uses hyperparameter optimization\n";
@@ -519,6 +447,7 @@ TadahCLI::TadahCLI():
hpo = app.add_subcommand("hpo", ss.str());
+#ifdef TADAH_ENABLE_HPO
ss.str(std::string());
ss << "A config file containing model inital parameters\n";
ss << "and training dataset(s).\n";
@@ -552,8 +481,6 @@ TadahCLI::TadahCLI():
//hpo->add_flag("-u,--uncertainty",
// "Dump uncertainty on weights."); // TODO check this
-#else
- ss << "(UNAVAILABLE)" << std::endl;
#endif
}
diff --git a/examples/ex_0/targets_new b/examples/ex_0/targets_new
index c4e2ed8..cce6296 100644
--- a/examples/ex_0/targets_new
+++ b/examples/ex_0/targets_new
@@ -1,19 +1,19 @@
# Global settings for the hyperparameter optimiser
-MAXCALLS 100 # Maximum number of optimisation cycles
-EPS 1e-3 # Accuracy convergence threshold.
- # Smaller values increase precision but slow global exploration.
+MAXCALLS 10 # Maximum number of optimisation cycles
+EPS 1e-3 # Accuracy convergence threshold. (default 1e-2)
+ # Smaller values increase precision but slow down global exploration.
# Model parameters to be optimised.
# Model is specified in a separate config file.
-OPTIM RCUT2B 5.0 6.0 # Optimise RCUT2B within these bounds
-OPTIM CGRID2B 1.1 6.0 # Optimise CGRID2B within these bounds
-OPTIM SGRID2B 0.1 4.11 # Optimise SGRID2B within these bounds
-OPTIM AGRIDMB 0 2 # Optimise AGRIDMB within these bounds
-OPTIM EWEIGHT 1e-2 1e0 # Optimise EWEIGHT within these bounds
+OPTIM RCUT2B 3.0 6.0 # Optimise RCUT2B within these bounds
+OPTIM CGRID2B 1.0 6.0 # Optimise CGRID2B within these bounds
+OPTIM SGRID2B 0.01 6.0 # Optimise SGRID2B within these bounds
+#OPTIM AGRIDMB 0 2 # Optimise AGRIDMB within these bounds
+#OPTIM EWEIGHT 1e-2 1e0 # Optimise EWEIGHT within these bounds
# Basic optimisation targets with weighted inclusion in global loss
ERMSE 0 100 # Energy RMSE with weight 100, printed to properties.dat
-SRMSE 0 20 # Stress RMSE with weight 20, printed to properties.dat
+#SRMSE 0 20 # Stress RMSE with weight 20, printed to properties.dat
# Default settings
LOSSFUNC 2 # Loss function type (default 1):
@@ -22,8 +22,8 @@ LOSSFUNC 2 # Loss function type (default 1):
# [3] -> log(tloss)
# Default settings for LAMMPS simulations
-FAILSCORE 100 # Score assigned if simulation fails (default 100)
-MAXTIME 2 1 # Maximum of 2 seconds for LAMMPS simulations, checked per step (default 10 1)
+FAILSCORE 100 # Score assigned if simulation fails (default 1000)
+MAXTIME 2 # Maximum of 2 seconds for LAMMPS simulations, checked every step (default 10)
# Advanced optimisation targets used with LAMMPS
# Default settings may be overridden with specific flags
@@ -37,16 +37,18 @@ MAXTIME 2 1 # Maximum of 2 seconds for LAMMPS simulations, checked per step (
# -o, --outvar # Output variable produced by the script, printed to properties.dat
# -w, --weight # Weighting factor for tloss: w*tloss
# -f, --failscore # Custom FAILSCORE for this simulation
+# -l, --lossfunc # Custom LOSSFUNC for this simulation
# -m, --maxtime # Max time for simulation; returns failscore if exceeded
+# -n, --ncpu # Number of cores to be used by this LAMMPS simulation, MPI only
# Simulation 1: Optimise lattice constant using FCC lattice.
# Executes LAMMPS with 'in.lata', fully defining the simulation and 'tloss'.
-LAMMPS --script in.lata --varloss tloss
+#LAMMPS --script in.lata --varloss tloss
-# Simulation 2: Similar setup as simulation 1, at varying pressures.
+# Simulation 2: Similar setup as simulation 1, but at different pressures.
# 'tloss' is defined in 'in.mod_lata' as:
# tloss = abs(lata_exp - lata) / lata_exp
-LAMMPS --script in.mod_lata --varloss tloss --invar P 1 --invar lata_exp 3.3 --outvar lata
+LAMMPS --script in.mod_lata --varloss tloss --invar P 1 --invar lata_exp 3.3 --outvar lata --ncpu 4
LAMMPS --script in.mod_lata --varloss tloss --invar P 100 --invar lata_exp 3.2 --outvar lata
-LAMMPS --script in.mod_lata --varloss tloss --invar P 1000 --invar lata_exp 3.1 --outvar lata
-
+LAMMPS --script in.mod_lata --varloss tloss --invar P 1000 --invar lata_exp 3.1 --outvar lata --ncpu 1
+LAMMPS --script in.mod2_lata --varloss tloss --invar P 1000 --invar lata_exp 3.1 --outvar lata --ncpu 4
--
GitLab