def run_simulations(args, save_result, local_mode, init_ray=True):
if init_ray:
start_ray(local_mode)
if save_result:
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
n_reps = args.n_reps
alg_param_grid = get_grid(args.param_grid_def)
n_grid = alg_param_grid.shape[0]
config_grid_vals = get_grid(args.config_grid_def)
n_config_grid = len(config_grid_vals)
planing_loss = np.zeros((n_reps, n_config_grid, n_grid))
info_dict = {}
# ----- Run simulation in parrnell process---------------------------------------------#
loss_rep_id_lst = []
for i_rep in range(n_reps):
# returns objects ids:
args_r = deepcopy(args)
planing_loss_rep_id = run_rep.remote(i_rep, alg_param_grid,
args_r.config_grid_def, args_r)
loss_rep_id_lst.append(planing_loss_rep_id)
# end for i_rep
# ----- get the results --------------------------------------------#
for i_rep in range(n_reps):
loss_rep = ray.get(loss_rep_id_lst[i_rep])
if i_rep % max(n_reps // 100, 1) == 0:
time_str = time.strftime(
"%H hours, %M minutes and %S seconds",
time.gmtime(timeit.default_timer() - start_time))
write_to_log(
'Finished: {} out of {} reps, time: {}'.format(
i_rep + 1, n_reps, time_str), args)
# end if
planing_loss[i_rep] = loss_rep
info_dict = {
'planing_loss_avg': planing_loss.mean(axis=0),
'planing_loss_std': planing_loss.std(axis=0),
'alg_param_grid': alg_param_grid,
'n_reps_finished': i_rep + 1
}
if save_result:
save_run_data(args, info_dict, verbose=0)
# end if
# end for i_rep
if save_result:
save_run_data(args, info_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args, save_result)
return info_dict
def run_simulations(args, save_result, local_mode):
args_def = deepcopy(args)
start_ray(local_mode)
if save_result:
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
n_reps = args.n_reps
param_val_grid = get_grid(args.param_grid_def)
n_grid = param_val_grid.shape[0]
config_grid = get_grid(args.config_grid_def)
n_configs = len(config_grid)
args.n_configs = n_configs
loss_mat = np.zeros((n_reps, n_configs, n_grid))
# ----- Run simulation in parrnell process---------------------------------------------#
loss_rep_id_lst = []
for i_rep in range(n_reps):
# returns objects ids:
loss_mat_rep_id = run_rep.remote(i_rep, param_val_grid, config_grid,
args)
loss_rep_id_lst.append(loss_mat_rep_id)
# ----- get the results --------------------------------------------#
for i_rep in range(n_reps):
loss_rep = ray.get(loss_rep_id_lst[i_rep])
write_to_log('Finished: {} out of {} reps'.format(i_rep + 1, n_reps),
args)
loss_mat[i_rep] = loss_rep
# end for i_rep
info_dict = {
'loss_avg': loss_mat.mean(axis=0),
'loss_std': loss_mat.std(axis=0),
'param_val_grid': param_val_grid,
'config_grid': config_grid
}
if save_result:
save_run_data(args, info_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args)
write_to_log(
['-' * 10 + 'Defined args: ',
pretty_print_args(args_def), '-' * 20], args)
return info_dict
def run_simulations(args, local_mode):
start_ray(local_mode)
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
create_result_dir(args)
set_random_seed(args.seed)
l2_grid = get_grid(args.l2_grid_def)
gam_grid = get_grid(args.gam_grid_def)
write_to_log('gamma_grid == {}'.format(gam_grid), args)
write_to_log('l2_grid == {}'.format(l2_grid), args)
grid_shape = (len(l2_grid), len(gam_grid))
loss_avg = np.zeros(grid_shape)
loss_std = np.zeros(grid_shape)
run_idx = 0
for i0 in range(grid_shape[0]):
for i1 in range(grid_shape[1]):
args_run = deepcopy(args)
args_run.param_grid_def = {
'type': 'L2_factor',
'spacing': 'list',
'list': [l2_grid[i0]]
}
args_run.default_gamma = gam_grid[i1]
info_dict = run_main_control(args_run,
save_result=False,
plot=False,
init_ray=False)
loss_avg[i0, i1] = info_dict['planing_loss_avg'][0]
loss_std[i0, i1] = info_dict['planing_loss_std'][0]
run_idx += 1
print("Finished {}/{}".format(run_idx, loss_avg.size))
# end for
# end for
grid_results_dict = {
'l2_grid': l2_grid,
'gam_grid': gam_grid,
'loss_avg': loss_avg,
'loss_std': loss_std
}
save_run_data(args, grid_results_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args)
return grid_results_dict
return results_dict
# end run_simulations
# -------------------------------------------------------------------------------------------
if __name__ == "__main__":
# *********************************
if run_mode == 'Load':
args, results_dict = load_run_data(result_dir_to_load)
# *********************************
elif run_mode == 'New':
hyper_grid_vals = get_grid(args.hyper_grid_def)
create_result_dir(args)
n_hyper_grid = len(hyper_grid_vals)
n_reps = args.n_reps
reg_types = args.reg_types
# define search grids for regularization parameters
reg_grids = dict()
for reg_type in reg_types:
reg_param_grid_def = args.search_grid_def[reg_type]
reg_grids[reg_type] = get_grid(reg_param_grid_def)
# init result matrix with nan (no result)
loss = {reg_type: np.full((n_hyper_grid, len(reg_grids[reg_type]), n_reps), np.nan) for reg_type in reg_types}
# run
results_dict = run_simulation(args, hyper_grid_vals, loss, reg_grids, local_mode)