def main(case_num=1, nx=11, interpolate_nan_outputs=True):
# when from the command line
case_num = int(case_num)
nx = int(nx)
interpolate_nan_outputs = bool(interpolate_nan_outputs)
# Start MPI communicator
comm, rank, num_procs = _init_mpi()
tic = time.time()
global_results, sweep_params = run_analysis(case_num, nx,
interpolate_nan_outputs)
print(global_results)
toc = time.time()
if rank == 0:
total_samples = 1
for k, v in sweep_params.items():
total_samples *= v.num_samples
print("Finished case_num = %d." % (case_num))
print("Processed %d swept parameters comprising %d total points." %
(len(sweep_params), total_samples))
print("Elapsed time = %.1f s." % (toc - tic))
return global_results, sweep_params
def test_parameter_sweep_bad_force_initialize(self, model, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
A = m.fs.input['a']
B = m.fs.input['b']
sweep_params = {A.name: (A, 0.1, 0.9, 3), B.name: (B, 0.0, 0.5, 3)}
results_file = os.path.join(tmp_path, 'global_results_recover.csv')
h5_fname = "output_dict_recover"
with pytest.raises(ValueError):
# Call the parameter_sweep function
parameter_sweep(m,
sweep_params,
outputs=None,
csv_results_file=results_file,
h5_results_file=h5_fname,
optimize_function=_optimization,
reinitialize_before_sweep=True,
reinitialize_function=None,
reinitialize_kwargs=None,
mpi_comm=comm)
def test_create_global_output(self, model):
comm, rank, num_procs = _init_mpi()
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
sweep_params = {
'input_a': NormalSample(m.fs.input['a'], 0.1, 0.9),
'input_b': NormalSample(m.fs.input['b'], 0.0, 0.5)
}
outputs = {
'output_c': m.fs.output['c'],
'output_d': m.fs.output['d'],
'performance': m.fs.performance
}
sweep_params, sampling_type = _process_sweep_params(sweep_params)
# Get the globale sweep param values
global_num_cases = 2 * num_procs
global_values = _build_combinations(sweep_params, sampling_type,
global_num_cases, comm, rank,
num_procs)
# divide the workload between processors
local_values = _divide_combinations(global_values, rank, num_procs)
local_num_cases = np.shape(local_values)[0]
local_output_dict, outputs = _create_local_output_skeleton(
model, sweep_params, None, local_num_cases)
# Manually update the values in the numpy array
for key, value in local_output_dict.items():
for subkey, subvalue in value.items():
subvalue['value'][:] = rank
# Local output dict also contains the solve_successful. The solve status is
# based on the
local_output_dict['solve_successful'] = [True] * local_num_cases
# Get the global output dictionary, This is properly created only on rank 0
global_output_dict = _create_global_output(local_output_dict,
global_num_cases, comm,
rank, num_procs)
if num_procs == 1:
assert local_output_dict == global_output_dict
else:
comm.Barrier()
if rank > 0:
assert global_output_dict == local_output_dict
else:
test_array = np.repeat(np.arange(0, num_procs, dtype=float), 2)
test_list = [True] * global_num_cases
for key, value in global_output_dict.items():
if key != 'solve_successful':
for subkey, subvalue in value.items():
assert np.allclose(subvalue['value'], test_array)
elif key == 'solve_successful':
assert list(value) == test_list
comm.Barrier()
def test_divide_combinations(self):
# _divide_combinations(global_combo_array, rank, num_procs)
comm, rank, num_procs = _init_mpi()
A_param = pyo.Param(initialize=0.0, mutable=True)
B_param = pyo.Param(initialize=1.0, mutable=True)
C_param = pyo.Param(initialize=2.0, mutable=True)
range_A = [0.0, 10.0]
range_B = [1.0, 20.0]
range_C = [2.0, 30.0]
nn_A = 4
nn_B = 5
nn_C = 6
param_dict = dict()
param_dict['var_A'] = LinearSample(A_param, range_A[0], range_A[1],
nn_A)
param_dict['var_B'] = LinearSample(B_param, range_B[0], range_B[1],
nn_B)
param_dict['var_C'] = LinearSample(C_param, range_C[0], range_C[1],
nn_C)
global_combo_array = _build_combinations(param_dict,
SamplingType.FIXED, None,
comm, rank, num_procs)
test = np.array_split(global_combo_array, num_procs, axis=0)[rank]
local_combo_array = _divide_combinations(global_combo_array, rank,
num_procs)
assert np.shape(local_combo_array)[1] == 3
assert np.allclose(test[:, 0], local_combo_array[:, 0])
assert np.allclose(test[:, 1], local_combo_array[:, 1])
assert np.allclose(test[:, 2], local_combo_array[:, 2])
if rank == 0:
assert local_combo_array[0, 0] == pytest.approx(range_A[0])
assert local_combo_array[0, 1] == pytest.approx(range_B[0])
assert local_combo_array[0, 2] == pytest.approx(range_C[0])
if rank == num_procs - 1:
assert local_combo_array[-1, 0] == pytest.approx(range_A[1])
assert local_combo_array[-1, 1] == pytest.approx(range_B[1])
assert local_combo_array[-1, 2] == pytest.approx(range_C[1])
def test_aggregate_results(self):
comm, rank, num_procs = _init_mpi()
# print('Rank %d, num_procs %d' % (rank, num_procs))
nn = 5
np.random.seed(1)
local_results = (rank + 1) * np.random.rand(nn, 2)
global_values = np.random.rand(nn * num_procs, 4)
global_results = _aggregate_results(local_results, global_values, comm,
num_procs)
assert np.shape(global_results)[1] == np.shape(local_results)[1]
assert np.shape(global_results)[0] == np.shape(global_values)[0]
if rank == 0:
assert global_results[0, 0] == pytest.approx(local_results[0, 0])
assert global_results[0, 1] == pytest.approx(local_results[0, 1])
assert global_results[-1, 0] == pytest.approx(num_procs *
local_results[-1, 0])
assert global_results[-1, 1] == pytest.approx(num_procs *
local_results[-1, 1])
def test_linear_build_combinations(self):
comm, rank, num_procs = _init_mpi()
A_param = pyo.Param(initialize=0.0, mutable=True)
B_param = pyo.Param(initialize=1.0, mutable=True)
C_param = pyo.Param(initialize=2.0, mutable=True)
range_A = [0.0, 10.0]
range_B = [1.0, 20.0]
range_C = [2.0, 30.0]
nn_A = 4
nn_B = 5
nn_C = 6
param_dict = dict()
param_dict['var_A'] = LinearSample(A_param, range_A[0], range_A[1],
nn_A)
param_dict['var_B'] = LinearSample(B_param, range_B[0], range_B[1],
nn_B)
param_dict['var_C'] = LinearSample(C_param, range_C[0], range_C[1],
nn_C)
global_combo_array = _build_combinations(param_dict,
SamplingType.FIXED, None,
comm, rank, num_procs)
assert np.shape(global_combo_array)[0] == nn_A * nn_B * nn_C
assert np.shape(global_combo_array)[1] == len(param_dict)
assert global_combo_array[0, 0] == pytest.approx(range_A[0])
assert global_combo_array[0, 1] == pytest.approx(range_B[0])
assert global_combo_array[0, 2] == pytest.approx(range_C[0])
assert global_combo_array[-1, 0] == pytest.approx(range_A[1])
assert global_combo_array[-1, 1] == pytest.approx(range_B[1])
assert global_combo_array[-1, 2] == pytest.approx(range_C[1])
def test_parameter_sweep_bad_recover(self, model, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
A = m.fs.input['a']
B = m.fs.input['b']
sweep_params = {A.name: (A, 0.1, 0.9, 3), B.name: (B, 0.0, 0.5, 3)}
results_file = os.path.join(tmp_path, 'global_results_bad_recover.csv')
h5_fname = "output_dict_bad_recover"
# Call the parameter_sweep function
parameter_sweep(m,
sweep_params,
outputs=None,
csv_results_file=results_file,
h5_results_file=h5_fname,
optimize_function=_optimization,
reinitialize_function=_bad_reinitialize,
reinitialize_kwargs={'slack_penalty': 10.},
mpi_comm=comm)
# NOTE: rank 0 "owns" tmp_path, so it needs to be
# responsible for doing any output file checking
# tmp_path can be deleted as soon as this method
# returns
if rank == 0:
# Check that the global results file is created
assert os.path.isfile(results_file)
# Attempt to read in the data
data = np.genfromtxt(results_file, skip_header=1, delimiter=',')
# Compare the last row of the imported data to truth
truth_data = [
0.9, 0.5, np.nan, np.nan, np.nan, np.nan, np.nan, np.nan
]
assert np.allclose(data[-1], truth_data, equal_nan=True)
if rank == 0:
truth_dict = {
'outputs': {
'fs.output[c]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0.2, 0.2, np.nan, 1., 1., np.nan, np.nan, np.nan,
np.nan
])
},
'fs.output[d]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0., 0.75, np.nan, 0., 0.75, np.nan, np.nan, np.nan,
np.nan
])
},
'fs.performance': {
'value':
np.array([
0.2, 0.95, np.nan, 1., 1.75, np.nan, np.nan,
np.nan, np.nan
])
},
'fs.slack[ab_slack]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0., 0., np.nan, 0., 0., np.nan, np.nan, np.nan,
np.nan
])
},
'fs.slack[cd_slack]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0., 0., np.nan, 0., 0., np.nan, np.nan, np.nan,
np.nan
])
},
'objective': {
'value':
np.array([
0.2, 0.95, np.nan, 1., 1.75, np.nan, np.nan,
np.nan, np.nan
])
}
},
'solve_successful':
[True, True, False, True, True, False, False, False, False],
'sweep_params': {
'fs.input[a]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0.1, 0.1, 0.1, 0.5, 0.5, 0.5, 0.9, 0.9, 0.9])
},
'fs.input[b]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0., 0.25, 0.5, 0., 0.25, 0.5, 0., 0.25, 0.5])
}
}
}
h5_fpath = os.path.join(tmp_path, '{0}.h5'.format(h5_fname))
read_dict = _read_output_h5(h5_fpath)
_assert_dictionary_correctness(truth_dict, read_dict)
_assert_h5_csv_agreement(results_file, read_dict)
sweep_params,
outputs,
csv_results_file_name=output_filename,
optimize_function=opt_function,
optimize_kwargs=optimize_kwargs,
debugging_data_dir=os.path.split(output_filename)[0] + "/local",
interpolate_nan_outputs=interp_nan_outputs,
)
return global_results, sweep_params
if __name__ == "__main__":
# Start MPI communicator
comm, rank, num_procs = _init_mpi()
# Get the case number to run
try:
case_num = int(sys.argv[1])
except:
# Default to running case 1
case_num = 1
# Get the default number of discretization points
try:
nx = int(sys.argv[2])
except:
# Default to a 4-point discretization
nx = 4
def test_init_mpi(self):
comm, rank, num_procs = _init_mpi()
assert type(rank) == int
assert type(num_procs) == int
assert 0 <= rank < num_procs
def test_parameter_sweep_optimize(self, model, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
A = m.fs.input['a']
B = m.fs.input['b']
sweep_params = {A.name: (A, 0.1, 0.9, 3), B.name: (B, 0.0, 0.5, 3)}
outputs = {
'output_c': m.fs.output['c'],
'output_d': m.fs.output['d'],
'performance': m.fs.performance,
'objective': m.objective
}
results_file = os.path.join(tmp_path, 'global_results_optimize.csv')
h5_fname = "output_dict_optimize"
# Call the parameter_sweep function
parameter_sweep(m,
sweep_params,
outputs=outputs,
csv_results_file=results_file,
h5_results_file=h5_fname,
optimize_function=_optimization,
optimize_kwargs={'relax_feasibility': True},
mpi_comm=comm)
# NOTE: rank 0 "owns" tmp_path, so it needs to be
# responsible for doing any output file checking
# tmp_path can be deleted as soon as this method
# returns
if rank == 0:
# Check that the global results file is created
assert os.path.isfile(results_file)
# Attempt to read in the data
data = np.genfromtxt(results_file, skip_header=1, delimiter=',')
# Compare the last row of the imported data to truth
truth_data = [
0.9, 0.5, 1.0, 1.0, 2.0,
2.0 - 1000. * ((2. * 0.9 - 1.) + (3. * 0.5 - 1.))
]
assert np.allclose(data[-1], truth_data, equal_nan=True)
# Check the h5
if rank == 0:
truth_dict = {
'outputs': {
'output_c': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value':
np.array([0.2, 0.2, 0.2, 1., 1., 1., 1., 1., 1.])
},
'output_d': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
9.98580690e-09, 0.75, 1., 9.99872731e-09, 0.75, 1.,
9.99860382e-09, 0.75, 1.
])
},
'performance': {
'value':
np.array([0.2, 0.95, 1.2, 1., 1.75, 2., 1., 1.75, 2.])
},
'objective': {
'value':
np.array([
0.2, 9.50000020e-01, -4.98799990e+02, 1., 1.75,
-4.97999990e+02, -7.98999990e+02, -7.98249990e+02,
2.0 - 1000. * ((2. * 0.9 - 1.) + (3. * 0.5 - 1.))
])
}
},
'solve_successful': [True] * 9,
'sweep_params': {
'fs.input[a]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0.1, 0.1, 0.1, 0.5, 0.5, 0.5, 0.9, 0.9, 0.9])
},
'fs.input[b]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0., 0.25, 0.5, 0., 0.25, 0.5, 0., 0.25, 0.5])
}
}
}
h5_fpath = os.path.join(tmp_path, '{0}.h5'.format(h5_fname))
read_dict = _read_output_h5(h5_fpath)
_assert_dictionary_correctness(truth_dict, read_dict)
_assert_h5_csv_agreement(results_file, read_dict)
def test_recursive_parameter_sweep(model, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
m = model
solver = pyo.SolverFactory("ipopt")
sweep_params = {}
sweep_params["a_val"] = UniformSample(m.fs.a, 0.0, 1.0)
outputs = {}
outputs["x_val"] = m.fs.x
# Run the parameter sweep study using num_samples randomly drawn from the above range
num_samples = 10
seed = 0
results_fname = os.path.join(tmp_path, "global_results")
csv_results_file = str(results_fname) + ".csv"
h5_results_file = str(results_fname) + ".h5"
# Run the parameter sweep
# recursive_parameter_sweep(m, sweep_params, outputs, results_file='recursive_sweep.csv',
# optimize_function=optimize, optimize_kwargs={'solver':solver}, req_num_samples=num_samples,
# seed=seed, reinitialize_before_sweep=False, reinitialize_function=initialize_system,
# reinitialize_kwargs={'solver':solver}
data = recursive_parameter_sweep(
m,
sweep_params,
outputs=outputs,
csv_results_file_name=csv_results_file,
h5_results_file_name=h5_results_file,
req_num_samples=num_samples,
debugging_data_dir=tmp_path,
seed=seed,
)
reference_save_data = np.array([
[0.38344152, 0.11655848],
[0.4236548, 0.0763452],
[0.43758721, 0.06241279],
[0.0187898, 0.4812102],
[0.0202184, 0.4797816],
[0.06022547, 0.43977453],
[0.07103606, 0.42896394],
[0.0871293, 0.4128707],
[0.10204481, 0.39795519],
[0.11827443, 0.38172557],
])
assert np.shape(data) == (10, 2)
assert np.allclose(reference_save_data, data, equal_nan=True)
assert np.allclose(np.sum(data, axis=1), value(m.fs.success_prob))
if rank == 0:
# Check that the global results file is created
assert os.path.isfile(csv_results_file)
# Check that all local output files have been created
for k in range(num_procs):
assert os.path.isfile(
os.path.join(tmp_path, f"local_results_{k:03}.h5"))
assert os.path.isfile(
os.path.join(tmp_path, f"local_results_{k:03}.csv"))
csv_data = np.genfromtxt(csv_results_file,
skip_header=1,
delimiter=",")
# Compare the last row of the imported data to truth
assert np.allclose(data[-1, :],
reference_save_data[-1, :],
equal_nan=True)
# Check for the h5 output
truth_dict = {
"outputs": {
"x_val": {
"lower bound":
-1.7976931348623157e308,
"units":
"None",
"upper bound":
1.7976931348623157e308,
"value":
np.array([
0.11655848,
0.0763452,
0.06241279,
0.4812102,
0.4797816,
0.43977453,
0.42896394,
0.4128707,
0.39795519,
0.38172557,
0.3710737,
0.35664671,
0.33869048,
0.29112324,
0.28961744,
0.23544439,
0.18457165,
0.1404921,
0.13628923,
0.08533806,
0.06296805,
0.06139849,
0.04384967,
0.03852064,
]),
}
},
"solve_successful": [
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
True,
],
"sweep_params": {
"fs.a": {
"units":
"None",
"value":
np.array([
0.38344152,
0.4236548,
0.43758721,
0.0187898,
0.0202184,
0.06022547,
0.07103606,
0.0871293,
0.10204481,
0.11827443,
0.1289263,
0.14335329,
0.16130952,
0.20887676,
0.21038256,
0.26455561,
0.31542835,
0.3595079,
0.36371077,
0.41466194,
0.43703195,
0.43860151,
0.45615033,
0.46147936,
]),
}
},
}
read_dict = _read_output_h5(h5_results_file)
_assert_dictionary_correctness(truth_dict, read_dict)
assert np.allclose(
data[:, -1], read_dict["outputs"]["x_val"]["value"][:num_samples])
# Check for the companion text file
txt_fpath = os.path.join(tmp_path, "{0}.txt".format(h5_results_file))
assert os.path.exists(txt_fpath)
truth_txt_dict = {
"outputs": ["x_val"],
"sweep_params": ["fs.a"],
}
with open(txt_fpath, "r") as f:
f_contents = f.read()
read_txt_dict = ast.literal_eval(f_contents)
assert read_txt_dict == truth_txt_dict
def test_parameter_sweep(self, model, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
A = m.fs.input['a']
B = m.fs.input['b']
sweep_params = {A.name: (A, 0.1, 0.9, 3), B.name: (B, 0.0, 0.5, 3)}
outputs = {
'output_c': m.fs.output['c'],
'output_d': m.fs.output['d'],
'performance': m.fs.performance
}
results_file = os.path.join(tmp_path, 'global_results.csv')
h5_fname = "output_dict"
# Call the parameter_sweep function
parameter_sweep(m,
sweep_params,
outputs=outputs,
csv_results_file=results_file,
h5_results_file=h5_fname,
optimize_function=_optimization,
debugging_data_dir=tmp_path,
mpi_comm=comm)
# NOTE: rank 0 "owns" tmp_path, so it needs to be
# responsible for doing any output file checking
# tmp_path can be deleted as soon as this method
# returns
if rank == 0:
# Check that the global results file is created
assert os.path.isfile(results_file)
# Check that all local output files have been created
for k in range(num_procs):
assert os.path.isfile(
os.path.join(tmp_path, f'local_results_{k:03}.csv'))
# Attempt to read in the data
data = np.genfromtxt(results_file, skip_header=1, delimiter=',')
# Compare the last row of the imported data to truth
truth_data = [0.9, 0.5, np.nan, np.nan, np.nan]
assert np.allclose(data[-1], truth_data, equal_nan=True)
# Check for the h5 output
if rank == 0:
truth_dict = {
'outputs': {
'output_c': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0.2, 0.2, np.nan, 1., 1., np.nan, np.nan, np.nan,
np.nan
])
},
'output_d': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([
0., 0.75, np.nan, 0., 0.75, np.nan, np.nan, np.nan,
np.nan
])
},
'performance': {
'value':
np.array([
0.2, 0.95, np.nan, 1., 1.75, np.nan, np.nan,
np.nan, np.nan
])
}
},
'solve_successful':
[True, True, False, True, True, False, False, False, False],
'sweep_params': {
'fs.input[a]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0.1, 0.1, 0.1, 0.5, 0.5, 0.5, 0.9, 0.9, 0.9])
},
'fs.input[b]': {
'lower bound':
0,
'units':
'None',
'upper bound':
0,
'value':
np.array([0., 0.25, 0.5, 0., 0.25, 0.5, 0., 0.25, 0.5])
}
}
}
h5_fpath = os.path.join(tmp_path, 'output_dict.h5')
read_dict = _read_output_h5(h5_fpath)
_assert_dictionary_correctness(truth_dict, read_dict)
_assert_h5_csv_agreement(results_file, read_dict)
# Check if there is a text file created
import ast
truth_txt_dict = {
'outputs': ['output_c', 'output_d', 'performance'],
'sweep_params': ['fs.input[a]', 'fs.input[b]']
}
txt_fpath = os.path.join(tmp_path, '{0}.txt'.format(h5_fname))
assert os.path.exists(txt_fpath)
f = open(txt_fpath)
f_contents = f.read()
read_txt_dict = ast.literal_eval(f_contents)
assert read_txt_dict == truth_txt_dict
def test_create_local_output_skeleton(self, model):
comm, rank, num_procs = _init_mpi()
m = model
m.fs.slack_penalty = 1000.
m.fs.slack.setub(0)
sweep_params = {
'input_a': (m.fs.input['a'], 0.1, 0.9, 3),
'input_b': (m.fs.input['b'], 0.0, 0.5, 3)
}
outputs = {
'output_c': m.fs.output['c'],
'output_d': m.fs.output['d'],
'performance': m.fs.performance
}
sweep_params, sampling_type = _process_sweep_params(sweep_params)
values = _build_combinations(sweep_params, sampling_type, None, comm,
rank, num_procs)
num_cases = np.shape(values)[0]
output_dict, outputs = _create_local_output_skeleton(
model, sweep_params, None, num_cases)
truth_dict = {
'outputs': {
'fs.output[c]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.output[d]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.performance': {
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.slack[ab_slack]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.slack[cd_slack]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'objective': {
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
}
},
'sweep_params': {
'fs.input[a]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.input[b]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
}
}
}
_assert_dictionary_correctness(truth_dict, output_dict)
def test_h5_read_write(self, tmp_path):
comm, rank, num_procs = _init_mpi()
tmp_path = _get_rank0_path(comm, tmp_path)
reference_dict = {
'outputs': {
'fs.input[a]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.input[b]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.output[c]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0.2, 0.2, 0., 1., 1., 0., 0., 0., 0.])
},
'fs.output[d]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0.75, 0., 0., 0.75, 0., 0., 0., 0.])
},
'fs.slack[ab_slack]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
},
'fs.slack[cd_slack]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0., 0., 0., 0., 0., 0., 0., 0.])
}
},
'solve_successful': [True] * 9,
'sweep_params': {
'fs.input[a]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0.1, 0.1, 0., 0.5, 0.5, 0., 0., 0., 0.])
},
'fs.input[b]': {
'lower bound': 0,
'units': 'None',
'upper bound': 0,
'value': np.array([0., 0.25, 0., 0., 0.25, 0., 0., 0., 0.])
}
}
}
h5_fname = "h5_test_{0}.h5".format(rank)
_write_output_to_h5(reference_dict,
output_directory=tmp_path,
fname=h5_fname)
read_dictionary = _read_output_h5(os.path.join(tmp_path, h5_fname))
_assert_dictionary_correctness(reference_dict, read_dictionary)
def test_aggregate_filtered_input_arr():
comm, rank, num_procs = _init_mpi()
input_dict = {
"outputs": {
"x_val": {
"lower bound":
None,
"units":
"None",
"upper bound":
None,
"value":
np.array([
0.31655848,
0.2763452,
0.26241279,
0.15511682,
0.1511865,
0.09723662,
0.05410589,
0.6797816,
0.62896394,
0.6128707,
0.23852064,
0.17110508,
0.13195544,
]),
}
},
"solve_successful": [True] * 13,
"sweep_params": {
"fs.a": {
"units":
"None",
"value":
np.array([
0.38344152,
0.4236548,
0.43758721,
0.54488318,
0.5488135,
0.60276338,
0.64589411,
0.0202184,
0.07103606,
0.0871293,
0.46147936,
0.52889492,
0.56804456,
]),
}
},
}
truth_arr = np.array([
0.38344152,
0.4236548,
0.43758721,
0.54488318,
0.5488135,
0.60276338,
0.64589411,
0.0202184,
0.07103606,
0.0871293,
0.46147936,
0.52889492,
0.56804456,
])
req_num_samples = 10
values_arr = _aggregate_filtered_input_arr(input_dict, req_num_samples,
comm, rank, num_procs)
assert np.allclose(values_arr[:, 0],
truth_arr[0:req_num_samples],
equal_nan=True)
def recursive_parameter_sweep(
model,
sweep_params,
outputs=None,
csv_results_file_name=None,
h5_results_file_name=None,
optimize_function=_default_optimize,
optimize_kwargs=None,
reinitialize_function=None,
reinitialize_kwargs=None,
reinitialize_before_sweep=False,
mpi_comm=None,
debugging_data_dir=None,
interpolate_nan_outputs=False,
req_num_samples=None,
seed=None,
):
# Get an MPI communicator
comm, rank, num_procs = _init_mpi(mpi_comm)
# Convert sweep_params to LinearSamples
sweep_params, sampling_type = _process_sweep_params(sweep_params)
# Set the seed before sampling
np.random.seed(seed)
# Set up optimize_kwargs
if optimize_kwargs is None:
optimize_kwargs = dict()
# Set up reinitialize_kwargs
if reinitialize_kwargs is None:
reinitialize_kwargs = dict()
n_samples_remaining = req_num_samples
num_total_samples = req_num_samples
local_output_collection = {}
loop_ctr = 0
while n_samples_remaining > 0 and loop_ctr < 10:
# Enumerate/Sample the parameter space
global_values = _build_combinations(sweep_params, sampling_type,
num_total_samples, comm, rank,
num_procs)
# divide the workload between processors
local_values = _divide_combinations(global_values, rank, num_procs)
local_num_cases = np.shape(local_values)[0]
if loop_ctr == 0:
true_local_num_cases = local_num_cases
local_output_collection[loop_ctr] = _do_param_sweep(
model,
sweep_params,
outputs,
local_values,
optimize_function,
optimize_kwargs,
reinitialize_function,
reinitialize_kwargs,
reinitialize_before_sweep,
comm,
)
# Get the number of successful solves on this proc (sum of boolean flags)
success_count = sum(
local_output_collection[loop_ctr]["solve_successful"])
failure_count = local_num_cases - success_count
# Get the global number of successful solves and update the number of remaining samples
if num_procs > 1: # pragma: no cover
global_success_count = np.zeros(1, dtype=np.float64)
global_failure_count = np.zeros(1, dtype=np.float64)
comm.Allreduce(np.array(success_count, dtype=np.float64),
global_success_count)
comm.Allreduce(np.array(failure_count, dtype=np.float64),
global_failure_count)
else:
global_success_count = success_count
global_failure_count = failure_count
success_prob = global_success_count / (global_failure_count +
global_success_count)
if success_prob < 0.1:
warnings.warn(
f"Success rate of solves = {100.0*success_prob}%, consider adjusting sweep limits."
)
n_samples_remaining -= global_success_count
# The total number of samples to generate at the next iteration is a multiple of the total remaining samples
scale_factor = 2.0 / max(success_prob, 0.10)
num_total_samples = int(np.ceil(scale_factor * n_samples_remaining))
loop_ctr += 1
# Now that we have all of the local output dictionaries, we need to construct
# a consolidated dictionary based on a filter, e.g., optimal solves.
local_filtered_dict, local_n_successful = _filter_recursive_solves(
model, sweep_params, outputs, local_output_collection, comm)
# if we are debugging
if debugging_data_dir is not None:
local_filtered_values = np.zeros(
(local_n_successful, len(local_filtered_dict["sweep_params"])),
dtype=np.float64,
)
for i, (key, item) in enumerate(
local_filtered_dict["sweep_params"].items()):
local_filtered_values[:, i] = item["value"][:]
else:
local_filtered_values = None
# Not that we have all of the successful outputs in a consolidated dictionary locally,
# we can now construct a global dictionary of successful solves.
(
global_filtered_dict,
global_filtered_results,
global_filtered_values,
) = _aggregate_filtered_results(local_filtered_dict, req_num_samples, comm,
rank, num_procs)
# Now we can save this
if num_procs > 1: # pragma: no cover
comm.Barrier()
global_save_data = _save_results(
sweep_params,
local_filtered_values,
global_filtered_values,
local_filtered_dict,
global_filtered_dict,
global_filtered_results,
csv_results_file_name,
h5_results_file_name,
debugging_data_dir,
comm,
rank,
num_procs,
interpolate_nan_outputs,
)
return global_save_data
def test_random_build_combinations(self):
comm, rank, num_procs = _init_mpi()
nn = int(1e5)
# Uniform random sampling [lower_limit, upper_limit]
A_param = pyo.Param(initialize=-10.0, mutable=True)
B_param = pyo.Param(initialize=0.0, mutable=True)
C_param = pyo.Param(initialize=10.0, mutable=True)
range_A = [-10.0, 0.0]
range_B = [0.0, 10.0]
range_C = [10.0, 20.0]
param_dict = dict()
param_dict['var_A'] = UniformSample(A_param, range_A[0], range_A[1])
param_dict['var_B'] = UniformSample(B_param, range_B[0], range_B[1])
param_dict['var_C'] = UniformSample(C_param, range_C[0], range_C[1])
global_combo_array = _build_combinations(param_dict,
SamplingType.RANDOM, nn, comm,
rank, num_procs)
assert np.shape(global_combo_array)[0] == nn
assert np.shape(global_combo_array)[1] == len(param_dict)
assert np.all(range_A[0] < global_combo_array[:, 0])
assert np.all(range_B[0] < global_combo_array[:, 1])
assert np.all(range_C[0] < global_combo_array[:, 2])
assert np.all(global_combo_array[:, 0] < range_A[1])
assert np.all(global_combo_array[:, 1] < range_B[1])
assert np.all(global_combo_array[:, 2] < range_C[1])
# Normal random sampling [mean, stdev]
A_param = pyo.Param(initialize=10.0, mutable=True)
B_param = pyo.Param(initialize=100.0, mutable=True)
C_param = pyo.Param(initialize=1000.0, mutable=True)
range_A = [10.0, 5.0]
range_B = [100.0, 50.0]
range_C = [1000.0, 0.0]
param_dict = dict()
param_dict['var_A'] = NormalSample(A_param, range_A[0], range_A[1])
param_dict['var_B'] = NormalSample(B_param, range_B[0], range_B[1])
param_dict['var_C'] = NormalSample(C_param, range_C[0], range_C[1])
global_combo_array = _build_combinations(param_dict,
SamplingType.RANDOM, nn, comm,
rank, num_procs)
assert np.shape(global_combo_array)[0] == nn
assert np.shape(global_combo_array)[1] == len(param_dict)
assert np.mean(global_combo_array[:, 0]) < (range_A[0] + range_A[1])
assert np.mean(global_combo_array[:, 1]) < (range_B[0] + range_B[1])
assert (range_A[0] - range_A[1]) < np.mean(global_combo_array[:, 0])
assert (range_B[0] - range_B[1]) < np.mean(global_combo_array[:, 1])
assert np.all(global_combo_array[:, 2] == range_C[0])
