def default_params_forecasts(model):
'''
Our baseline: forecasts with default SWAN params
'''
closest_params = model.closest_params(
params=SWANParams(drf=1.0, cfw=0.015, stpm=0.00302))
default_params = SWANParams(drf=closest_params[0],
cfw=closest_params[1],
stpm=closest_params[2])
drf_idx, cfw_idx, stpm_idx, fid_time_idx, fid_space_idx = model.params_idxs(
default_params)
forecasts = model.grid[drf_idx, cfw_idx, stpm_idx, fid_time_idx,
fid_space_idx]
return forecasts
def save_archive_history(history, file_name='history.csv'):
test_model = model_all_stations()
with open(file_name, 'w', newline='') as csvfile:
fieldnames = ['idx', 'gen_idx'] + [f'err_{idx + 1}' for idx in range(len(ALL_STATIONS))] \
+ ['drf', 'stpm', 'cfw']
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writeheader()
for gen_idx, gen in enumerate(history):
for ind_idx, ind in enumerate(gen):
row_to_write = {}
idx = gen_idx * len(gen) + ind_idx
row_to_write['idx'] = idx
row_to_write['gen_idx'] = gen_idx
params = test_model.closest_params(ind.genotype)
closest_params = SWANParams(drf=params[0],
cfw=params[1],
stpm=params[2],
fidelity_time=params[3],
fidelity_space=params[4])
# TODO: bug here
metrics = test_model.output(params=closest_params)
for err_idx, err_value in enumerate(metrics):
row_to_write[f'err_{err_idx + 1}'] = err_value
row_to_write['drf'] = ind.genotype.drf
row_to_write['stpm'] = ind.genotype.stpm
row_to_write['cfw'] = ind.genotype.cfw
writer.writerow(row_to_write)
def error_grid(noise_case=0):
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
stations = [1, 2, 3, 4, 5, 6, 7, 8, 9]
ww3_obs_all = \
[obs.time_series() for obs in
wave_watch_results(path_to_results='../../samples/ww-res/', stations=stations)]
model_all = FidelityFakeModel(grid_file=grid,
observations=ww3_obs_all,
stations_to_out=stations,
error=error_rmse_all,
forecasts_path='../../../wind-postproc/out')
with open(f'../../samples/params_rmse_{noise_case}.csv',
mode='w',
newline='') as csv_file:
writer = csv.writer(csv_file, delimiter=',')
header = ['DRF', 'CFW', 'STPM']
error_columns = [f'ERROR_K{station}' for station in stations]
header.extend(error_columns)
writer.writerow(header)
fidelity = 240
for row in grid.rows:
metrics = model_all.output(
params=SWANParams(drf=row.model_params.drf,
cfw=row.model_params.cfw,
stpm=row.model_params.stpm,
fidelity_time=fidelity))
row_to_write = row.model_params.params_list()
row_to_write.extend(metrics)
writer.writerow(row_to_write)
def train(self, mode='lhs', **kwargs):
target = []
if mode is 'lhs':
for feature in self.features:
params = SWANParams(drf=feature[0],
cfw=feature[1],
stpm=feature[2],
fidelity_time=self.fidelity[0],
fidelity_space=self.fidelity[1])
target.append(
self.fake_model.output_from_model(
params=params)[self.station])
target = np.asarray(target)
start_time = datetime.now().strftime(DATE_FORMAT)
print(
f'{start_time}: starting to train kriging model '
f'with {self.points_to_train} points for station: {self.station}')
krig = kriging(self.features, target, name='multikrieg')
krig.train(optimizer='ga')
self.krig = krig
end_time = datetime.now().strftime(DATE_FORMAT)
print(f'{end_time}: finished to train kriging model with'
f' {self.points_to_train} points for station: {self.station}')
def run_genetic_opt(max_gens, pop_size, archive_size, crossover_rate,
mutation_rate, mutation_value_rate, stations, **kwargs):
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in wave_watch_results(path_to_results='../../samples/ww-res/', stations=stations)]
train_model = FidelityFakeModel(grid_file=grid,
observations=ww3_obs,
stations_to_out=stations,
error=error_rmse_all,
forecasts_path='../../../wind-fidelity/*')
test_range = (0, 1)
test_model = model_all_stations(forecasts_range=test_range)
operators = default_operators()
history, archive_history = DefaultSPEA2(
params=DefaultSPEA2.Params(max_gens=max_gens,
pop_size=pop_size,
archive_size=archive_size,
crossover_rate=crossover_rate,
mutation_rate=mutation_rate,
mutation_value_rate=mutation_value_rate),
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators).solution(verbose=False)
exptime2 = str(datetime.datetime.now().time()).replace(":", "-")
# save_archive_history(archive_history, f'rob-exp-bl-{exptime2}.csv')
params = history.last().genotype
if 'save_figures' in kwargs and kwargs['save_figures'] is True:
params = test_model.closest_params(params)
closest_params = SWANParams(drf=params[0],
cfw=params[1],
stpm=params[2],
fidelity_time=params[3],
fidelity_space=params[4])
drf_idx, cfw_idx, stpm_idx, fid_time_idx, fid_space_idx = test_model.params_idxs(
closest_params)
forecasts = test_model.grid[drf_idx, cfw_idx, stpm_idx, fid_time_idx,
fid_space_idx]
plot_results(forecasts=forecasts,
observations=wave_watch_results(
path_to_results='../../samples/ww-res/',
stations=ALL_STATIONS),
baseline=default_params_forecasts(test_model),
save=True,
file_path=kwargs['figure_path'],
values_range=test_range)
return history.last()
def _fixed_params(self, params):
params_fixed = SWANParams(
drf=min(max(params.drf, min(self.grid_file.drf_grid)),
max(self.grid_file.drf_grid)),
cfw=min(max(params.cfw, min(self.grid_file.cfw_grid)),
max(self.grid_file.cfw_grid)),
stpm=min(max(params.stpm, min(self.grid_file.stpm_grid)),
max(self.grid_file.stpm_grid)),
fidelity_time=params.fid_time,
fidelity_space=params.fid_space)
return params_fixed
def crossover(p1, p2, rate):
if random.random() >= rate:
return p1
part1_rate = abs(random.random())
part2_rate = 1 - part1_rate
child_params = SWANParams(
drf=abs(p1.drf * part1_rate + p2.drf * part2_rate),
cfw=abs(p1.cfw * part1_rate + p2.cfw * part2_rate),
stpm=abs(p1.stpm * part1_rate + p2.stpm * part2_rate))
return child_params
def params_and_error_grid(model, station_index, stpm_index):
drf, cfw = np.meshgrid(model.grid_file.drf_grid, model.grid_file.cfw_grid)
stpm_fixed = model.grid_file.stpm_grid[stpm_index]
error = np.ones(shape=drf.shape)
for i in range(drf.shape[0]):
for j in range(drf.shape[1]):
out_by_stations = model.output(params=SWANParams(
drf=drf[i, j], cfw=cfw[i, j], stpm=stpm_fixed))
error[i, j] = out_by_stations[station_index]
return drf, cfw, error
def default_points_by_fidelity(size=10):
points_by_fidelity = {}
for fidelity in fidelity_combinations():
points = []
for _ in range(size):
params = SWANParams.new_instance()
params.fid_time = fidelity[0]
params.fid_space = fidelity[1]
points.append(params)
points_by_fidelity[fidelity] = points
return points_by_fidelity
def retrain_with_new_points(self, new_points):
extended_features = self.features.tolist()
for point in new_points:
extended_features.append([point.drf, point.cfw, point.stpm])
params = SWANParams(drf=point.drf,
cfw=point.cfw,
stpm=point.stpm,
fidelity_time=self.fidelity[0],
fidelity_space=self.fidelity[1])
print(self.fake_model.output_from_model(params=params)[0])
self.points_to_train += len(new_points)
print(f'retrain with new {len(new_points)} features')
self.features = np.asarray(extended_features)
self.train()
def initial_pop_lhs(size, **kwargs):
samples_grid = lhs(PARAMS, size, 'center')
for idx, params_range in enumerate([drf_range, cfw_range, stpm_range]):
samples_grid[:, idx] = norm(loc=np.mean(params_range),
scale=np.std(params_range)).ppf(
samples_grid[:, idx])
population = [
SWANParams(drf=sample[0], cfw=sample[1], stpm=sample[2])
for sample in samples_grid
]
# population = [SWANParams(drf=1.0, cfw=0.015, stpm=0.00302) for sample in samples_grid]
if 'dump' in kwargs and kwargs['dump'] is True:
dump_population(population, kwargs['file_path'])
return population
def plot_params_space():
grid = CSVGridFile('../../samples/wind-exp-params.csv')
forecasts_path = '../../../samples/wind-noice-runs/results/1/'
fake = FidelityFakeModel(grid_file=grid, forecasts_path=forecasts_path)
drf, cfw = np.meshgrid(fake.grid_file.drf_grid, fake.grid_file.cfw_grid)
stpm_fixed = fake.grid_file.stpm_grid[0]
stations = 3
error = np.ones(shape=(stations, drf.shape[0], drf.shape[1]))
for i in range(drf.shape[0]):
for j in range(drf.shape[1]):
diff = fake.output(params=SWANParams(
drf=drf[i, j], cfw=cfw[i, j], stpm=stpm_fixed))
for station in range(stations):
error[station, i, j] = diff[station]
def optimize_test(train_stations,
max_gens,
pop_size,
archive_size,
crossover_rate,
mutation_rate,
mutation_value_rate,
sur_points,
plot_figures=True):
train_range = (0, 1)
test_range = (0, 1)
grid = CSVGridFile('../../samples/wind-exp-params-new.csv')
ww3_obs = \
[obs.time_series() for obs in
wave_watch_results(path_to_results='../../samples/ww-res/', stations=train_stations)]
error = error_rmse_all
train_model = FidelityFakeModel(grid_file=grid,
observations=ww3_obs,
stations_to_out=train_stations,
error=error,
forecasts_path='../../../2fidelity/*',
forecasts_range=train_range,
is_surrogate=True,
sur_points=sur_points)
operators = default_operators()
history, archive_history = DefaultSPEA2(
params=DefaultSPEA2.Params(max_gens,
pop_size=pop_size,
archive_size=archive_size,
crossover_rate=crossover_rate,
mutation_rate=mutation_rate,
mutation_value_rate=mutation_value_rate),
objectives=partial(calculate_objectives_interp, train_model),
evolutionary_operators=operators).solution(verbose=True)
params = history.last().genotype
if plot_figures:
test_model = model_all_stations(forecasts_range=test_range)
params = test_model.closest_params(params)
closest_params = SWANParams(drf=params[0],
cfw=params[1],
stpm=params[2],
fidelity_time=params[3],
fidelity_space=params[4])
drf_idx, cfw_idx, stpm_idx, fid_time_idx, fid_space_idx = test_model.params_idxs(
closest_params)
forecasts = test_model.grid[drf_idx, cfw_idx, stpm_idx, fid_time_idx,
fid_space_idx]
plot_results(forecasts=forecasts,
observations=wave_watch_results(
path_to_results='../../samples/ww-res/',
stations=ALL_STATIONS),
baseline=default_params_forecasts(test_model))
plot_population_movement(archive_history, grid)
return history.last().error_value
def reference_metrics():
return all_error_metrics(params=SWANParams(drf=1.0,
cfw=0.015,
stpm=0.00302),
models_to_tests=init_models_to_tests())
def default_initial_pop(size):
return [SWANParams.new_instance() for _ in range(size)]
def _swan_params(cls, csv_row):
return SWANParams(drf=float(csv_row['DRF']),
cfw=float(csv_row['CFW']),
stpm=float(csv_row['STPM']))
def _init_grids(self):
self.grid = self._empty_grid()
files = forecast_files_from_dir(self.forecasts_path)
if not files:
raise FileNotFoundError("EMPTY FORECAST")
stations = files_by_stations(
files,
noise_run=self.noise_run,
stations=[str(st) for st in self.stations])
files_by_run_idx = dict()
for station in stations:
for file in station:
_, name = os.path.split(file)
_, _, run_idx = extracted_forecast_params(file_name=name)
files_by_run_idx[file] = run_idx
for row in self.grid_file.rows:
run_idx = row.id
forecasts_files = sorted([
key for key in files_by_run_idx.keys()
if files_by_run_idx[key] == run_idx
])
files_by_fidelity = self._files_by_fidelity(forecasts_files)
for fidelity in files_by_fidelity:
files = files_by_fidelity[fidelity]
forecasts = []
for idx, file_name in enumerate(files):
forecasts.append(
FidelityFakeModel.Forecast(
self.stations[idx],
ForecastFile(path=file_name),
range_values=self.forecasts_range))
fid_time, fid_space = fidelity
drf_idx, cfw_idx, stpm_idx, fid_time_idx, fid_space_idx = self.params_idxs(
params=SWANParams(drf=row.model_params.drf,
cfw=row.model_params.cfw,
stpm=row.model_params.stpm,
fidelity_time=fid_time,
fidelity_space=fid_space))
self.grid[drf_idx, cfw_idx, stpm_idx, fid_time_idx,
fid_space_idx] = forecasts
# empty array
self.err_grid = np.zeros(shape=self.grid.shape + (len(stations), ))
# calc fitness for every point
st_set_id = ("-".join(str(self.stations)))
file_path = f'grid-saved-{self.error.__name__}_range_{self.forecasts_range}_st{st_set_id}.pik'
grid_file_path = os.path.join(GRID_PATH, file_path)
if not os.path.isfile(grid_file_path):
grid_idxs = self.__grid_idxs()
for i, j, k, m, n in grid_idxs:
forecasts = [forecast for forecast in self.grid[i, j, k, m, n]]
for forecast, observation in zip(forecasts, self.observations):
station_idx = forecasts.index(forecast)
obs_in_range = observations_from_range(
observation, self.forecasts_range)
self.err_grid[i, j, k, m, n, station_idx] = self.error(
forecast, obs_in_range)
pickle_out = open(grid_file_path, 'wb')
pickle.dump(self.err_grid, pickle_out)
pickle_out.close()
print(f"FITNESS GRID SAVED, file_name: {grid_file_path}")
else:
with open(grid_file_path, 'rb') as f:
self.err_grid = pickle.load(f)