def __init__(self,
exp_prefix,
est_params,
sim_params,
observations,
keys_of_interest,
n_mc_samples=10**7,
n_x_cond=5,
n_seeds=5,
use_gpu=True,
tail_measures=True):
assert est_params and exp_prefix and sim_params and keys_of_interest
assert observations.all()
# every simulator configuration will be run multiple times with different randomness seeds
sim_params = _add_seeds_to_sim_params(n_seeds, sim_params)
self.observations = observations
self.n_mc_samples = n_mc_samples
self.n_x_cond = n_x_cond
self.keys_of_interest = keys_of_interest
self.exp_prefix = exp_prefix
self.use_gpu = use_gpu
self.tail_measures = tail_measures
logger.configure(log_directory=config.DATA_DIR,
prefix=exp_prefix,
color='green')
''' ---------- Either load or generate the configs ----------'''
config_pkl_path = os.path.join(logger.log_directory, logger.prefix,
EXP_CONFIG_FILE)
if os.path.isfile(config_pkl_path):
logger.log("{:<70s} {:<30s}".format(
"Loading experiment previous configs from file: ",
config_pkl_path))
self.configs = logger.load_pkl(EXP_CONFIG_FILE)
else:
logger.log("{:<70s} {:<30s}".format(
"Generating and storing experiment configs under: ",
config_pkl_path))
self.configs = self._generate_configuration_variants(
est_params, sim_params)
logger.dump_pkl(data=self.configs, path=EXP_CONFIG_FILE)
''' ---------- Either load already existing results or start a new result collection ---------- '''
results_pkl_path = os.path.join(logger.log_directory, logger.prefix,
RESULTS_FILE)
if os.path.isfile(results_pkl_path):
logger.log_line("{:<70s} {:<30s}".format("Continue with: ",
results_pkl_path))
self.gof_single_res_collection = dict(
logger.load_pkl_log(RESULTS_FILE))
else: # start from scratch
self.gof_single_res_collection = {}
self.gof_results = GoodnessOfFitResults(self.gof_single_res_collection)
def test_load_pkl_log(setup):
import numpy
d1 = numpy.random.randn(20, 10)
logger.log_pkl(d1, 'test_file.pkl')
sleep(1.0)
d2 = numpy.random.randn(20, 10)
logger.log_pkl(d2, 'test_file.pkl')
sleep(1.0)
data = logger.load_pkl_log('test_file.pkl')
assert len(data) == 2, "data should contain two arrays"
assert numpy.array_equal(data[0], d1), "first should be the same as d1"
assert numpy.array_equal(data[1], d2), "first should be the same as d2"
def test_log_data(setup):
import numpy
d1 = numpy.random.randn(20, 10)
logger.log_pkl(d1, 'test_file.pkl')
sleep(1.0)
d2 = numpy.random.randn(20, 10)
logger.dump_pkl(d2, 'test_file.pkl')
sleep(1.0)
data = logger.load_pkl_log('test_file.pkl')
assert len(
data
) == 1, "data should contain only one array because we overwrote it."
assert numpy.array_equal(data[0], d2), "first should be the same as d2"
def get_density_plots(estimators_list,
simulators_dict,
path_to_results,
exp_prefix="question1_noise_reg_x",
task_ids=None):
"""
This function allows to compare plots from estimators and simulators (i.e. fitted and true densities). Two modes are currently available:
1) by specifying estimators and simulator, the function picks one result pair randomly that matches the given simulator/estimator
selection
2) by specifying the task_ids as list, it is possible to pick specific plots to compare
Args:
estimators: a list containing strings of estimators to be evaluated, e.g. ['KernelMixtureNetwork', 'MixtureDensityNetwork']
simulators: a dict containing specifications of a simulator under which the estimators shall be compared, e.g.
{'heteroscedastic': True, 'random_seed': 20, 'std': 1, 'simulator': 'EconDensity'}
path_to_results: absolute path to where the dumped model files are stored
exp_prefix: specifies the task question
Returns:
A list of figures for fitted and true densities.
"""
if task_ids is not None:
assert type(task_ids) == list
assert len(task_ids) == len(estimators_list)
RESULTS_FILE = 'results.pkl'
logger.configure(path_to_results, exp_prefix)
results_from_pkl_file = dict(logger.load_pkl_log(RESULTS_FILE))
gof_result = GoodnessOfFitResults(
single_results_dict=results_from_pkl_file)
results_df = gof_result.generate_results_dataframe(
base_experiment.KEYS_OF_INTEREST)
""" load model's estimators """
if task_ids is None:
models_of_interest = {
k: v
for k, v in gof_result.single_results_dict.items()
if v.probabilistic_model_params == simulators_dict and v.ndim_x +
v.ndim_y == 2
}
models = [
ConfigRunner.load_dumped_estimator(
take_of_type(1, estimator_str, models_of_interest))
for estimator_str in estimators_list
]
else:
models = [
ConfigRunner.load_dumped_estimators(gof_result, task_id=task_ids)
]
""" load model's simulators """
# todo: implement when simulator dumps exist
figs = []
for model in models:
graph = model.estimator.sess.graph
sess = tf.Session(graph=graph)
with sess:
sess.run(tf.global_variables_initializer())
model.estimator.sess = sess
""" fitted density figures"""
plt.suptitle(model.estimator.name)
fig_fitted = model.estimator.plot3d()
figs.append(fig_fitted)
""" true density figures """
# todo: use newly dumped simulators
sess.close()
return figs
'y_noise_std': [0.1],
},
}
simulators_params = {'LinearStudentT': {'ndim_x': [10]}}
observations = 100 * np.logspace(2, 6, num=8, base=2.0, dtype=np.int32)
return estimator_params, simulators_params, observations
if __name__ == '__main__':
estimator_params, simulators_params, observations = question4()
load = base_experiment.launch_experiment(estimator_params,
simulators_params,
observations,
EXP_PREFIX,
n_mc_samples=N_MC_SAMPLES,
tail_measures=False)
if load:
logger.configure(config.DATA_DIR, EXP_PREFIX)
results_from_pkl_file = dict(logger.load_pkl_log(RESULTS_FILE))
gof_result = GoodnessOfFitResults(
single_results_dict=results_from_pkl_file)
results_df = gof_result.generate_results_dataframe(
base_experiment.KEYS_OF_INTEREST)
gof_result = ConfigRunner.load_dumped_estimators(gof_result)
def test_load_module(setup, test_module):
result, = logger.load_pkl_log(f"modules/{0:04d}_Test.pkl")
import numpy as np
assert (result['var_1'] == np.ones(
[100, 2])).all(), "should be the same as test data"