adelorenzo/stable
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

better defaults

Browse Source
This commit is contained in:
orehunt 2020-03-17 14:20:41 +01:00
parent a5b44de0f6
commit 8d03887b02
1 changed files with 19 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
freqtrade/optimize/hyperopt.py
View File

@ -154,6 +154,8 @@ class Hyperopt:
backend.manager = Manager() backend.manager = Manager()
self.mode = self.config.get('mode', 'single') self.mode = self.config.get('mode', 'single')
self.shared = False self.shared = False
# models are only needed for posterior eval
self.n_models = 1
if self.mode in ('multi', 'shared'): if self.mode in ('multi', 'shared'):
self.multi = True self.multi = True
if self.mode == 'shared': if self.mode == 'shared':
@ -161,19 +163,19 @@ class Hyperopt:
backend.optimizers = backend.manager.Queue() backend.optimizers = backend.manager.Queue()
backend.results_board = backend.manager.Queue(maxsize=1) backend.results_board = backend.manager.Queue(maxsize=1)
backend.results_board.put({}) backend.results_board.put({})
default_n_points = 2
self.opt_base_estimator = 'GBRT' self.opt_base_estimator = 'GBRT'
self.opt_acq_optimizer = 'sampling' self.opt_acq_optimizer = 'sampling'
# in multi opt one model is enough
self.n_models = 1
default_n_points = 2
else: else:
self.multi = False self.multi = False
backend.results = backend.manager.Queue() backend.results = backend.manager.Queue()
self.opt_base_estimator = 'GP' self.opt_base_estimator = 'ET'
self.opt_acq_optimizer = 'lbfgs' self.opt_acq_optimizer = 'sampling'
# models are only needed for posterior eval
self.n_models = min(16, self.n_jobs)
default_n_points = 1 default_n_points = 1
# The GaussianProcessRegressor is heavy, which makes it not a good default
# however longer backtests might make it a better tradeoff
# self.opt_base_estimator = 'GP'
# self.opt_acq_optimizer = 'lbfgs'
# in single opt assume runs are expensive so default to 1 point per ask # in single opt assume runs are expensive so default to 1 point per ask
self.n_points = self.config.get('n_points', default_n_points) self.n_points = self.config.get('n_points', default_n_points)
@ -235,17 +237,21 @@ class Hyperopt:
num_trials = len(self.trials) num_trials = len(self.trials)
print() print()
if num_trials > self.num_trials_saved: if num_trials > self.num_trials_saved:
logger.info(f"Saving {num_trials} {plural(num_trials, 'epoch')}.") logger.debug(f"Saving {num_trials} {plural(num_trials, 'epoch')}.")
dump(self.trials, self.trials_file) dump(self.trials, self.trials_file)
self.num_trials_saved = num_trials self.num_trials_saved = num_trials
self.save_opts() self.save_opts()
if final: if final:
logger.info(f"{num_trials} {plural(num_trials, 'epoch')} " logger.debug(f"{num_trials} {plural(num_trials, 'epoch')} "
f"saved to '{self.trials_file}'.") f"saved to '{self.trials_file}'.")
def save_opts(self) -> None: def save_opts(self) -> None:
""" Save optimizers state to disk. The minimum required state could also be constructed """
from the attributes [ models, space, rng ] with Xi, yi loaded from trials """ Save optimizers state to disk. The minimum required state could also be constructed
from the attributes [ models, space, rng ] with Xi, yi loaded from trials.
All we really care about are [rng, Xi, yi] since models are never passed over queues
and space is dependent on dimensions matching with hyperopt config
"""
# synchronize with saved trials # synchronize with saved trials
opts = [] opts = []
n_opts = 0 n_opts = 0
@ -259,7 +265,7 @@ class Hyperopt:
if self.opt: if self.opt:
n_opts = 1 n_opts = 1
opts = [self.opt] opts = [self.opt]
logger.info(f"Saving {n_opts} {plural(n_opts, 'optimizer')}.") logger.debug(f"Saving {n_opts} {plural(n_opts, 'optimizer')}.")
dump(opts, self.opts_file) dump(opts, self.opts_file)
@staticmethod @staticmethod
@ -874,7 +880,7 @@ class Hyperopt:
if self.max_epoch > self.search_space_size: if self.max_epoch > self.search_space_size:
self.max_epoch = self.search_space_size self.max_epoch = self.search_space_size
print() print()
logger.info(f'Max epoch set to: {self.epochs_limit()}') logger.debug(f'Max epoch set to: {self.epochs_limit()}')
def setup_optimizers(self): def setup_optimizers(self):
""" Setup the optimizers objects, try to load from disk, or create new ones """ """ Setup the optimizers objects, try to load from disk, or create new ones """