Merge pull request #7258 from freqtrade/feat/hyp_optinal_indicator

Add flag to move hyperopt populate_indicators to epoch
This commit is contained in:
Matthias 2022-08-27 09:21:16 +02:00 committed by GitHub
commit 6686489c06
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
10 changed files with 105 additions and 22 deletions

View File

@ -40,7 +40,8 @@ pip install -r requirements-hyperopt.txt
```
usage: freqtrade hyperopt [-h] [-v] [--logfile FILE] [-V] [-c PATH] [-d PATH]
[--userdir PATH] [-s NAME] [--strategy-path PATH]
[--recursive-strategy-search] [-i TIMEFRAME]
[--recursive-strategy-search] [--freqaimodel NAME]
[--freqaimodel-path PATH] [-i TIMEFRAME]
[--timerange TIMERANGE]
[--data-format-ohlcv {json,jsongz,hdf5}]
[--max-open-trades INT]
@ -53,7 +54,7 @@ usage: freqtrade hyperopt [-h] [-v] [--logfile FILE] [-V] [-c PATH] [-d PATH]
[--print-all] [--no-color] [--print-json] [-j JOBS]
[--random-state INT] [--min-trades INT]
[--hyperopt-loss NAME] [--disable-param-export]
[--ignore-missing-spaces]
[--ignore-missing-spaces] [--analyze-per-epoch]
optional arguments:
-h, --help show this help message and exit
@ -129,6 +130,7 @@ optional arguments:
--ignore-missing-spaces, --ignore-unparameterized-spaces
Suppress errors for any requested Hyperopt spaces that
do not contain any parameters.
--analyze-per-epoch Run populate_indicators once per epoch.
Common arguments:
-v, --verbose Verbose mode (-vv for more, -vvv to get all messages).
@ -154,6 +156,10 @@ Strategy arguments:
--recursive-strategy-search
Recursively search for a strategy in the strategies
folder.
--freqaimodel NAME Specify a custom freqaimodels.
--freqaimodel-path PATH
Specify additional lookup path for freqaimodels.
```
### Hyperopt checklist
@ -185,7 +191,7 @@ Rarely you may also need to create a [nested class](advanced-hyperopt.md#overrid
### Hyperopt execution logic
Hyperopt will first load your data into memory and will then run `populate_indicators()` once per Pair to generate all indicators.
Hyperopt will first load your data into memory and will then run `populate_indicators()` once per Pair to generate all indicators, unless `--analyze-per-epoch` is specified.
Hyperopt will then spawn into different processes (number of processors, or `-j <n>`), and run backtesting over and over again, changing the parameters that are part of the `--spaces` defined.
@ -426,9 +432,10 @@ While this strategy is most likely too simple to provide consistent profit, it s
`range` property may also be used with `DecimalParameter` and `CategoricalParameter`. `RealParameter` does not provide this property due to infinite search space.
??? Hint "Performance tip"
By doing the calculation of all possible indicators in `populate_indicators()`, the calculation of the indicator happens only once for every parameter.
While this may slow down the hyperopt startup speed, the overall performance will increase as the Hyperopt execution itself may pick the same value for multiple epochs (changing other values).
You should however try to use space ranges as small as possible. Every new column will require more memory, and every possibility hyperopt can try will increase the search space.
During normal hyperopting, indicators are calculated once and supplied to each epoch, linearly increasing RAM usage as a factor of increasing cores. As this also has performance implications, hyperopt provides `--analyze-per-epoch` which will move the execution of `populate_indicators()` to the epoch process, calculating a single value per parameter per epoch instead of using the `.range` functionality. In this case, `.range` functionality will only return the actually used value. This will reduce RAM usage, but increase CPU usage. However, your hyperopting run will be less likely to fail due to Out Of Memory (OOM) issues.
In either case, you should try to use space ranges as small as possible this will improve CPU/RAM usage in both scenarios.
## Optimizing protections
@ -879,6 +886,7 @@ To combat these, you have multiple options:
* Avoid using `--timeframe-detail` (this loads a lot of additional data into memory).
* Reduce the number of parallel processes (`-j <n>`).
* Increase the memory of your machine.
* Use `--analyze-per-epoch` if you're using a lot of parameters with `.range` functionality.
## The objective has been evaluated at this point before.

View File

@ -34,7 +34,7 @@ ARGS_HYPEROPT = ARGS_COMMON_OPTIMIZE + ["hyperopt", "hyperopt_path",
"print_colorized", "print_json", "hyperopt_jobs",
"hyperopt_random_state", "hyperopt_min_trades",
"hyperopt_loss", "disableparamexport",
"hyperopt_ignore_missing_space"]
"hyperopt_ignore_missing_space", "analyze_per_epoch"]
ARGS_EDGE = ARGS_COMMON_OPTIMIZE + ["stoploss_range"]

View File

@ -255,6 +255,13 @@ AVAILABLE_CLI_OPTIONS = {
nargs='+',
default='default',
),
"analyze_per_epoch": Arg(
'--analyze-per-epoch',
help='Run populate_indicators once per epoch.',
action='store_true',
default=False,
),
"print_all": Arg(
'--print-all',
help='Print all results, not only the best ones.',

View File

@ -302,6 +302,9 @@ class Configuration:
self._args_to_config(config, argname='spaces',
logstring='Parameter -s/--spaces detected: {}')
self._args_to_config(config, argname='analyze_per_epoch',
logstring='Parameter --analyze-per-epoch detected.')
self._args_to_config(config, argname='print_all',
logstring='Parameter --print-all detected ...')

View File

@ -3,6 +3,7 @@ from freqtrade.enums.backteststate import BacktestState
from freqtrade.enums.candletype import CandleType
from freqtrade.enums.exitchecktuple import ExitCheckTuple
from freqtrade.enums.exittype import ExitType
from freqtrade.enums.hyperoptstate import HyperoptState
from freqtrade.enums.marginmode import MarginMode
from freqtrade.enums.ordertypevalue import OrderTypeValues
from freqtrade.enums.rpcmessagetype import RPCMessageType

View File

@ -0,0 +1,12 @@
from enum import Enum
class HyperoptState(Enum):
""" Hyperopt states """
STARTUP = 1
DATALOAD = 2
INDICATORS = 3
OPTIMIZE = 4
def __str__(self):
return f"{self.name.lower()}"

View File

@ -24,13 +24,15 @@ from pandas import DataFrame
from freqtrade.constants import DATETIME_PRINT_FORMAT, FTHYPT_FILEVERSION, LAST_BT_RESULT_FN
from freqtrade.data.converter import trim_dataframes
from freqtrade.data.history import get_timerange
from freqtrade.enums import HyperoptState
from freqtrade.exceptions import OperationalException
from freqtrade.misc import deep_merge_dicts, file_dump_json, plural
from freqtrade.optimize.backtesting import Backtesting
# Import IHyperOpt and IHyperOptLoss to allow unpickling classes from these modules
from freqtrade.optimize.hyperopt_auto import HyperOptAuto
from freqtrade.optimize.hyperopt_loss_interface import IHyperOptLoss
from freqtrade.optimize.hyperopt_tools import HyperoptTools, hyperopt_serializer
from freqtrade.optimize.hyperopt_tools import (HyperoptStateContainer, HyperoptTools,
hyperopt_serializer)
from freqtrade.optimize.optimize_reports import generate_strategy_stats
from freqtrade.resolvers.hyperopt_resolver import HyperOptLossResolver
@ -74,10 +76,14 @@ class Hyperopt:
self.dimensions: List[Dimension] = []
self.config = config
self.min_date: datetime
self.max_date: datetime
self.backtesting = Backtesting(self.config)
self.pairlist = self.backtesting.pairlists.whitelist
self.custom_hyperopt: HyperOptAuto
self.analyze_per_epoch = self.config.get('analyze_per_epoch', False)
HyperoptStateContainer.set_state(HyperoptState.STARTUP)
if not self.config.get('hyperopt'):
self.custom_hyperopt = HyperOptAuto(self.config)
@ -290,6 +296,7 @@ class Hyperopt:
Called once per epoch to optimize whatever is configured.
Keep this function as optimized as possible!
"""
HyperoptStateContainer.set_state(HyperoptState.OPTIMIZE)
backtest_start_time = datetime.now(timezone.utc)
params_dict = self._get_params_dict(self.dimensions, raw_params)
@ -321,6 +328,10 @@ class Hyperopt:
with self.data_pickle_file.open('rb') as f:
processed = load(f, mmap_mode='r')
if self.analyze_per_epoch:
# Data is not yet analyzed, rerun populate_indicators.
processed = self.advise_and_trim(processed)
bt_results = self.backtesting.backtest(
processed=processed,
start_date=self.min_date,
@ -406,22 +417,33 @@ class Hyperopt:
def _set_random_state(self, random_state: Optional[int]) -> int:
return random_state or random.randint(1, 2**16 - 1)
def prepare_hyperopt_data(self) -> None:
data, timerange = self.backtesting.load_bt_data()
self.backtesting.load_bt_data_detail()
logger.info("Dataload complete. Calculating indicators")
def advise_and_trim(self, data: Dict[str, DataFrame]) -> Dict[str, DataFrame]:
preprocessed = self.backtesting.strategy.advise_all_indicators(data)
# Trim startup period from analyzed dataframe to get correct dates for output.
processed = trim_dataframes(preprocessed, timerange, self.backtesting.required_startup)
processed = trim_dataframes(preprocessed, self.timerange, self.backtesting.required_startup)
self.min_date, self.max_date = get_timerange(processed)
return processed
logger.info(f'Hyperopting with data from {self.min_date.strftime(DATETIME_PRINT_FORMAT)} '
f'up to {self.max_date.strftime(DATETIME_PRINT_FORMAT)} '
f'({(self.max_date - self.min_date).days} days)..')
# Store non-trimmed data - will be trimmed after signal generation.
dump(preprocessed, self.data_pickle_file)
def prepare_hyperopt_data(self) -> None:
HyperoptStateContainer.set_state(HyperoptState.DATALOAD)
data, self.timerange = self.backtesting.load_bt_data()
self.backtesting.load_bt_data_detail()
logger.info("Dataload complete. Calculating indicators")
if not self.analyze_per_epoch:
HyperoptStateContainer.set_state(HyperoptState.INDICATORS)
preprocessed = self.advise_and_trim(data)
logger.info(f'Hyperopting with data from '
f'{self.min_date.strftime(DATETIME_PRINT_FORMAT)} '
f'up to {self.max_date.strftime(DATETIME_PRINT_FORMAT)} '
f'({(self.max_date - self.min_date).days} days)..')
# Store non-trimmed data - will be trimmed after signal generation.
dump(preprocessed, self.data_pickle_file)
else:
dump(data, self.data_pickle_file)
def get_asked_points(self, n_points: int) -> Tuple[List[List[Any]], List[bool]]:
"""

View File

@ -13,6 +13,7 @@ from colorama import Fore, Style
from pandas import isna, json_normalize
from freqtrade.constants import FTHYPT_FILEVERSION, USERPATH_STRATEGIES
from freqtrade.enums import HyperoptState
from freqtrade.exceptions import OperationalException
from freqtrade.misc import deep_merge_dicts, round_coin_value, round_dict, safe_value_fallback2
from freqtrade.optimize.hyperopt_epoch_filters import hyperopt_filter_epochs
@ -32,6 +33,15 @@ def hyperopt_serializer(x):
return str(x)
class HyperoptStateContainer():
""" Singleton class to track state of hyperopt"""
state: HyperoptState = HyperoptState.OPTIMIZE
@classmethod
def set_state(cls, value: HyperoptState):
cls.state = value
class HyperoptTools():
@staticmethod

View File

@ -7,6 +7,9 @@ from abc import ABC, abstractmethod
from contextlib import suppress
from typing import Any, Optional, Sequence, Union
from freqtrade.enums.hyperoptstate import HyperoptState
from freqtrade.optimize.hyperopt_tools import HyperoptStateContainer
with suppress(ImportError):
from skopt.space import Integer, Real, Categorical
@ -57,6 +60,13 @@ class BaseParameter(ABC):
Get-space - will be used by Hyperopt to get the hyperopt Space
"""
def can_optimize(self):
return (
self.in_space
and self.optimize
and HyperoptStateContainer.state != HyperoptState.OPTIMIZE
)
class NumericParameter(BaseParameter):
""" Internal parameter used for Numeric purposes """
@ -133,7 +143,7 @@ class IntParameter(NumericParameter):
Returns a List with 1 item (`value`) in "non-hyperopt" mode, to avoid
calculating 100ds of indicators.
"""
if self.in_space and self.optimize:
if self.can_optimize():
# Scikit-optimize ranges are "inclusive", while python's "range" is exclusive
return range(self.low, self.high + 1)
else:
@ -212,7 +222,7 @@ class DecimalParameter(NumericParameter):
Returns a List with 1 item (`value`) in "non-hyperopt" mode, to avoid
calculating 100ds of indicators.
"""
if self.in_space and self.optimize:
if self.can_optimize():
low = int(self.low * pow(10, self._decimals))
high = int(self.high * pow(10, self._decimals)) + 1
return [round(n * pow(0.1, self._decimals), self._decimals) for n in range(low, high)]
@ -261,7 +271,7 @@ class CategoricalParameter(BaseParameter):
Returns a List with 1 item (`value`) in "non-hyperopt" mode, to avoid
calculating 100ds of indicators.
"""
if self.in_space and self.optimize:
if self.can_optimize():
return self.opt_range
else:
return [self.value]

View File

@ -12,7 +12,9 @@ from freqtrade.configuration import TimeRange
from freqtrade.data.dataprovider import DataProvider
from freqtrade.data.history import load_data
from freqtrade.enums import ExitCheckTuple, ExitType, SignalDirection
from freqtrade.enums.hyperoptstate import HyperoptState
from freqtrade.exceptions import OperationalException, StrategyError
from freqtrade.optimize.hyperopt_tools import HyperoptStateContainer
from freqtrade.optimize.space import SKDecimal
from freqtrade.persistence import PairLocks, Trade
from freqtrade.resolvers import StrategyResolver
@ -859,7 +861,9 @@ def test_strategy_safe_wrapper_trade_copy(fee):
def test_hyperopt_parameters():
HyperoptStateContainer.set_state(HyperoptState.INDICATORS)
from skopt.space import Categorical, Integer, Real
with pytest.raises(OperationalException, match=r"Name is determined.*"):
IntParameter(low=0, high=5, default=1, name='hello')
@ -937,6 +941,12 @@ def test_hyperopt_parameters():
assert list(boolpar.range) == [True, False]
HyperoptStateContainer.set_state(HyperoptState.OPTIMIZE)
assert len(list(intpar.range)) == 1
assert len(list(fltpar.range)) == 1
assert len(list(catpar.range)) == 1
assert len(list(boolpar.range)) == 1
def test_auto_hyperopt_interface(default_conf):
default_conf.update({'strategy': 'HyperoptableStrategyV2'})