issue resolved: opt.tell() is now called in the parent process

freqtrade/optimize/hyperopt.py

@@ -3,6 +3,7 @@
 """
 This module contains the hyperopt logic
 """
+import functools
 import logging
 import multiprocessing
 import os
@@ -16,6 +17,8 @@ from pprint import pprint
 from typing import Any, Dict, List
 
 from joblib import Parallel, delayed, dump, load, wrap_non_picklable_objects
+from joblib._parallel_backends import LokyBackend
+from joblib import register_parallel_backend, parallel_backend
 from pandas import DataFrame
 from skopt import Optimizer
 from skopt.space import Dimension
@@ -37,6 +40,9 @@ TICKERDATA_PICKLE = os.path.join('user_data', 'hyperopt_tickerdata.pkl')
 EVALS_FRAME = 100
 
 
+_hyperopt = None
+
+
 class Hyperopt(Backtesting):
     """
     Hyperopt class, this class contains all the logic to run a hyperopt simulation
@@ -69,6 +75,9 @@ class Hyperopt(Backtesting):
         self.trials_file = os.path.join('user_data', 'hyperopt_results.pickle')
         self.trials: List = []
 
+        self.opt = None
+        self.f_val: List = []
+
     def get_args(self, params):
         dimensions = self.hyperopt_space()
         # Ensure the number of dimensions match
@@ -114,7 +123,7 @@ class Hyperopt(Backtesting):
             logger.info('ROI table:')
             pprint(self.custom_hyperopt.generate_roi_table(best_result['params']), indent=4)
 
-    def log_results_immediate(self) -> None:
+    def log_results_immediate(self, n) -> None:
         print('.', end='')
         sys.stdout.flush()
 
@@ -169,7 +178,21 @@ class Hyperopt(Backtesting):
         return spaces
 
     def generate_optimizer(self, _params: Dict) -> Dict:
+        """
+        Process params (a point in the hyper options space generated by the
+        optimizer.ask() optimizer method) and run backtest() for it.
+
+        This method is executed in one of the joblib worker processes.
+
+        The optimizer.tell() function is executed in the joblib parent process
+        in order to tell results to single instance of the optimizer rather than
+        its copies serialized in the joblib worker processes.
+
+        It receives result of this method as a part of the joblib batch completion
+        callback function.
+        """
         params = self.get_args(_params)
+
         if self.has_space('roi'):
             self.strategy.minimal_roi = self.custom_hyperopt.generate_roi_table(params)
 
@@ -203,19 +226,14 @@ class Hyperopt(Backtesting):
         trade_count = len(results.index)
         trade_duration = results.trade_duration.mean()
 
-        if trade_count == 0:
+        loss = MAX_LOSS if trade_count == 0 else \
-            return {
+            self.calculate_loss(total_profit, trade_count, trade_duration)
-                'loss': MAX_LOSS,
-                'params': params,
-                'result': result_explanation,
-            }
-
-        loss = self.calculate_loss(total_profit, trade_count, trade_duration)
 
         return {
             'loss': loss,
             'params': params,
             'result': result_explanation,
+            'asked': _params,
         }
 
     def format_results(self, results: DataFrame) -> str:
@@ -239,20 +257,35 @@ class Hyperopt(Backtesting):
             base_estimator="ET",
             acq_optimizer="auto",
             n_initial_points=30,
-            acq_optimizer_kwargs={'n_jobs': cpu_count}
+            acq_optimizer_kwargs={'n_jobs': cpu_count},
+            random_state=777,
         )
 
-    def run_optimizer_parallel(self, parallel, opt, tries: int, first_try: int) -> List:
+    def run_optimizer_parallel(self, parallel, tries: int, first_try: int) -> List:
         result = parallel(delayed(
-                          wrap_non_picklable_objects(self.parallel_tell_and_log))
+                          wrap_non_picklable_objects(self.parallel_objective))
-                          (opt, i, opt.ask()) for i in range(first_try, first_try + tries))
+                          (asked, i) for asked, i in
+                          zip(self.opt_generator(), range(first_try, first_try + tries)))
         return result
 
-    def parallel_tell_and_log(self, opt, i, asked):
+    def opt_generator(self):
-        f_val = self.generate_optimizer(asked)
+        while True:
-        opt.tell(asked, f_val['loss'])
+            if self.f_val:
-        self.log_results_immediate()
+                # print("opt.tell(): ",
-        return f_val
+                #       [v['asked'] for v in self.f_val], [v['loss'] for v in self.f_val])
+                functools.partial(
+                    self.opt.tell,
+                    ([v['asked'] for v in self.f_val], [v['loss'] for v in self.f_val])
+                )
+                self.f_val = []
+            yield self.opt.ask()
+
+    def parallel_objective(self, asked, n):
+        self.log_results_immediate(n)
+        return self.generate_optimizer(asked)
+
+    def parallel_callback(self, f_val):
+        self.f_val.extend(f_val)
 
     def load_previous_results(self):
         """ read trials file if we have one """
@@ -283,12 +316,14 @@ class Hyperopt(Backtesting):
         cpus = multiprocessing.cpu_count()
         logger.info(f'Found {cpus} CPU cores. Let\'s make them scream!')
 
-        opt = self.get_optimizer(cpus)
+        self.opt = self.get_optimizer(cpus)
 
         frames = self.total_tries // EVALS_FRAME
         last_frame_len = self.total_tries % EVALS_FRAME
 
         try:
+            register_parallel_backend('custom', CustomImmediateResultBackend)
+            with parallel_backend('custom'):
                 with Parallel(n_jobs=cpus, verbose=0) as parallel:
                     for frame in range(frames + 1):
                         frame_start = frame * EVALS_FRAME
@@ -296,7 +331,7 @@ class Hyperopt(Backtesting):
                         print(f"\n{frame_start+1}-{frame_start+frame_len}"
                               f"/{self.total_tries}: ", end='')
                         f_val = self.run_optimizer_parallel(
-                            parallel, opt,
+                            parallel,
                             frame_len,
                             frame_start
                         )
@@ -337,5 +372,32 @@ def start(args: Namespace) -> None:
             "to understand how to configure hyperopt.")
         raise ValueError("--strategy configured but not supported for hyperopt")
     # Initialize backtesting object
-    hyperopt = Hyperopt(config)
+    global _hyperopt
-    hyperopt.start()
+    _hyperopt = Hyperopt(config)
+    _hyperopt.start()
+
+
+class MultiCallback:
+    def __init__(self, *callbacks):
+        self.callbacks = [cb for cb in callbacks if cb]
+
+    def __call__(self, out):
+        for cb in self.callbacks:
+            cb(out)
+
+
+class CustomImmediateResultBackend(LokyBackend):
+    def callback(self, result):
+        """
+        Our custom completion callback. Executed in the parent process.
+        Use it to run Optimizer.tell() with immediate results of the backtest()
+        evaluated in the joblib worker process.
+        """
+        # Fetch results from the Future object passed to us.
+        # Future object is assumed to be 'done' already.
+        f_val = result.result().copy()
+        _hyperopt.parallel_callback(f_val)
+
+    def apply_async(self, func, callback=None):
+        cbs = MultiCallback(callback, self.callback)
+        return super().apply_async(func, cbs)

freqtrade/tests/optimize/test_hyperopt.py

@@ -358,7 +358,12 @@ def test_generate_optimizer(mocker, default_conf) -> None:
         'loss': 1.9840569076926293,
         'result': '     1 trades. Avg profit  2.31%. Total profit  0.00023300 BTC '
                   '(0.0231Σ%). Avg duration 100.0 mins.',
-        'params': optimizer_param
+        'params': optimizer_param,
+        'asked': [
+            0, 35, 0, 0, False, True, False, False, 'macd_cross_signal', 0, 75, 0, 0,
+            False, True, False, False, 'macd_cross_signal', 60.0, 30.0, 20.0,
+            0.01, 0.01, 0.1, -0.4
+        ]
     }
 
     hyperopt = Hyperopt(default_conf)