From 6062e5085b40225a52d8813c979b6d97d43cf34c Mon Sep 17 00:00:00 2001
From: Floris-Jan Willemsen <fjwillemsen@icloud.com>
Date: Tue, 22 Oct 2024 00:29:40 -0700
Subject: [PATCH] Avoid duplicate execution, resolved an issue which sometimes
 led to hyperparamtuner tuning returning more results than the searchspace
 size

---
 kernel_tuner/backends/hypertuner.py |  1 +
 kernel_tuner/hyper.py               | 29 ++++++++++++++++++++++++++---
 kernel_tuner/interface.py           |  2 +-
 test/test_hyper.py                  |  4 ++--
 4 files changed, 30 insertions(+), 6 deletions(-)

diff --git a/kernel_tuner/backends/hypertuner.py b/kernel_tuner/backends/hypertuner.py
index ab89c39b..b5cf2b6a 100644
--- a/kernel_tuner/backends/hypertuner.py
+++ b/kernel_tuner/backends/hypertuner.py
@@ -38,6 +38,7 @@ def __init__(self, iterations):
         self.observers = [ScoreObserver(self)]
         self.name = platform.processor()
         self.max_threads = 1024
+        self.last_score = None
 
         # set the environment options
         env = dict()
diff --git a/kernel_tuner/hyper.py b/kernel_tuner/hyper.py
index 289fcd27..0aacca6c 100644
--- a/kernel_tuner/hyper.py
+++ b/kernel_tuner/hyper.py
@@ -1,10 +1,21 @@
 """Module for functions related to hyperparameter optimization."""
 
 
+from pathlib import Path
+from random import randint
 
 import kernel_tuner
 
 
+def get_random_unique_filename(prefix = '', suffix=''):
+    """Get a random, unique filename that does not yet exist."""
+    def randpath():
+        return Path(f"{prefix}{randint(1000, 9999)}{suffix}")
+    path = randpath()
+    while path.exists():
+        path = randpath()
+    return path
+
 def tune_hyper_params(target_strategy: str, hyper_params: dict, *args, **kwargs):
     """Tune hyperparameters for a given strategy and kernel.
 
@@ -46,8 +57,10 @@ def tune_hyper_params(target_strategy: str, hyper_params: dict, *args, **kwargs)
     if "iterations" in kwargs:
         iterations = kwargs['iterations']
         del kwargs['iterations']
-    if "cache" in kwargs:
-        del kwargs['cache']
+
+    # pass a temporary cache file to avoid duplicate execution
+    cachefile = get_random_unique_filename('temp_', '.json')
+    kwargs['cache'] = str(cachefile)
 
     def put_if_not_present(target_dict, key, value):
         target_dict[key] = value if key not in target_dict else target_dict[key]
@@ -59,8 +72,18 @@ def put_if_not_present(target_dict, key, value):
     kwargs['verify'] = None
     arguments = [target_strategy]
 
-    return kernel_tuner.tune_kernel('hyperparamtuning', None, [], arguments, hyper_params, *args, lang='Hypertuner',
+    # execute the hyperparameter tuning
+    result, env = kernel_tuner.tune_kernel('hyperparamtuning', None, [], arguments, hyper_params, *args, lang='Hypertuner',
                                     objective='score', objective_higher_is_better=True, iterations=iterations, **kwargs)
+    
+    # remove the temporary cachefile and return only unique results in order
+    cachefile.unlink()
+    result_unique = dict()
+    for r in result:
+        config_id = ",".join(str(r[k]) for k in hyper_params.keys())
+        if config_id not in result_unique:
+            result_unique[config_id] = r
+    return list(result_unique.values()), env
 
 if __name__ == "__main__":  # TODO remove in production
     # hyperparams = {
diff --git a/kernel_tuner/interface.py b/kernel_tuner/interface.py
index 3fcb769a..7570c321 100644
--- a/kernel_tuner/interface.py
+++ b/kernel_tuner/interface.py
@@ -404,7 +404,7 @@ def __deepcopy__(self, _):
             All strategies support the following two options:
 
             1. "max_fevals": the maximum number of unique valid function evaluations (i.e. compiling and
-            benchmarking a kernel configuration the strategy is allowed to perform as part of the optimization.
+            benchmarking a kernel configuration) the strategy is allowed to perform as part of the optimization.
             Note that some strategies implement a default max_fevals of 100.
 
             2. "time_limit": the maximum amount of time in seconds the strategy is allowed to spent on trying to
diff --git a/test/test_hyper.py b/test/test_hyper.py
index 1b7ccc55..7251cb3e 100644
--- a/test/test_hyper.py
+++ b/test/test_hyper.py
@@ -15,6 +15,6 @@ def test_hyper(env):
 
     target_strategy = "genetic_algorithm"
 
-    result, env = tune_hyper_params(target_strategy, hyper_params, iterations=1, verbose=True, cache=cache_filename)
-    assert len(result) >= 2     # Look into why the hyperparamtuner returns more results than the searchspace size
+    result, env = tune_hyper_params(target_strategy, hyper_params, iterations=1, verbose=True)
+    assert len(result) == 2
     assert 'best_config' in env