Executor - Save benchmark results to file (#86)

* Save benchmark results to json file.

superbench/executor/executor.py

@@ -3,6 +3,8 @@
 
 """SuperBench Executor."""
 
+import json
+import itertools
 from pathlib import Path
 
 from omegaconf import ListConfig
@@ -94,6 +96,9 @@ def __exec_benchmark(self, context, log_suffix):
         Args:
             context (BenchmarkContext): Benchmark context to launch.
             log_suffix (str): Log string suffix.
+
+        Return:
+            dict: Benchmark results.
         """
         try:
             benchmark = BenchmarkRegistry.launch_benchmark(context)
@@ -106,10 +111,39 @@ def __exec_benchmark(self, context, log_suffix):
                     logger.info('Executor succeeded in %s.', log_suffix)
                 else:
                     logger.error('Executor failed in %s.', log_suffix)
+                return json.loads(benchmark.serialized_result)
             else:
                 logger.error('Executor failed in %s, invalid context.', log_suffix)
-        except Exception:
+        except Exception as e:
+            logger.error(e)
             logger.error('Executor failed in %s.', log_suffix)
+        return None
+
+    def __create_benchmark_dir(self, benchmark_name):
+        """Create output directory for benchmark.
+
+        Args:
+            benchmark_name (str): Benchmark name.
+        """
+        benchmark_output_dir = Path(self._output_dir, 'benchmarks', benchmark_name)
+        if benchmark_output_dir.is_dir() and any(benchmark_output_dir.iterdir()):
+            logger.warn('Benchmark output directory %s is not empty.', str(benchmark_output_dir))
+            for i in itertools.count(start=1):
+                backup_dir = benchmark_output_dir.with_name('{}.{}'.format(benchmark_name, i))
+                if not backup_dir.is_dir():
+                    benchmark_output_dir.rename(backup_dir)
+                    break
+        benchmark_output_dir.mkdir(mode=0o755, parents=True, exist_ok=True)
+
+    def __write_benchmark_results(self, benchmark_name, benchmark_results):
+        """Write benchmark results.
+
+        Args:
+            benchmark_name (str): Benchmark name.
+            benchmark_results (dict): Benchmark results.
+        """
+        with Path(self._output_dir, 'benchmarks', benchmark_name, 'results.json').open(mode='w') as f:
+            json.dump(benchmark_results, f, indent=2)
 
     def exec(self):
         """Run the SuperBench benchmarks locally."""
@@ -117,6 +151,8 @@ def exec(self):
             if benchmark_name not in self._sb_enabled:
                 continue
             benchmark_config = self._sb_benchmarks[benchmark_name]
+            benchmark_results = {}
+            self.__create_benchmark_dir(benchmark_name)
             for framework in benchmark_config.frameworks or [Framework.NONE]:
                 if benchmark_name.endswith('_models'):
                     for model in benchmark_config.models:
@@ -128,7 +164,11 @@ def exec(self):
                             framework=Framework(framework.lower()),
                             parameters=self.__get_arguments(benchmark_config.parameters)
                         )
-                        self.__exec_benchmark(context, log_suffix)
+                        result = self.__exec_benchmark(context, log_suffix)
+                        if framework != Framework.NONE:
+                            benchmark_results['{}/{}'.format(framework, model)] = result
+                        else:
+                            benchmark_results[model] = result
                 else:
                     log_suffix = 'micro-benchmark {}'.format(benchmark_name)
                     logger.info('Executor is going to execute %s.', log_suffix)
@@ -138,4 +178,9 @@ def exec(self):
                         framework=Framework(framework.lower()),
                         parameters=self.__get_arguments(benchmark_config.parameters)
                     )
-                    self.__exec_benchmark(context, log_suffix)
+                    result = self.__exec_benchmark(context, log_suffix)
+                    if framework != Framework.NONE:
+                        benchmark_results[framework] = result
+                    else:
+                        benchmark_results = result
+            self.__write_benchmark_results(benchmark_name, benchmark_results)

tests/executor/test_executor.py

@@ -3,10 +3,12 @@
 
 """SuperBench Executor test."""
 
+import json
 import unittest
 import shutil
 import tempfile
 from pathlib import Path
+from unittest import mock
 
 from omegaconf import OmegaConf
 
@@ -74,7 +76,59 @@ def test_get_arguments(self):
             ), expected_bert_models_args
         )
 
+    def test_create_benchmark_dir(self):
+        """Test __create_benchmark_dir."""
+        foo_path = Path(self.output_dir, 'benchmarks', 'foo')
+        self.executor._SuperBenchExecutor__create_benchmark_dir('foo')
+        self.assertTrue(foo_path.is_dir())
+        self.assertFalse(any(foo_path.iterdir()))
+
+        (foo_path / 'bar.txt').touch()
+        self.executor._SuperBenchExecutor__create_benchmark_dir('foo')
+        self.assertTrue(foo_path.is_dir())
+        self.assertFalse(any(foo_path.iterdir()))
+        self.assertFalse((foo_path / 'bar.txt').is_file())
+        self.assertTrue(foo_path.with_name('foo.1').is_dir())
+        self.assertTrue((foo_path.with_name('foo.1') / 'bar.txt').is_file())
+
+        (foo_path / 'bar.json').touch()
+        self.executor._SuperBenchExecutor__create_benchmark_dir('foo')
+        self.assertTrue(foo_path.is_dir())
+        self.assertFalse(any(foo_path.iterdir()))
+        self.assertFalse((foo_path / 'bar.json').is_file())
+        self.assertTrue(foo_path.with_name('foo.2').is_dir())
+        self.assertTrue((foo_path.with_name('foo.2') / 'bar.json').is_file())
+
+    def test_write_benchmark_results(self):
+        """Test __write_benchmark_results."""
+        foobar_path = Path(self.output_dir, 'benchmarks', 'foobar')
+        foobar_results_path = foobar_path / 'results.json'
+        self.executor._SuperBenchExecutor__create_benchmark_dir('foobar')
+        foobar_results = {
+            'sum': 1,
+            'avg': 1.1,
+        }
+        self.executor._SuperBenchExecutor__write_benchmark_results('foobar', foobar_results)
+        self.assertTrue(foobar_results_path.is_file())
+        with foobar_results_path.open(mode='r') as f:
+            self.assertDictEqual(json.load(f), foobar_results)
+
     def test_exec_empty_benchmarks(self):
         """Test execute empty benchmarks, nothing should happen."""
         self.executor._sb_enabled = []
         self.executor.exec()
+
+    @mock.patch('superbench.executor.SuperBenchExecutor._SuperBenchExecutor__exec_benchmark')
+    def test_exec_default_benchmarks(self, mock_exec_benchmark):
+        """Test execute default benchmarks, mock exec function.
+
+        Args:
+            mock_exec_benchmark (function): Mocked __exec_benchmark function.
+        """
+        mock_exec_benchmark.return_value = {}
+        self.executor.exec()
+
+        self.assertTrue(Path(self.output_dir, 'benchmarks').is_dir())
+        for benchmark_name in self.executor._sb_benchmarks:
+            self.assertTrue(Path(self.output_dir, 'benchmarks', benchmark_name).is_dir())
+            self.assertTrue(Path(self.output_dir, 'benchmarks', benchmark_name, 'results.json').is_file())