Add tolerable_loss to TuningConfig (#1579)

Signed-off-by: Kaihui-intel <[email protected]>

Author: Kaihui-intel

neural_compressor/common/base_tuning.py

@@ -213,14 +213,57 @@ class TuningConfig:
         config_set: quantization configs. Default value is empty.
         timeout: Tuning timeout (seconds). Default value is 0 which means early stop.
         max_trials: Max tuning times. Default value is 100. Combine with timeout field to decide when to exit.
+        tolerable_loss: This float indicates how much metric loss we can accept. \
+            The metric loss is relative, it can be both positive and negative. Default is 0.01.
+
+    Examples:
+        from neural_compressor import TuningConfig
+        tune_config = TuningConfig(
+            config_set=[config1, config2, ...],
+            max_trials=3,
+            tolerable_loss=0.01
+        )
+
+        # Case 1: Tolerable Loss
+        fp32_baseline = 100
+        config1_metric, config2_metric, ... = 98, 99, ...
+
+        # Tuning result of case 1:
+        # The best tuning config is config2, because config2_metric >= fp32_baseline * (1 - tolerable_loss)
+
+        # Case 2: Maximum Trials
+        fp32_baseline = 100
+        config1_metric, config2_metric, config3_metric, ... = 98, 98, 97, ...
+
+        # Tuning result of case 2:
+        # The best tuning config is config2, because of the following:
+        # 1. Not achieving the set goal. (config_metric < fp32_baseline * (1 - tolerable_loss))
+        # 2. Reached maximum tuning times.
+
+        # Case 3: Timeout
+        tune_config = TuningConfig(
+            config_set=[config1, config2, ...],
+            timeout=10, # seconds
+            max_trials=3,
+            tolerable_loss=0.01
+        )
+        config1_tuning_time, config2_tuning_time, config3_tuning_time, ... = 4, 5, 6, ... # seconds
+        fp32_baseline = 100
+        config1_metric, config2_metric, config3_metric, ... = 98, 98, 97, ...
+
+        # Tuning result of case 3:
+        # The best tuning config is config2, due to timeout, the third trial was forced to exit.
     """
 
-    def __init__(self, config_set=None, timeout=0, max_trials=100, sampler: Sampler = None) -> None:
+    def __init__(
+        self, config_set=None, timeout=0, max_trials=100, sampler: Sampler = None, tolerable_loss=0.01
+    ) -> None:
         """Init a TuneCriterion object."""
         self.config_set = config_set
         self.timeout = timeout
         self.max_trials = max_trials
         self.sampler = sampler
+        self.tolerable_loss = tolerable_loss
 
 
 class _TrialRecord:
@@ -242,12 +285,17 @@ def __init__(self, tuning_config: TuningConfig) -> None:
         self.tuning_config = tuning_config
         self.trial_cnt = 0
         self.tuning_history: List[_TrialRecord] = []
+        self.baseline = None
 
     def add_trial_result(self, trial_index: int, trial_result: Union[int, float], quant_config: BaseConfig) -> None:
         self.trial_cnt += 1
         trial_record = _TrialRecord(trial_index, trial_result, quant_config)
         self.tuning_history.append(trial_record)
 
+    def set_baseline(self, baseline: float):
+        self.baseline = baseline
+        logger.info(f"Fp32 baseline is {self.baseline}")
+
     def get_number_of_trials(self):
         return len(self.tuning_history)
 
@@ -260,8 +308,23 @@ def get_best_quant_config(self) -> BaseConfig:
         return sorted_trials_records[0].quant_config
 
     def need_stop(self) -> bool:
-        # TODO Support more stop criteria in the next PR, such as `reach accuracy goal`, `timeout`, and so on.
-        return self.trial_cnt >= self.tuning_config.max_trials
+        """Check if need to stop tuning. Either accuracy goal is met, max trials is reached or timeout is reached.
+
+        Returns:
+            bool: True if need to stop, otherwise False.
+        """
+
+        # TODO: Support more stop criteria in the next PR, such as `timeout`, and so on.
+        # reach max trials
+        reach_max_trials = self.trial_cnt >= self.tuning_config.max_trials
+        # reach accuracy goal
+        meet_accuracy_goal = (
+            False
+            if self.baseline is None
+            else self.tuning_history[-1].trial_result >= (self.baseline * (1 - self.tuning_config.tolerable_loss))
+        )
+        # [-1] is the last element representing the latest trail record.
+        return reach_max_trials or meet_accuracy_goal
 
 
 def init_tuning(tuning_config: TuningConfig) -> Tuple[ConfigLoader, TuningLogger, TuningMonitor]:

neural_compressor/torch/quantization/autotune.py

@@ -46,6 +46,8 @@ def autotune(
     evaluator.set_eval_fn_registry(eval_fns)
     evaluator.self_check()
     config_loader, tuning_logger, tuning_monitor = init_tuning(tuning_config=tune_config)
+    baseline: float = evaluator.evaluate(model)
+    tuning_monitor.set_baseline(baseline)
     tuning_logger.tuning_start()
     for trial_index, quant_config in enumerate(config_loader):
         tuning_logger.trial_start(trial_index=trial_index)
@@ -58,11 +60,12 @@ def autotune(
         eval_result: float = evaluator.evaluate(q_model)
         tuning_logger.evaluation_end()
         tuning_monitor.add_trial_result(trial_index, eval_result, quant_config)
+        tuning_logger.trial_end(trial_index)
         if tuning_monitor.need_stop():
             best_quant_config: BaseConfig = tuning_monitor.get_best_quant_config()
             # !!! Make sure to use deepcopy only when inplace is set to `True`.
             quantize(deepcopy(model), quant_config=best_quant_config, run_fn=run_fn, run_args=run_args, inplace=True)
             best_quant_model = model  # quantize model inplace
-        tuning_logger.trial_end(trial_index)
+            break
     tuning_logger.tuning_end()
     return best_quant_model

test/3x/common/test_common.py

@@ -42,7 +42,13 @@
 
 from typing import Any, Callable, List, Optional, Tuple, Union
 
-from neural_compressor.common.base_config import BaseConfig, get_all_config_set_from_config_registry, register_config
+from neural_compressor.common.base_config import (
+    BaseConfig,
+    ComposableConfig,
+    get_all_config_set_from_config_registry,
+    register_config,
+)
+from neural_compressor.common.base_tuning import ConfigLoader, Sampler
 from neural_compressor.common.utils import DEFAULT_WHITE_LIST, OP_NAME_OR_MODULE_TYPE
 
 PRIORITY_FAKE_ALGO = 100
@@ -137,5 +143,36 @@ def test_api(self):
         self.assertEqual(config_set[0].weight_bits, DEFAULT_WEIGHT_BITS)
 
 
+class TestConfigLoader(unittest.TestCase):
+    def setUp(self):
+        self.config_set = [get_default_fake_config(), get_default_fake_config()]
+        self.loader = ConfigLoader(self.config_set, Sampler())
+
+    def test_parse_quant_config_single(self):
+        quant_config = get_default_fake_config()
+        result = ConfigLoader.parse_quant_config(quant_config)
+        self.assertEqual(str(result), str(quant_config.expand()))
+
+    def test_parse_quant_config_composable(self):
+        quant_config = get_default_fake_config()
+        composable_config = ComposableConfig(get_default_fake_config())
+        composable_config.config_list = [quant_config]
+        result = ConfigLoader.parse_quant_config(composable_config)
+        self.assertEqual(str(result), str(quant_config.expand()))
+
+    def test_parse_quant_configs(self):
+        quant_configs = [get_default_fake_config(), get_default_fake_config()]
+        self.config_set[0].expand = lambda: quant_configs
+        self.config_set[1].expand = lambda: []
+        result = self.loader.parse_quant_configs()
+        self.assertEqual(result, quant_configs)
+
+    def test_iteration(self):
+        quant_configs = [get_default_fake_config(), get_default_fake_config()]
+        self.loader.parse_quant_configs = lambda: quant_configs
+        result = list(self.loader)
+        self.assertEqual(result, quant_configs)
+
+
 if __name__ == "__main__":
     unittest.main()

test/3x/common/test_utility.py

@@ -0,0 +1,58 @@
+"""Tests for common components.
+
+!!! Please do not import any framework-specific modules in this file. !!!
+* Note, we may need to add some auto check mechanisms to ensure this.
+
+These tests aim to assess the fundamental functionalities of common utils and enhance code coverage.
+All tests will be included for each framework CI.
+"""
+import unittest
+
+from neural_compressor.common import options
+from neural_compressor.common.utils import set_random_seed, set_resume_from, set_tensorboard, set_workspace
+
+
+class TestOptions(unittest.TestCase):
+    def test_set_random_seed(self):
+        seed = 12345
+        set_random_seed(seed)
+        self.assertEqual(options.random_seed, seed)
+
+        # non int type
+        seed = "12345"
+        with self.assertRaises(AssertionError):
+            set_random_seed(seed)
+
+    def test_set_workspace(self):
+        workspace = "/path/to/workspace"
+        set_workspace(workspace)
+        self.assertEqual(options.workspace, workspace)
+
+        # non String type
+        workspace = 12345
+        with self.assertRaises(AssertionError):
+            set_workspace(workspace)
+
+    def test_set_resume_from(self):
+        resume_from = "/path/to/resume"
+        set_resume_from(resume_from)
+        self.assertEqual(options.resume_from, resume_from)
+
+        # non String type
+        resume_from = 12345
+        with self.assertRaises(AssertionError):
+            set_resume_from(resume_from)
+
+    def test_set_tensorboard(self):
+        tensorboard = True
+        set_tensorboard(tensorboard)
+        self.assertEqual(options.tensorboard, tensorboard)
+
+        # non bool type
+        tensorboard = 123
+        with self.assertRaises(AssertionError):
+            set_tensorboard(tensorboard)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/3x/torch/test_autotune.py

@@ -200,6 +200,45 @@ def test_autotune_not_eval_func(self):
             str(context.exception), "Please ensure that you register at least one evaluation metric for auto-tune."
         )
 
+    def test_autotune_baseline(self):
+        logger.info("test_autotune_api")
+        from neural_compressor.common.base_tuning import evaluator
+
+        baseline = [1.0]
+
+        # case 1
+        # Where default tolerable_loss is 0.01, we expect the tuning to end with a "2-trail end" output logged.
+        acc_res_lst = baseline + [0.9] * 2 + [0.99]
+
+        def eval_acc_fn(model):
+            res = acc_res_lst.pop(0)
+            return res
+
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6, 5, 8])], max_trials=6)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fns=eval_acc_fn)
+        self.assertIsNotNone(best_model)
+
+        # case 2
+        # Where tolerable_loss is 0.1, we expect the tuning to end with a "0-trail end" output logged.
+        acc_res_lst = baseline + [0.9] * 2 + [0.99] + [1.01]
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6, 5, 8])], tolerable_loss=0.1)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fns=eval_acc_fn)
+        self.assertIsNotNone(best_model)
+
+        # case 3
+        # Where tolerable_loss is -0.01, we expect the tuning to end with a "3-trail end" output logged.
+        acc_res_lst = baseline + [0.9] * 2 + [0.99] + [1.01]
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6, 5, 8])], tolerable_loss=-0.01)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fns=eval_acc_fn)
+        self.assertIsNotNone(best_model)
+
+        # case 4
+        # Where tolerable_loss is 0.01 and accuracy meets the goal, we expect best model is None.
+        acc_res_lst = baseline + [0.9] * 2 + [0.9] + [0.9]
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6, 5, 8])], tolerable_loss=0.01)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fns=eval_acc_fn)
+        self.assertIsNone(best_model)
+
 
 if __name__ == "__main__":
     unittest.main()