Enhance autotune to return the best q_model directly (#1875)

Signed-off-by: yiliu30 <[email protected]>

Author: yiliu30

neural_compressor/common/base_tuning.py

@@ -336,13 +336,17 @@ def set_baseline(self, baseline: float):
     def get_number_of_trials(self):
         return len(self.tuning_history)
 
-    def get_best_quant_config(self) -> BaseConfig:
+    def get_best_trial_record(self) -> _TrialRecord:
         assert self.get_number_of_trials() > 0, "No trial record in tuning monitor."
         # Put the record with a higher score at the beginning
         sorted_trials_records: List[_TrialRecord] = sorted(
             self.tuning_history, key=lambda x: x.trial_result, reverse=True
         )
-        return sorted_trials_records[0].quant_config
+        return sorted_trials_records[0]
+
+    def get_best_quant_config(self) -> BaseConfig:
+        best_trial_record = self.get_best_trial_record()
+        return best_trial_record.quant_config
 
     def need_stop(self) -> bool:
         """Check if need to stop tuning. Either accuracy goal is met, max trials is reached or timeout is reached.

neural_compressor/common/utils/utility.py

@@ -15,9 +15,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import collections
 import importlib
 import subprocess
 import time
+from typing import Dict
 
 import cpuinfo
 import psutil
@@ -35,6 +37,7 @@
     "LazyImport",
     "CpuInfo",
     "default_tuning_logger",
+    "call_counter",
 ]
 
 
@@ -225,3 +228,15 @@ def inner_wrapper(*args, **kwargs):
         return inner_wrapper
 
     return log_process_wrapper
+
+
+# decorator for recording number of times a function is called
+FUNC_CALL_COUNTS: Dict[str, int] = collections.defaultdict(int)
+
+
+def call_counter(func):
+    def wrapper(*args, **kwargs):
+        FUNC_CALL_COUNTS[func.__name__] += 1
+        return func(*args, **kwargs)
+
+    return wrapper

neural_compressor/tensorflow/quantization/autotune.py

@@ -20,7 +20,7 @@
 from neural_compressor.common import logger
 from neural_compressor.common.base_config import BaseConfig, get_all_config_set_from_config_registry
 from neural_compressor.common.base_tuning import EvaluationFuncWrapper, TuningConfig, init_tuning
-from neural_compressor.common.utils import dump_elapsed_time
+from neural_compressor.common.utils import call_counter, dump_elapsed_time
 from neural_compressor.tensorflow.quantization import quantize_model
 from neural_compressor.tensorflow.quantization.config import FRAMEWORK_NAME, StaticQuantConfig
 from neural_compressor.tensorflow.utils import BaseModel, Model, constants
@@ -36,6 +36,7 @@ def get_all_config_set() -> Union[BaseConfig, List[BaseConfig]]:
 
 
 @dump_elapsed_time("Pass auto-tune")
+@call_counter
 def autotune(
     model: Union[str, tf.keras.Model, BaseModel],
     tune_config: TuningConfig,
@@ -52,7 +53,7 @@ def autotune(
     baseline: float = eval_func_wrapper.evaluate(model)
     tuning_monitor.set_baseline(baseline)
     tuning_logger.tuning_start()
-    for trial_index, quant_config in enumerate(config_loader):
+    for trial_index, quant_config in enumerate(config_loader, 1):
         tuning_logger.trial_start(trial_index=trial_index)
         tuning_logger.execution_start()
         logger.info(quant_config.to_dict())
@@ -65,8 +66,14 @@ def autotune(
         tuning_logger.trial_end(trial_index)
         if tuning_monitor.need_stop():
             logger.info("Stopped tuning.")
-            best_quant_config: BaseConfig = tuning_monitor.get_best_quant_config()
-            best_quant_model = quantize_model(model, quant_config, calib_dataloader, calib_iteration)
+            best_trial_record = tuning_monitor.get_best_trial_record()
+            if best_trial_record.trial_index != trial_index:
+                logger.info("Re-quantizing with best quantization config...")
+                del q_model
+                best_quant_config: BaseConfig = best_trial_record.quant_config
+                best_quant_model = quantize_model(model, best_quant_config, calib_dataloader, calib_iteration)
+            else:
+                best_quant_model = q_model
             break
     tuning_logger.tuning_end()
     return best_quant_model

neural_compressor/torch/quantization/autotune.py

@@ -72,7 +72,7 @@ def autotune(
     baseline: float = eval_func_wrapper.evaluate(model)
     tuning_monitor.set_baseline(baseline)
     tuning_logger.tuning_start()
-    for trial_index, quant_config in enumerate(config_loader):
+    for trial_index, quant_config in enumerate(config_loader, 1):
         tuning_logger.trial_start(trial_index=trial_index)
         tuning_logger.execution_start()
         logger.info(quant_config.to_dict())
@@ -93,10 +93,11 @@ def autotune(
         tuning_logger.trial_end(trial_index)
         if tuning_monitor.need_stop():
             logger.info("Stopped tuning.")
-            if trial_index == 0:  # recover the best q_model from previous results.
-                logger.info("Reconvering the best quantized model...")
+            best_trial_record = tuning_monitor.get_best_trial_record()
+            if best_trial_record.trial_index != trial_index:
+                logger.info("Re-quantizing with best quantization config...")
                 del q_model  # maybe gc.collect() is needed for memory release
-                best_quant_config: BaseConfig = tuning_monitor.get_best_quant_config()
+                best_quant_config: BaseConfig = best_trial_record.quant_config
                 # !!! Make sure to use deepcopy only when inplace is set to `True`.
                 q_model = quantize(
                     deepcopy(model),

neural_compressor/torch/quantization/quantize.py

@@ -18,7 +18,7 @@
 import torch
 
 from neural_compressor.common.base_config import BaseConfig, ComposableConfig, config_registry
-from neural_compressor.common.utils import Mode, log_process
+from neural_compressor.common.utils import Mode, call_counter, log_process
 from neural_compressor.torch.quantization.config import SmoothQuantConfig, StaticQuantConfig
 from neural_compressor.torch.utils import is_ipex_available, logger
 from neural_compressor.torch.utils.utility import WHITE_MODULE_LIST, algos_mapping, get_model_info
@@ -31,6 +31,7 @@ def need_apply(configs_mapping: Dict[Tuple[str, callable], BaseConfig], algo_nam
 
 
 @log_process(mode=Mode.QUANTIZE)
+@call_counter
 def quantize(
     model: torch.nn.Module,
     quant_config: BaseConfig,

test/3x/common/test_utility.py

@@ -11,6 +11,7 @@
 import unittest
 from unittest.mock import MagicMock, patch
 
+import neural_compressor.common.utils.utility as inc_utils
 from neural_compressor.common import options
 from neural_compressor.common.utils import (
     CpuInfo,
@@ -166,5 +167,26 @@ def __init__(self):
         assert instance2.value == 1, "Singleton should return the same instance"
 
 
+class TestCallCounter(unittest.TestCase):
+    def test_call_counter(self):
+        # empty dict
+        inc_utils.FUNC_CALL_COUNTS.clear()
+
+        @inc_utils.call_counter
+        def add(a, b):
+            return a + b
+
+        # Initial count should be 0
+        self.assertEqual(inc_utils.FUNC_CALL_COUNTS["add"], 0)
+
+        # Call the function multiple times
+        add(1, 2)
+        add(3, 4)
+        add(5, 6)
+
+        # Count should be incremented accordingly
+        self.assertEqual(inc_utils.FUNC_CALL_COUNTS["add"], 3)
+
+
 if __name__ == "__main__":
     unittest.main()

test/3x/torch/test_autotune.py

@@ -6,6 +6,7 @@
 import torch
 import transformers
 
+import neural_compressor.common.utils.utility as inc_utils
 from neural_compressor.common import logger
 from neural_compressor.torch.quantization import (
     MixPrecisionConfig,
@@ -163,6 +164,43 @@ def eval_acc_fn(model) -> float:
 
         custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6])], max_trials=2)
         best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fn=eval_acc_fn)
+        print(inc_utils.FUNC_CALL_COUNTS)
+        self.assertIsNotNone(best_model)
+
+    def test_autotune_return_qmodel_directly(self):
+        inc_utils.FUNC_CALL_COUNTS.clear()
+
+        baseline = 1
+        eval_result = [0.9, 1.1]
+        acc_list = [baseline] + eval_result
+
+        def eval_acc_fn(model) -> float:
+            acc = acc_list.pop(0)
+            return acc
+
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6])], max_trials=2)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fn=eval_acc_fn)
+        assert (
+            inc_utils.FUNC_CALL_COUNTS.get("quantize") == 2
+        ), f"quantize should be called twice, but got {inc_utils.FUNC_CALL_COUNTS.get('quantize')}"
+        self.assertIsNotNone(best_model)
+
+    def test_autotune_return_re_quant_qmodel(self):
+        inc_utils.FUNC_CALL_COUNTS.clear()
+
+        baseline = 1
+        eval_result = [0.9, 0.8]
+        acc_list = [baseline] + eval_result
+
+        def eval_acc_fn(model) -> float:
+            acc = acc_list.pop(0)
+            return acc
+
+        custom_tune_config = TuningConfig(config_set=[RTNConfig(bits=[4, 6])], max_trials=2)
+        best_model = autotune(model=build_simple_torch_model(), tune_config=custom_tune_config, eval_fn=eval_acc_fn)
+        assert (
+            inc_utils.FUNC_CALL_COUNTS.get("quantize") == 3
+        ), f"quantize should be called three times, but got {inc_utils.FUNC_CALL_COUNTS.get('quantize')}"
         self.assertIsNotNone(best_model)
 
     @reset_tuning_target