Merge pull request #1088 from openvinotoolkit/class-incr-learning-validation

Applying #1084 to develop branch

Author: Songki Choi

external/mmsegmentation/tests/test_ote_training.py

@@ -276,5 +276,8 @@ def test(self,
              test_parameters,
              test_case_fx, data_collector_fx,
              cur_test_expected_metrics_callback_fx):
+        if "18_OCR" in test_parameters["model_name"] \
+                or "x-mod3_OCR" in test_parameters["model_name"]:
+            pytest.skip("Known issue CVS-83781")
         test_case_fx.run_stage(test_parameters['test_stage'], data_collector_fx,
                                cur_test_expected_metrics_callback_fx)

external/model-preparation-algorithm/mpa_tasks/apis/classification/task.py

@@ -15,6 +15,7 @@
 
 from ote_sdk.entities.datasets import DatasetEntity
 from ote_sdk.entities.inference_parameters import InferenceParameters, default_progress_callback
+from ote_sdk.entities.train_parameters import default_progress_callback as train_default_progress_callback
 from ote_sdk.entities.model import ModelEntity, ModelPrecision  # ModelStatus
 from ote_sdk.entities.resultset import ResultSetEntity
 from mmcv.utils import ConfigDict
@@ -217,6 +218,7 @@ def cancel_training(self):
         The stopping mechanism allows stopping after each iteration, but validation will still be carried out. Stopping
         will therefore take some time.
         """
+        self._should_stop = True
         logger.info("Cancel training requested.")
         if self.cancel_interface is not None:
             self.cancel_interface.cancel()
@@ -229,17 +231,34 @@ def train(self,
               output_model: ModelEntity,
               train_parameters: Optional[TrainParameters] = None):
         logger.info('train()')
+        # Check for stop signal between pre-eval and training. 
+        # If training is cancelled at this point,
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
+
         # Set OTE LoggerHook & Time Monitor
-        update_progress_callback = default_progress_callback
+        update_progress_callback = train_default_progress_callback
         if train_parameters is not None:
             update_progress_callback = train_parameters.update_progress
         self._time_monitor = TrainingProgressCallback(update_progress_callback)
         self._learning_curves = defaultdict(OTELoggerHook.Curve)
 
         stage_module = 'ClsTrainer'
         self._data_cfg = self._init_train_data_cfg(dataset)
+        self._is_training = True
         results = self._run_task(stage_module, mode='train', dataset=dataset, parameters=train_parameters)
 
+        # Check for stop signal between pre-eval and training. 
+        # If training is cancelled at this point,
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
+
         # get output model
         model_ckpt = results.get('final_ckpt')
         if model_ckpt is None:
@@ -257,6 +276,7 @@ def train(self,
                                   dashboard_metrics=training_metrics)
         logger.info(f'Final model performance: {str(performance)}')
         output_model.performance = performance
+        self._is_training = False
         logger.info('train done.')
 
     def _init_train_data_cfg(self, dataset: DatasetEntity):

external/model-preparation-algorithm/mpa_tasks/apis/detection/task.py

@@ -11,7 +11,7 @@
 import torch
 from mmcv.utils import ConfigDict
 from detection_tasks.apis.detection.config_utils import remove_from_config
-from detection_tasks.apis.detection.ote_utils import TrainingProgressCallback
+from detection_tasks.apis.detection.ote_utils import TrainingProgressCallback, InferenceProgressCallback
 from detection_tasks.extension.utils.hooks import OTELoggerHook
 from mpa_tasks.apis import BaseTask, TrainType
 from mpa_tasks.apis.detection import DetectionConfig
@@ -67,6 +67,11 @@ def infer(self,
               ) -> DatasetEntity:
         logger.info('infer()')
 
+        update_progress_callback = default_progress_callback
+        if inference_parameters is not None:
+            update_progress_callback = inference_parameters.update_progress
+
+        self._time_monitor = InferenceProgressCallback(len(dataset), update_progress_callback)
         # If confidence threshold is adaptive then up-to-date value should be stored in the model
         # and should not be changed during inference. Otherwise user-specified value should be taken.
         if not self._hyperparams.postprocessing.result_based_confidence_threshold:
@@ -75,7 +80,7 @@ def infer(self,
 
         stage_module = 'DetectionInferrer'
         self._data_cfg = self._init_test_data_cfg(dataset)
-        results = self._run_task(stage_module, mode='train', dataset=dataset)
+        results = self._run_task(stage_module, mode='train', dataset=dataset, parameters=inference_parameters)
         # TODO: InferenceProgressCallback register
         logger.debug(f'result of run_task {stage_module} module = {results}')
         output = results['outputs']
@@ -310,7 +315,7 @@ def cancel_training(self):
         will therefore take some time.
         """
         logger.info("Cancel training requested.")
-        # self._should_stop = True
+        self._should_stop = True
         # stop_training_filepath = os.path.join(self._training_work_dir, '.stop_training')
         # open(stop_training_filepath, 'a').close()
         if self.cancel_interface is not None:
@@ -324,6 +329,14 @@ def train(self,
               output_model: ModelEntity,
               train_parameters: Optional[TrainParameters] = None):
         logger.info('train()')
+        # Check for stop signal when training has stopped. 
+        # If should_stop is true, training was cancelled and no new
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
+
         # Set OTE LoggerHook & Time Monitor
         update_progress_callback = default_progress_callback
         if train_parameters is not None:
@@ -333,8 +346,15 @@ def train(self,
 
         stage_module = 'DetectionTrainer'
         self._data_cfg = self._init_train_data_cfg(dataset)
+        self._is_training = True
         results = self._run_task(stage_module, mode='train', dataset=dataset, parameters=train_parameters)
-        # logger.info(f'result of run_task {stage_module} module = {results}')
+
+        # Check for stop signal when training has stopped. If should_stop is true, training was cancelled and no new
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
 
         # get output model
         model_ckpt = results.get('final_ckpt')
@@ -389,6 +409,7 @@ def train(self,
         self.save_model(output_model)
         output_model.performance = performance
         # output_model.model_status = ModelStatus.SUCCESS
+        self._is_training = False
         logger.info('train done.')
 
     def _init_train_data_cfg(self, dataset: DatasetEntity):

external/model-preparation-algorithm/mpa_tasks/apis/segmentation/task.py

@@ -11,7 +11,7 @@
 import torch
 from mmcv.utils import ConfigDict
 from segmentation_tasks.apis.segmentation.config_utils import remove_from_config
-from segmentation_tasks.apis.segmentation.ote_utils import TrainingProgressCallback
+from segmentation_tasks.apis.segmentation.ote_utils import TrainingProgressCallback, InferenceProgressCallback
 from segmentation_tasks.extension.utils.hooks import OTELoggerHook
 from mpa import MPAConstants
 from mpa_tasks.apis import BaseTask, TrainType
@@ -22,6 +22,7 @@
 from ote_sdk.configuration.helper.utils import ids_to_strings
 from ote_sdk.entities.datasets import DatasetEntity
 from ote_sdk.entities.inference_parameters import InferenceParameters
+from ote_sdk.entities.inference_parameters import default_progress_callback as default_infer_progress_callback
 from ote_sdk.entities.label import Domain
 from ote_sdk.entities.metrics import (CurveMetric, InfoMetric, LineChartInfo,
                                       MetricsGroup, Performance, ScoreMetric,
@@ -48,8 +49,6 @@
     create_annotation_from_segmentation_map,
     create_hard_prediction_from_soft_prediction)
 
-# from mmdet.apis import export_model
-
 
 logger = get_logger()
 
@@ -70,12 +69,14 @@ def infer(self,
         logger.info('infer()')
 
         if inference_parameters is not None:
-            # update_progress_callback = inference_parameters.update_progress
+            update_progress_callback = inference_parameters.update_progress
             is_evaluation = inference_parameters.is_evaluation
         else:
-            # update_progress_callback = default_infer_progress_callback
+            update_progress_callback = default_infer_progress_callback
             is_evaluation = False
 
+        self._time_monitor = InferenceProgressCallback(len(dataset), update_progress_callback)
+
         stage_module = 'SegInferrer'
         self._data_cfg = self._init_test_data_cfg(dataset)
         self._label_dictionary = dict(enumerate(self._labels, 1))
@@ -187,8 +188,10 @@ def _init_test_data_cfg(self, dataset: DatasetEntity):
         data_cfg = ConfigDict(
             data=ConfigDict(
                 train=ConfigDict(
-                    ote_dataset=None,
-                    labels=self._labels,
+                    dataset=ConfigDict(
+                        ote_dataset=None,
+                        labels=self._labels,
+                    )
                 ),
                 test=ConfigDict(
                     ote_dataset=dataset,
@@ -311,7 +314,7 @@ def cancel_training(self):
         will therefore take some time.
         """
         logger.info("Cancel training requested.")
-        # self._should_stop = True
+        self._should_stop = True
         # stop_training_filepath = os.path.join(self._training_work_dir, '.stop_training')
         # open(stop_training_filepath, 'a').close()
         if self.cancel_interface is not None:
@@ -325,6 +328,14 @@ def train(self,
               output_model: ModelEntity,
               train_parameters: Optional[TrainParameters] = None):
         logger.info('train()')
+        # Check for stop signal between pre-eval and training. 
+        # If training is cancelled at this point,
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
+
         # Set OTE LoggerHook & Time Monitor
         if train_parameters is not None:
             update_progress_callback = train_parameters.update_progress
@@ -336,8 +347,17 @@ def train(self,
         # learning_curves = defaultdict(OTELoggerHook.Curve)
         stage_module = 'SegTrainer'
         self._data_cfg = self._init_train_data_cfg(dataset)
+        self._is_training = True
         results = self._run_task(stage_module, mode='train', dataset=dataset, parameters=train_parameters)
 
+        # Check for stop signal when training has stopped. 
+        # If should_stop is true, training was cancelled and no new
+        if self._should_stop:
+            logger.info('Training cancelled.')
+            self._should_stop = False
+            self._is_training = False
+            return
+
         # get output model
         model_ckpt = results.get('final_ckpt')
         if model_ckpt is None:
@@ -358,6 +378,7 @@ def train(self,
         self.save_model(output_model)
         output_model.performance = performance
         # output_model.model_status = ModelStatus.SUCCESS
+        self._is_training = False
         logger.info('train done.')
 
     def _init_train_data_cfg(self, dataset: DatasetEntity):

external/model-preparation-algorithm/mpa_tasks/apis/task.py

@@ -29,27 +29,6 @@
 logger = get_logger()
 
 
-class _MPAUpdateProgressCallbackWrapper(UpdateProgressCallback):
-    """ UpdateProgressCallback wrapper
-        just wrapping the callback instance and provides error free representation as 'pretty_text'
-    """
-
-    def __init__(self, callback, **kwargs):
-        if not callable(callback):
-            raise RuntimeError(f'cannot accept a not callable object!! {callback}')
-        self._callback = callback
-        super().__init__(**kwargs)
-
-    def __repr__(self):
-        return f"'{__name__}._MPAUpdateProgressCallbackWrapper'"
-
-    def __reduce__(self):
-        return (self.__class__, (id(self),))
-
-    def __call__(self, progress: float, score: Optional[float] = None):
-        self._callback(progress, score)
-
-
 class BaseTask:
     def __init__(self, task_config, task_environment: TaskEnvironment):
         self._task_config = task_config
@@ -83,6 +62,8 @@ def __init__(self, task_config, task_environment: TaskEnvironment):
         self._mode = None
         self._time_monitor = None
         self._learning_curves = None
+        self._is_training = False
+        self._should_stop = False
         self.cancel_interface = None
         self.reserved_cancel = False
         self.on_hook_initialized = self.OnHookInitialized(self)
@@ -104,30 +85,9 @@ def _run_task(self, stage_module, mode=None, dataset=None, parameters=None, **kw
             raise RuntimeError(
                 "'recipe_cfg' is not initialized yet."
                 "call prepare() method before calling this method")
-        # self._stage_module = stage_module
+
         if mode is not None:
             self._mode = mode
-        if parameters is not None:
-            if isinstance(parameters, TrainParameters):
-                hook_name = 'TrainProgressUpdateHook'
-                progress_callback = _MPAUpdateProgressCallbackWrapper(parameters.update_progress)
-                # TODO: update recipe to do RESUME
-                if parameters.resume:
-                    pass
-            elif isinstance(parameters, InferenceParameters):
-                hook_name = 'InferenceProgressUpdateHook'
-                progress_callback = _MPAUpdateProgressCallbackWrapper(parameters.update_progress)
-        else:
-            hook_name = 'ProgressUpdateHook'
-            progress_callback = None
-        logger.info(f'progress callback = {progress_callback}, hook name = {hook_name}')
-        if progress_callback is not None:
-            progress_update_hook_cfg = ConfigDict(
-                type='ProgressUpdateHook',
-                name=hook_name,
-                callback=progress_callback
-            )
-            update_or_add_custom_hook(self._recipe_cfg, progress_update_hook_cfg)
 
         common_cfg = ConfigDict(dict(output_path=self._output_path))
 
@@ -152,6 +112,14 @@ def finalize(self):
                 if os.path.exists(self._output_path):
                     shutil.rmtree(self._output_path, ignore_errors=False)
 
+    def _delete_scratch_space(self):
+        """
+        Remove model checkpoints and mpa logs
+        """
+
+        if os.path.exists(self._output_path):
+            shutil.rmtree(self._output_path, ignore_errors=False)
+
     def __del__(self):
         self.finalize()
 

external/model-preparation-algorithm/ote_tests_pytest.ini

@@ -1,2 +1,2 @@
 [pytest]
-python_files = test_*_cls_il.py
+python_files = test_ote_*.py

external/model-preparation-algorithm/submodule

@@ -0,0 +1,11 @@
+# Copyright (C) 2022 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+
+try:
+    import os
+    from e2e import config as config_e2e
+
+    config_e2e.repository_name = os.environ.get('TT_REPOSITORY_NAME', 'ote/training_extensions/external/model-preparation-algorithm')
+except ImportError:
+    pass