Remove 1.x API (1/N) (#1323)

* Update all experimental APIs caller
---------
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Signed-off-by: bmyrcha <bartosz.myrcha@intel.com>

neural_compressor/adaptor/tensorflow.py

@@ -190,7 +190,7 @@ def train(self, model, dataloader, optimizer_tuple, criterion_tuple, hooks, post
         callbacks = kwargs["kwargs"].get("callbacks", None)
         execution_mode = kwargs["kwargs"].get("execution_mode", None)
         distributed = getattr(dataloader, "distributed", False)
-        from neural_compressor.experimental.common.criterion import TensorflowKnowledgeDistillationLoss
+        from neural_compressor.compression.distillation.criterions import TensorflowKnowledgeDistillationLoss
 
         if isinstance(criterion, TensorflowKnowledgeDistillationLoss):
             input_model = model._model
@@ -1757,8 +1757,8 @@ def _get_mse_order(
 
     def _partial_dataset_of(self, dataloader, confidence_batches):
         """Partial dataset."""
+        from neural_compressor.data.datasets.dummy_dataset import DummyDataset
         from neural_compressor.data.datasets.dummy_dataset import DummyDataset as DummyDataset_v2_x
-        from neural_compressor.experimental.data.datasets.dummy_dataset import DummyDataset
 
         if isinstance(dataloader.dataset, DummyDataset) or isinstance(dataloader.dataset, DummyDataset_v2_x):
             assert isinstance(confidence_batches, int)

neural_compressor/compression/distillation/criterions.py

@@ -17,6 +17,8 @@
 """Initialize critetion classes.
 
 Classes includes:
+    TensorFlowCrossEntropyLoss, PyTorchCrossEntropyLoss,
+    TensorFlowSparseCategoricalCrossentropy,
     TensorflowKnowledgeDistillationLoss, PyTorchKnowledgeDistillationLoss,
     PyTorchIntermediateLayersKnowledgeDistillationLoss.
 """
@@ -91,7 +93,12 @@ def __getitem__(self, criterion_type):
         Returns:
             cls: criterion class.
         """
-        assert criterion_type in self.criterions.keys(), "only support criterions in {}".format(self.criterions.keys())
+        assert (
+            criterion_type in self.criterions.keys()
+        ), "only support criterions in {} \
+            , but got criterion type {}".format(
+            self.criterions.keys(), criterion_type
+        )
 
         return self.criterions[criterion_type]
 
@@ -130,6 +137,119 @@ def decorator_criterion(cls):
     return decorator_criterion
 
 
+@criterion_registry("CrossEntropyLoss", "tensorflow")
+class TensorFlowCrossEntropyLoss(object):
+    """TensorFlow CrossEntropyLoss criterion."""
+
+    def __init__(self, param_dict):
+        """Initialize the Datasets class.
+
+        Args:
+            param_dict (dict): The dict of parameters setting by user for CrossEntropyLoss criterion.
+        """
+        assert isinstance(param_dict, dict), "This criterion constructor parameter must be a dict"
+        self._param_dict = param_dict
+
+    def _mapping(self):
+        _param_map = {"reduction": "reduction", "from_logits": "from_logits"}
+        _dict = {}
+        for key in self._param_dict:
+            if key in _param_map:
+                if key == "reduction":
+                    assert self._param_dict[key] in [
+                        "auto",
+                        "none",
+                        "sum",
+                        "sum_over_batch_size",
+                    ], "Supported reduction value for tensorflow is auto, none, sum, sum_over_batch_size"
+                _dict.update({_param_map[key]: self._param_dict[key]})
+        return _dict
+
+    def __call__(self):
+        """Call the TensorFlowCrossEntropyLoss.
+
+        Returns:
+            cls: criterion class.
+            param_dict(dict): param_dict
+        """
+        return tf.keras.losses.CategoricalCrossentropy, self._mapping()
+
+
+@criterion_registry("SparseCategoricalCrossentropy", "tensorflow")
+class TensorFlowSparseCategoricalCrossentropy(object):
+    """TensorFlow SparseCategoricalCrossentropyLoss criterion."""
+
+    def __init__(self, param_dict):
+        """Initialize the Datasets class.
+
+        Args:
+            param_dict (string): param_dict.
+        """
+        assert isinstance(param_dict, dict), "This criterion constructor parameter must be a dict"
+        self._param_dict = param_dict
+
+    def _mapping(self):
+        _param_map = {"reduction": "reduction", "from_logits": "from_logits"}
+        _dict = {}
+        for key in self._param_dict:
+            if key in _param_map:
+                if key == "reduction":
+                    assert self._param_dict[key] in [
+                        "auto",
+                        "none",
+                        "sum",
+                        "sum_over_batch_size",
+                    ], "Supported reduction value for tensorflow is auto, none, sum, sum_over_batch_size"
+                _dict.update({_param_map[key]: self._param_dict[key]})
+        return _dict
+
+    def __call__(self):
+        """Call the TensorFlowSparseCategoricalCrossentropy.
+
+        Returns:
+            cls: criterion class.
+            param_dict(dict): param_dict
+        """
+        return tf.keras.losses.SparseCategoricalCrossentropy, self._mapping()
+
+
+@criterion_registry("CrossEntropyLoss", "pytorch")
+class PyTorchCrossEntropyLoss(object):
+    """PyTorch CrossEntropyLoss criterion."""
+
+    def __init__(self, param_dict):
+        """Initialize the PyTorchCrossEntropyLoss class.
+
+        Args:
+            param_dict (string): param_dict.
+        """
+        assert isinstance(param_dict, dict), "This criterion constructor parameter must be a dict"
+        self._param_dict = param_dict
+
+    def _mapping(self):
+        _param_map = {"reduction": "reduction"}
+        _dict = {}
+        for key in self._param_dict:
+            if key in _param_map:
+                if key == "reduction":
+                    assert self._param_dict[key] in [
+                        "none",
+                        "mean",
+                        "sum",
+                    ], "Supported reduction value is none, mean, sum"
+                _dict.update({_param_map[key]: self._param_dict[key]})
+        return _dict
+
+    def __call__(self):
+        """Call the PyTorchCrossEntropyLoss.
+
+        Returns:
+            cls: criterion class.
+            param_dict(dict): param_dict
+        """
+        return torch.nn.CrossEntropyLoss, self._mapping()
+
+
 class KnowledgeDistillationFramework(object):
     """Knowledge Distillation Framework."""
 
@@ -916,7 +1036,7 @@ def register_hooks_for_models(self):
         Raises:
             AttributeError: AttributeError
         """
-        from neural_compressor.experimental.common import torch_utils
+        from neural_compressor.compression.distillation import utility
 
         def register_model_forward_hook(model, path, output_process="", student=False):
             module = model
@@ -927,7 +1047,7 @@ def register_model_forward_hook(model, path, output_process="", student=False):
                         module = module.__getattr__(node)
                     except:
                         raise AttributeError("There is no path {} in the model.".format(path))
-            return module.register_forward_hook(torch_utils.get_activation(path, output_process, student))
+            return module.register_forward_hook(utility.get_activation(path, output_process, student))
 
         assert isinstance(self.student_model, torch.nn.Module) and isinstance(self.teacher_model, torch.nn.Module), (
             "Expect student_model and teacher_model to be an torch.nn.Module object, "
@@ -939,8 +1059,8 @@ def register_model_forward_hook(model, path, output_process="", student=False):
             student_output_process, teacher_output_process = self.layer_output_process[idx]
             st_handle = register_model_forward_hook(self.student_model, student_layer, student_output_process, True)
             te_handle = register_model_forward_hook(self.teacher_model, teacher_layer, teacher_output_process)
-            torch_utils.STUDENT_FEATURES = self.student_features
-            torch_utils.TEACHER_FEATURES = self.teacher_features
+            utility.STUDENT_FEATURES = self.student_features
+            utility.TEACHER_FEATURES = self.teacher_features
             self.hook_handles.extend([st_handle, te_handle])
 
     def remove_all_hooks(self):