New API rewrite to separate code from exprimental API. (#559)

Signed-off-by: Cheng, Penghui <[email protected]> Signed-off-by: Lv, Liang1 <[email protected]> Signed-off-by: Mengni Wang <[email protected]> Signed-off-by: yiliu30 <[email protected]> Signed-off-by: Xinyu Ye <[email protected]> Signed-off-by: Zhang, Weiwei1 <[email protected]> Signed-off-by: zehao-intel <[email protected]> Signed-off-by: wenhuach21 <[email protected]>
intel · Feb 28, 2023 · 6e10efd · 6e10efd
1 parent 73675ef
commit 6e10efd
Show file tree

Hide file tree

Showing 100 changed files with 4,714 additions and 937 deletions.
diff --git a/examples/pytorch/image_recognition/CNN-2/distillation/eager/main.py b/examples/pytorch/image_recognition/CNN-2/distillation/eager/main.py
@@ -261,7 +261,7 @@ def train(train_loader, model, criterion, optimizer, scheduler, compression_mana
                         loss=losses, top1=top1, scheduler=scheduler))
 
         compression_manager.callbacks.on_epoch_end()
-        best_score = validate(val_loader, model, epoch + 1)
+        best_score = validate(val_loader, model, epoch + 1, accelerator)
         # remember best prec@1 and save checkpoint
         is_best = best_score > best_prec1
         best_prec1 = max(best_score, best_prec1)

diff --git a/neural_compressor/__init__.py b/neural_compressor/__init__.py
@@ -25,4 +25,5 @@
 from .conf.config import conf
 from .conf.pythonic_config import config
 from .config import DistillationConfig, PostTrainingQuantConfig, \
-                    WeightPruningConfig, QuantizationAwareTrainingConfig
+                    WeightPruningConfig, QuantizationAwareTrainingConfig, \
+                    MixedPrecisionConfig
diff --git a/neural_compressor/adaptor/keras.py b/neural_compressor/adaptor/keras.py
@@ -28,7 +28,7 @@
 from ..utils.utility import Statistics, GLOBAL_STATE, MODE, version1_lt_version2
 from ..utils import logger
 from ..conf.dotdict import deep_get
-from ..experimental.data.dataloaders.base_dataloader import BaseDataLoader
+from ..data.dataloaders.base_dataloader import BaseDataLoader
 tf = LazyImport('tensorflow')
 
 def _add_supported_quantized_objects(custom_objects):

diff --git a/neural_compressor/adaptor/onnxrt.py b/neural_compressor/adaptor/onnxrt.py
@@ -61,7 +61,7 @@ def __init__(self, framework_specific_info):
         self.device = framework_specific_info["device"]
         self.static = framework_specific_info["approach"] == "post_training_static_quant"
         self.dynamic = framework_specific_info["approach"] == "post_training_dynamic_quant"
-        self.domain = framework_specific_info["domain"]
+        self.domain = framework_specific_info.get("domain", "auto")
         self.recipes = framework_specific_info["recipes"]
         self.backend = PROVIDERS[framework_specific_info["backend"]]
         self.performance_only = framework_specific_info.get("performance_only", False)

diff --git a/neural_compressor/adaptor/ox_utils/util.py b/neural_compressor/adaptor/ox_utils/util.py
@@ -180,10 +180,20 @@ def cast_tensor(tensor, dtype): # pragma: no cover
             new_val = float_to_bfloat16(val)
         else:
             raise ValueError('Expect fp16 or bf16 but get {}.'.format(dtype))
-        tensor.float_data[:] = []
-        tensor.int32_data[:] = []
-        tensor.raw_data = new_val.tostring()
-        tensor.data_type = dtype_mapping[dtype]
+        try:
+            new_tensor = helper.make_tensor(
+                    name=tensor.name,
+                    data_type=dtype_mapping[dtype],
+                    dims=numpy_helper.to_array(tensor).shape if \
+                        len(numpy_helper.to_array(tensor).shape) != 0 else [],
+                    vals=new_val if \
+                        len(numpy_helper.to_array(tensor)) != 0 else [numpy_helper.to_array(tensor)])
+            tensor.CopyFrom(new_tensor)
+        except:
+            tensor.float_data[:] = []
+            tensor.int32_data[:] = []
+            tensor.raw_data = new_val.tostring()
+            tensor.data_type = dtype_mapping[dtype]
         return True
     return False
 

diff --git a/neural_compressor/adaptor/pytorch.py b/neural_compressor/adaptor/pytorch.py
@@ -28,7 +28,7 @@
 from ..utils.utility import Statistics
 from ..utils import logger
 from .query import QueryBackendCapability
-from ..experimental.data.dataloaders.base_dataloader import BaseDataLoader
+from ..data.dataloaders.base_dataloader import BaseDataLoader
 from .torch_utils.smooth_quant import TorchSmoothQuant
 torch = LazyImport("torch")
 json = LazyImport("json")
@@ -1202,7 +1202,6 @@ def smooth_quant(self, model, dataloader, calib_iter, tune_cfg=None, alpha=0.5,
             model: A modified fp32 model
         """
         if not hasattr(self, 'sq') or force_re_smooth:
-            ##self.sq = TorchSmoothQuant(model._model, dataloader=dataloader)
             self.sq = TorchSmoothQuant(model._model, dataloader=dataloader)
         args = {}  ##different backends may have different default values
         if op_types != None:

diff --git a/neural_compressor/adaptor/tensorflow.py b/neural_compressor/adaptor/tensorflow.py
@@ -30,7 +30,7 @@
 from ..utils.utility import version1_lt_version2, version1_gte_version2, version1_eq_version2
 from ..utils import logger
 from ..conf.dotdict import deep_get
-from ..experimental.data.dataloaders.base_dataloader import BaseDataLoader
+from ..data.dataloaders.base_dataloader import BaseDataLoader
 
 tensorflow = LazyImport('tensorflow')
 spr_base_verions = ('2.11.0202242', '2.11.0202250')
@@ -1148,7 +1148,7 @@ def _inspect_tensor_inference(self, inspect_node_dict,  model, dataloader, itera
     def inspect_activation(self, node_list, graph_def, graph_node_name_mapping, quantization_cfg,
                            dataloader, iteration_list, graph_info):
         """Inspect the activation."""
-        from neural_compressor.experimental.common import Model
+        from neural_compressor.model import Model
         original_graph_node_mapping = {}
         for node in graph_def.node:
             original_graph_node_mapping[node.name] = node

diff --git a/neural_compressor/adaptor/tf_utils/graph_converter.py b/neural_compressor/adaptor/tf_utils/graph_converter.py
@@ -31,7 +31,7 @@
 from neural_compressor.utils.utility import combine_histogram
 from neural_compressor.utils.utility import CaptureOutputToFile
 from neural_compressor.conf.dotdict import deep_get
-from neural_compressor.experimental.common import Model
+from neural_compressor.model import Model
 from .transform_graph.insert_logging import InsertLogging
 from .transform_graph.rerange_quantized_concat import RerangeQuantizedConcat
 from .transform_graph.bias_correction import BiasCorrection

diff --git a/neural_compressor/adaptor/tf_utils/graph_converter_without_calib.py b/neural_compressor/adaptor/tf_utils/graph_converter_without_calib.py
@@ -24,7 +24,7 @@
 
 from tensorflow.python.platform import gfile
 from neural_compressor.conf.dotdict import deep_get
-from neural_compressor.experimental.common import Model
+from neural_compressor.model import Model
 from .transform_graph.rerange_quantized_concat import RerangeQuantizedConcat
 from .transform_graph.bias_correction import BiasCorrection
 from .quantize_graph.quantize_graph_for_intel_cpu import QuantizeGraphForIntel

diff --git a/neural_compressor/adaptor/tf_utils/graph_rewriter/generic/pre_optimize.py b/neural_compressor/adaptor/tf_utils/graph_rewriter/generic/pre_optimize.py
@@ -104,7 +104,7 @@ def get_optimized_model(self, itex_mode=False):
         Returns:
             [graphdef]: the optimized graphdef object.
         """
-        from neural_compressor.experimental.common import Model
+        from neural_compressor.model import Model
 
         origin_model = Model(self.model._model, **self.model.kwargs)
         origin_model.name = self.model.name

diff --git a/neural_compressor/adaptor/tf_utils/util.py b/neural_compressor/adaptor/tf_utils/util.py
@@ -431,7 +431,7 @@ def is_equivalent_input(input_tensor_list_1, input_tensor_list_2):
 # THIS API IS TO BE DEPRECATED!
 def get_graph_def(model, outputs=[], auto_input_output=False):
     """Get the model's graph_def."""
-    from neural_compressor.experimental.common import Model as NCModel
+    from neural_compressor.model import Model as NCModel
     if not isinstance(model, NCModel):
         model = NCModel(model)
         model.output_tensor_names = outputs