minor

Author: andrey-churkin

.ci/cspell_dict.txt

@@ -223,9 +223,9 @@ logicalor
 logicalxor
 logit
 loglikelihoods
+lspec
 lstmsequence
 lstsq
-lspec
 lyalyushkin
 mapillary
 maskrcnn

nncf/onnx/graph/model_transformer.py

@@ -47,8 +47,9 @@ class ONNXModelTransformer(ModelTransformer):
     ZERO_POINT_NAME_PREFIX = "zero_point_"
 
     def __init__(self, model: onnx.ModelProto):
-        super().__init__(model)
-        self.onnx_model_extractor = onnx.utils.Extractor(model)
+        inferred_model = onnx.shape_inference.infer_shapes(model)
+        super().__init__(inferred_model)
+        self.onnx_model_extractor = onnx.utils.Extractor(inferred_model)
 
     def _get_target_edge(
         self,

nncf/onnx/graph/passes.py

@@ -61,7 +61,7 @@ def eliminate_nop_cast(model: onnx.ModelProto) -> onnx.ModelProto:
     return model
 
 
-def apply_preprocess_passes(model: onnx.ModelProto) -> None:
+def apply_preprocess_passes(model: onnx.ModelProto) -> onnx.ModelProto:
     """
     Preprocesses the provided ONNX model for quantization.
 
@@ -73,5 +73,7 @@ def apply_preprocess_passes(model: onnx.ModelProto) -> None:
     :return: A preprocessed ONNX model, ready for quantization.
     """
     preprocessed_model = onnx.shape_inference.infer_shapes(model)
+    # The `eliminate_nop_cast` pass should be applied after onnx.shape_inference.infer_shapes() call.
+    # Otherwise, not all no-op Cast nodes will be found.
     preprocessed_model = eliminate_nop_cast(preprocessed_model)
     return preprocessed_model

nncf/onnx/quantization/quantize_model.py

@@ -12,7 +12,6 @@
 import sys
 from pathlib import Path
 from typing import Any, Callable, Iterable, Optional, TypeVar, Union
-import tempfile
 
 import onnx
 from onnx.external_data_helper import ExternalDataInfo
@@ -29,8 +28,8 @@
 from nncf.onnx.graph.model_metadata import MetadataKey
 from nncf.onnx.graph.model_metadata import remove_metadata
 from nncf.onnx.graph.model_metadata import set_metadata
-from nncf.onnx.graph.model_utils import eliminate_nop_cast
 from nncf.onnx.graph.nncf_graph_builder import GraphConverter
+from nncf.onnx.graph.passes import apply_preprocess_passes
 from nncf.onnx.quantization.backend_parameters import get_external_data_dir
 from nncf.parameters import BackupMode
 from nncf.parameters import CompressionFormat
@@ -157,6 +156,7 @@ def quantize_impl(
     external_data_dir = check_external_data_location(model, external_data_dir)
     if external_data_dir:
         set_metadata(model, MetadataKey.EXTERNAL_DATA_DIR, external_data_dir)
+    model = apply_preprocess_passes(model)
 
     quantization_algorithm = PostTrainingQuantization(
         preset=preset,