ONNX Export for Refactored Modifiers (#1769)

* ONNX export for modifier refactor

* export script runs for new modifier framework

* add cli alias

* style

* reload about QAT adjustment

* style

* rename session alias

integrations/torchvision/modifiers_refactor_example/e2e_test.py

@@ -24,7 +24,7 @@ def main():
     from torch.utils.data import DataLoader
     from torchvision import transforms
 
-    import sparseml.core.session as sml
+    import sparseml.core.session as session_manager
     from sparseml.core.event import EventType
     from sparseml.core.framework import Framework
     from sparseml.pytorch.utils import (
@@ -40,8 +40,8 @@ def main():
     device = "cuda:0"
 
     # set up SparseML session
-    sml.create_session()
-    session = sml.active_session()
+    session_manager.create_session()
+    session = session_manager.active_session()
 
     # download model
     model = torchvision.models.mobilenet_v2(

setup.py

@@ -187,8 +187,11 @@ def _setup_entry_points() -> Dict:
             ]
         )
 
-    entry_points["console_scripts"].append(
-        "sparseml.transformers.export_onnx=sparseml.transformers.export:main"
+    entry_points["console_scripts"].extend(
+        [
+            "sparseml.transformers.export_onnx=sparseml.transformers.export:main",
+            "sparseml.transformers.export_onnx_refactor=sparseml.transformers.sparsification.obcq.export:main",  # noqa 501
+        ]
     )
 
     # image classification integration

src/sparseml/modifiers/quantization/base.py

@@ -14,7 +14,7 @@
 
 from typing import Any, Dict, List, Optional
 
-from sparseml.core import Event, Modifier, State
+from sparseml.core import Event, Modifier
 
 
 __all__ = ["QuantizationModifier"]
@@ -136,6 +136,3 @@ def check_should_disable_observer(self, event: Event) -> bool:
         if event.current_index >= disable_epoch:
             return True
         return False
-
-    def on_initialize_structure(self, state: State, **kwargs):
-        pass  # nothing needed for this modifier

src/sparseml/modifiers/quantization/pytorch.py

@@ -71,6 +71,11 @@ def __init__(self, **kwargs):
             self.scheme_overrides, self.scheme
         )
 
+    def on_initialize_structure(self, state: State, **kwargs):
+        module = state.model.model
+        self._enable_module_qat(module)
+        state.model.model.apply(torch.quantization.disable_observer)
+
     def on_initialize(self, state: State, **kwargs) -> bool:
         raise_if_torch_quantization_not_available()
         if self.end and self.end != -1:
@@ -84,6 +89,7 @@ def on_initialize(self, state: State, **kwargs) -> bool:
 
         if self.calculate_start() == -1:  # one-shot
             self._enable_module_qat(module)
+            self._calibrate_if_possible(module)
             self._disable_quantization_observer(module)
 
         return True
@@ -122,30 +128,31 @@ def _disable_quantization_observer(self, model: Module):
         self.quantization_observer_disabled_ = True
 
     def _enable_module_qat(self, module: Module):
-        # fuse conv-bn-relu blocks prior to quantization emulation
-        self._fuse(module)
-
-        # add quantization_schemes to target submodules
-        set_quantization_schemes(
-            module,
-            scheme=self.scheme,
-            scheme_overrides=self.scheme_overrides,
-            ignore=self.ignore,
-            strict=self.strict,
-        )
+        module.apply(torch.quantization.enable_observer)
 
-        # fix for freezing batchnorm statistics when not fusing BN with convs.
-        # pytorch only supports freezing batchnorm statistics for fused modules.
-        # this fix wraps BN modules adding with a new module class that supports
-        # methods related to freezing/unfreezing BN statistics.
-        configure_module_bn_wrappers(module)
+        if not self.qat_enabled_:
+            # fuse conv-bn-relu blocks prior to quantization emulation
+            self._fuse(module)
+
+            # add quantization_schemes to target submodules
+            set_quantization_schemes(
+                module,
+                scheme=self.scheme,
+                scheme_overrides=self.scheme_overrides,
+                ignore=self.ignore,
+                strict=self.strict,
+            )
 
-        # convert target qconfig layers to QAT modules with FakeQuantize
-        convert_module_qat_from_schemes(module)
+            # fix for freezing batchnorm statistics when not fusing BN with convs.
+            # pytorch only supports freezing batchnorm statistics for fused modules.
+            # this fix wraps BN modules adding with a new module class that supports
+            # methods related to freezing/unfreezing BN statistics.
+            configure_module_bn_wrappers(module)
 
-        self.qat_enabled_ = True
+            # convert target qconfig layers to QAT modules with FakeQuantize
+            convert_module_qat_from_schemes(module)
 
-        self._calibrate_if_possible(module)
+        self.qat_enabled_ = True
 
     def _fuse(self, module: Module):
         if self.model_fuse_fn_name in [None, "conv_bn_relus"]: