def optimize(model: onnx.ModelProto, skip_fuse_bn: bool,
skipped_optimizers: Optional[Sequence[str]]) -> onnx.ModelProto:
"""
:param model: The onnx model.
:return: The optimized onnx model.
Before simplifying, use this method to generate value_info, which is used in `forward_all`
After simplifying, use this method to fold constants generated in previous step into initializer,
and eliminate unused constants.
"""
onnx.checker.check_model(model)
onnx.helper.strip_doc_string(model)
optimizers_list = onnxoptimizer.get_fuse_and_elimination_passes()
if not skip_fuse_bn:
optimizers_list.append('fuse_bn_into_conv')
if skipped_optimizers is not None:
for opt in skipped_optimizers:
try:
optimizers_list.remove(opt)
except ValueError:
pass
model = onnxoptimizer.optimize(model, optimizers_list, fixed_point=True)
onnx.checker.check_model(model)
return model
def optimize(self, optimizations: List[str] = None, fixed_point: bool = False):
"""
Use ONNX optimizer to optimize the ONNX model. The optimizations supported can be seen by calling
'onnxoptimizer.get_available_passes()'
:param optimizations: List of possible optimizations. If None, all of the optimizations will be used. Defaulted
to None.
:param fixed_point: Optimize the weights using fixed point. Defaulted to False.
"""
# Set the ONNX optimizations list:
onnx_optimizations = onnxoptimizer.get_fuse_and_elimination_passes()
if optimizations is None:
# Set to all optimizations:
optimizations = onnx_optimizations
# Optimize the model:
self._model = onnxoptimizer.optimize(
self._model, passes=optimizations, fixed_point=fixed_point
)
def optimize(model: onnx.ModelProto, skip_fuse_bn: bool,
skipped_optimizers: Optional[Sequence[str]]) -> onnx.ModelProto:
"""
:model参数: 待优化的ONXX模型.
:return: 优化之后的ONNX模型.
简化之前, 使用这个方法产生会在'forward_all'用到的ValueInfo
简化之后，使用这个方法去折叠前一步产生的常量到initializer中并且消除没被使用的常量
"""
onnx.checker.check_model(model)
onnx.helper.strip_doc_string(model)
optimizers_list = onnxoptimizer.get_fuse_and_elimination_passes()
if not skip_fuse_bn:
optimizers_list.append('fuse_bn_into_conv')
if skipped_optimizers is not None:
for opt in skipped_optimizers:
try:
optimizers_list.remove(opt)
except ValueError:
pass
model = onnxoptimizer.optimize(model, optimizers_list, fixed_point=True)
onnx.checker.check_model(model)
return model