quark.onnx.graph_transformations.transforms

Module Contents

Classes

class quark.onnx.graph_transformations.transforms.OpTypePattern(op_type: str = '', inputs: Optional[List[OpTypePattern]] = None, config: Optional[Dict[str, Any]] = None)

Defines a tree sub-graph pattern of onnx nodes to match in a model.

OpTypePattern can be used to describe various common patterns in model graphs that we need to find.

Examples:

Matches a Conv+BN+ReLU6 and DepthwiseConv+BN+ReLU6 pattern. pattern = OpTypePattern(‘ReLU’, {‘max_value’: 6.0}, [

OpTypePattern(‘BatchNormalization’, {}, [

OpTypePattern(‘Conv2D|DepthwiseConv2D’, {} [])

])

])

Matches multiple Conv2Ds feeding into a Concat. pattern = OpTypePattern(‘Concat’, {}, [

OpTypePattern(‘Conv2D’, {}, []), OpTypePattern(‘Conv2D’, {}, [])

])

class quark.onnx.graph_transformations.transforms.NodeTree(node: Optional[onnx.NodeProto] = None, weights: Union[OrderedDict[str, onnx.TensorProto], List[Any], None] = None, input_nodes: Optional[List[NodeTree]] = None, metadata: Optional[Dict[str, Any]] = None)

Represents a pattern matching results in a node containing a tree.

NodeTree is used to represent a tree of nodes in a model. It contains the NodeDef which describes the node, and other input nodes feeding into it.

It is used as a generic class to represent both sets of nodes which have been found in a model, and nodes which should be replaced inside the model.

class quark.onnx.graph_transformations.transforms.Transform

Defines a transform to be applied to a onnx model graph.

A transform is a combination of ‘Find + Replace’ which describes how to find a pattern of nodes in a model, and what to replace those nodes with.

A pattern is described using OpTypePattern. The replacement function receives a NodeTree which contains the matched nodes and should return a NodeTree which contains the set of nodes which replaced the matched nodes.

property allow_multi_consumers: bool

Whether to allow the internal node have multiple consuming nodes.

E.g.

B B

// //

A – to E –

C –> D F

Should set allow_mulit_consumers if you want to match pattern “A –> C –> D”. Please be careful to handle the transformation to not break the input connection of consumers outside the pattern, otherwise will lead to unknown input tensors.

abstract pattern() → OpTypePattern

Return the OpTypePattern to find in the model graph.

abstract replacement(match_node: NodeTree) → Any

Generate a replacement sub-graph for the matched sub-graph.

The fundamental constraint of the replacement is that the replacement sub-graph should consume the same input tensors as the original sub-graph and also produce a final list of tensors which are same in number and shape as the original sub-graph. Not following this could crash model creation, or introduce bugs in the new model graph.

Args:

match_nodes: Matched NodeTree based on self.pattern().