Module furiosa.quantizer.frontend.onnx.transformer.fuse_depth_to_space

Classes

class FuseDepthToSpace (*args, **kwds)

Abstract base class for generic types.

A generic type is typically declared by inheriting from this class parameterized with one or more type variables. For example, a generic mapping type might be defined as::

class Mapping(Generic[KT, VT]): def getitem(self, key: KT) -> VT: … # Etc.

This class can then be used as follows::

def lookup_name(mapping: Mapping[KT, VT], key: KT, default: VT) -> VT: try: return mapping[key] except KeyError: return default

Expand source code

class FuseDepthToSpace(Transformer):
    def transform(self, model: onnx.ModelProto) -> onnx.ModelProto:
        for transformer in [
            Pattern_1,
        ]:
            model = transformer(model).transform()

        return model

Ancestors

• furiosa_sdk_quantizer.interfaces.transformer.Transformer
• typing.Generic

Methods

def transform(self, model: onnx.onnx_ml_pb2.ModelProto) ‑> onnx.onnx_ml_pb2.ModelProto

Expand source code

def transform(self, model: onnx.ModelProto) -> onnx.ModelProto:
    for transformer in [
        Pattern_1,
    ]:
        model = transformer(model).transform()

    return model

class Pattern_1 (model)

transform prev –> Reshape –> Transpose –> Reshape –> next to prev –> DepthToSpace –> next

if Transpose.perm == [0, 1, 4, 2, 5, 3] or == [0, 3, 4, 1, 5, 2]

Expand source code

class Pattern_1(ONNXTransformer, abc.ABC):
    """
        transform
            prev --> Reshape --> Transpose --> Reshape --> next
        to
            prev --> DepthToSpace --> next

        if Transpose.perm == [0, 1, 4, 2, 5, 3] or == [0, 3, 4, 1, 5, 2]
    """

    def pattern_matching(self, base_node):
        inputs = base_node.input

        pattern_to_match = ['Reshape', 'Transpose', 'Reshape']
        matched_nodes = self.pattern_matcher(base_node, pattern_to_match)
        if not matched_nodes:
            return inputs

        if not self.pattern_condition_checker(matched_nodes):
            return inputs

        top_node, mid_node, _ = matched_nodes
        self.transform_to_fuse(matched_nodes,
                               nodes_to_add=[
                                   self.make_node('DepthToSpace', [top_node.input[0]], [base_node.output[0]],
                                                  top_node.name,
                                                  **self.get_attrs(top_node, mid_node))
                               ])

        return top_node.input

    def pattern_condition_checker(self, nodes_to_check):
        _, mid_node, _ = nodes_to_check
        perm = mid_node.attribute[0].ints
        if perm == [0, 1, 4, 2, 5, 3] or perm == [0, 3, 4, 1, 5, 2]:
            return True
        return False

    def get_attrs(self, top_node, mid_node):
        permutation = mid_node.attribute[0].ints
        reshaped_shape = self.get_value_info_shape(top_node.output[0])
        if all(x == y for (x, y) in zip(permutation, [0, 1, 4, 2, 5, 3])):
            mode = 'CRD'
            blocksize = reshaped_shape[2]
        elif all(x == y for (x, y) in zip(permutation, [0, 3, 4, 1, 5, 2])):
            mode = 'DCR'
            blocksize = reshaped_shape[1]
        else:
            raise Exception()

        return {'blocksize': blocksize, 'mode': mode}

Ancestors

• furiosa_sdk_quantizer.frontend.onnx.transformer.ONNXTransformer
• abc.ABC

Methods

def get_attrs(self, top_node, mid_node)

Expand source code

def get_attrs(self, top_node, mid_node):
    permutation = mid_node.attribute[0].ints
    reshaped_shape = self.get_value_info_shape(top_node.output[0])
    if all(x == y for (x, y) in zip(permutation, [0, 1, 4, 2, 5, 3])):
        mode = 'CRD'
        blocksize = reshaped_shape[2]
    elif all(x == y for (x, y) in zip(permutation, [0, 3, 4, 1, 5, 2])):
        mode = 'DCR'
        blocksize = reshaped_shape[1]
    else:
        raise Exception()

    return {'blocksize': blocksize, 'mode': mode}

def pattern_condition_checker(self, nodes_to_check)

Expand source code

def pattern_condition_checker(self, nodes_to_check):
    _, mid_node, _ = nodes_to_check
    perm = mid_node.attribute[0].ints
    if perm == [0, 1, 4, 2, 5, 3] or perm == [0, 3, 4, 1, 5, 2]:
        return True
    return False

def pattern_matching(self, base_node)

Expand source code

def pattern_matching(self, base_node):
    inputs = base_node.input

    pattern_to_match = ['Reshape', 'Transpose', 'Reshape']
    matched_nodes = self.pattern_matcher(base_node, pattern_to_match)
    if not matched_nodes:
        return inputs

    if not self.pattern_condition_checker(matched_nodes):
        return inputs

    top_node, mid_node, _ = matched_nodes
    self.transform_to_fuse(matched_nodes,
                           nodes_to_add=[
                               self.make_node('DepthToSpace', [top_node.input[0]], [base_node.output[0]],
                                              top_node.name,
                                              **self.get_attrs(top_node, mid_node))
                           ])

    return top_node.input