Module furiosa.quantizer.frontend.onnx.transformer.convert_clip_attr_to_input

Classes

class ConvertClipAttrToInput (*args, **kwds)

https://github.com/onnx/onnx/blob/master/docs/Operators.md#softmax from: [max, min] in node.attribute to: [min, max] in node.input

Assume NCHW Input

Expand source code

class ConvertClipAttrToInput(Transformer):
    """
    https://github.com/onnx/onnx/blob/master/docs/Operators.md#softmax
    from: [max, min] in node.attribute
    to: [min, max] in node.input

    Assume NCHW Input
    """

    def transform(self, model: onnx.ModelProto) -> onnx.ModelProto:
        optimized_nodes = []
        for node in model.graph.node:
            if node.op_type != 'Clip':
                optimized_nodes.append(node)
                continue

            if len(node.input) >= 2:
                optimized_nodes.append(node)
                continue

            node_input = node.input[0]
            node_output = node.output[0]

            input_names = dict()
            added_inits = dict()

            input_names['min'] = ''
            input_names['max'] = ''
            # The filter() method constructs an iterator from elements of an iterable for which a function returns true.
            for attr in filter(lambda x: x.name == 'min' or x.name == 'max', node.attribute):
                tensor_name = f'{node.input[0]}_clip_{attr.name}'
                tensor = make_tensor(tensor_name, TensorProto.FLOAT, (), [attr.f])
                input_names[attr.name] = tensor_name
                added_inits[attr.name] = tensor

            model.graph.initializer.extend([*added_inits.values()])

            new_node = make_node('Clip',
                                 inputs=[node_input, input_names['min'], input_names['max']],
                                 outputs=[node_output])

            optimized_nodes.append(new_node)

        # remove duplicate node(s) in optimized nodes
        seen = []
        for op_node in optimized_nodes:
            if op_node in seen:
                continue
            seen.append(op_node)
        optimized_nodes = seen

        model = utils.rebuild_model(model, optimized_nodes)
        check_model(model)

        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:
    optimized_nodes = []
    for node in model.graph.node:
        if node.op_type != 'Clip':
            optimized_nodes.append(node)
            continue

        if len(node.input) >= 2:
            optimized_nodes.append(node)
            continue

        node_input = node.input[0]
        node_output = node.output[0]

        input_names = dict()
        added_inits = dict()

        input_names['min'] = ''
        input_names['max'] = ''
        # The filter() method constructs an iterator from elements of an iterable for which a function returns true.
        for attr in filter(lambda x: x.name == 'min' or x.name == 'max', node.attribute):
            tensor_name = f'{node.input[0]}_clip_{attr.name}'
            tensor = make_tensor(tensor_name, TensorProto.FLOAT, (), [attr.f])
            input_names[attr.name] = tensor_name
            added_inits[attr.name] = tensor

        model.graph.initializer.extend([*added_inits.values()])

        new_node = make_node('Clip',
                             inputs=[node_input, input_names['min'], input_names['max']],
                             outputs=[node_output])

        optimized_nodes.append(new_node)

    # remove duplicate node(s) in optimized nodes
    seen = []
    for op_node in optimized_nodes:
        if op_node in seen:
            continue
        seen.append(op_node)
    optimized_nodes = seen

    model = utils.rebuild_model(model, optimized_nodes)
    check_model(model)

    return model