Pattern match on default attributes?

[leshabirukov]

Problem: rewrite Conv nodes, but only with dilations==[1,1]. Issue - default value. To check dilation, I need include it in pattern, but this cause pattern match to fail on nodes with default dilations.

Probably related: #1627

Possible fix: substitute defaults on deserialization, I suspect, such a function already exists.

rewr_c.py.txt

[justinchuby]

Thanks. So a node with default and without default is store differently in graph. They are semantically the same but syntactically different. We need to decide what the behavior of the matcher should be.

[leshabirukov]

Well, it can be outside of matcher, if one could force defaults for node.

But as I found, it may be not formalized at all!

from onnx import helper, reference

schema = reference.op_run.get_schema('Conv', 15)
if schema.attributes:
    for _, attr in sorted(schema.attributes.items()):
        default_value = helper.get_attribute_value(attr.default_value)
        print(f'{attr.name=:20}{default_value=}')

attr.name='auto_pad'
default_value=b'NOTSET'
attr.name='dilations'
default_value=None
attr.name='group'
default_value=1
attr.name='kernel_shape'
default_value=None
attr.name='pads'
default_value=None
attr.name='strides'
default_value=None

onnx.reference.op_run.get_schema('Conv', 11).attributes['dilations']

OpSchema.Attribute(name='dilations', type=<AttrType.INTS: 7>, description='dilation value along each spatial axis of the filter. If not present, the dilation defaults is 1 along each spatial axis.', default_value=, required=False)

I wonder, how it handled in onnxruntime, just hardcoded?

[leshabirukov]

I would rather put defaults handler outside of matcher, with function like force_defaults(node) =>node, forced keys probably paired with strip_defaults

[Johansmm]

I had the same problem. For now I solved it through a condition function:

def target_pattern(op, x, w):
        return op.Conv(x, w, _outputs=["y"])

def replacement_pattern(op, x, w, **__):
        # Naïve pattern
       return op.Identity(op.Conv(x, w))

def condition_fn(*_, y, **__):
       ir_node = y.producer()
       return (dilatation := ir_node.attributes.get("dilatation", None)) and dilatation.value == [1,1]

rule = RewriteRule(target_pattern, replacement_pattern, condition_fn)

I tried the following target_pattern:

def target_pattern(op, x, w, dilatation=[1,1]):
        return op.Conv(x, w, dilatation=dilatation)

But unfortunately the condition was never triggered.

Let me know if there is a simpler solution.

[leshabirukov]

@Johansmm hey, yes, it is better then my force/strip_defaults!