op.Pad: inconsistency between @script and rewriter

[leshabirukov]

@script()
def dw_patch( in_X  ):
    X = op.Pad( in_X, pads=[0, 0, 1, 1, 0, 0, 0, 0] )
    return X

This one translates to:

dw_patch ( in_X) => ( return_val) {
   const = Constant <value: tensor = int64[8] const {0,0,1,1,0,0,0,0}> ()
   return_val = Pad (in_X, const)
}

def patch( op, in_X, ):
    return op.Pad( in_X, pads=[0, 0, 1, 1, 0, 0, 0, 0] )

And this one, (used as replace part in rewriter) to

ini_mod (float[1,3,28,28] in_X) => (float[1,3,29,29] return_val) {
   return_val = Pad <pads: ints = [0, 0, 1, 1, 0, 0, 0, 0]> (in_X)
}

Here, input becomes attribute.

what_the_pad.py.txt

log.txt

[justinchuby]

In the rewriter, all kwargs are turned into attributes, whereas in script mode the logic understands the op schema and is able to know that it should be an input. We will work the bridge this gap but it may take some time

[leshabirukov]

Can I pass @script()-ed function to rewriter.pattern.RewriteRule?

Get

TypeError: Unexpected input type <class 'onnxscript.rewriter.pattern.OpsetPatternBuilder'>.

[justinchuby]

Not yet. We are going to bridge this gap later