[Rewriter]: add fusion rules for successive Min/Max patterns

[AyoubMDL]

This PR adds the following transformation:

• Min(Min(X)) -> Min(X)
• Max(Max(X)) -> Max(X)
• Min(Max(X)) -> Clip(X)
• Max(Min(X)) -> Clip(X)

[Copilot]

Pull Request Overview

This PR implements fusion rules for successive Min/Max patterns in the ONNX rewriter. The PR adds optimizations to simplify chains of Min/Max operations by collapsing redundant operations and converting specific patterns to more efficient Clip operations.

• Adds fusion of successive Min operations: Min(Min(X)) -> Min(X)
• Adds fusion of successive Max operations: Max(Max(X)) -> Max(X)
• Adds conversion of Min(Max(X)) and Max(Min(X)) patterns to Clip operations

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
onnxscript/rewriter/min_max_to_clip.py	Implements the core rewrite rules with abstract base class and concrete fusion implementations
onnxscript/rewriter/min_max_to_clip_test.py	Comprehensive test suite covering successful fusion cases and failure conditions

[codecov]

Codecov Report

❌ Patch coverage is 98.85057% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 70.17%. Comparing base (1934901) to head (e2125dc).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
...ipt/rewriter/rules/common/_min_max_to_clip_test.py	98.00%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2500      +/-   ##
==========================================
+ Coverage   69.99%   70.17%   +0.18%     
==========================================
  Files         216      218       +2     
  Lines       26074    26248     +174     
  Branches     2618     2625       +7     
==========================================
+ Hits        18250    18420     +170     
- Misses       6921     6923       +2     
- Partials      903      905       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

Thanks! I think we can enable these patterns by default. Could you share on what models these patterns will appear?

[AyoubMDL]

Thanks! I think we can enable these patterns by default. Could you share on what models these patterns will appear?

I originally added the rewrite for the Max(Min) → Clip fusion to handle cases where I needed to export a Keras model with a bounded ReLU to ONNX. For example, the following model uses ReLU(max_value=4.0):

import keras
import tensorflow as tf
import tf2onnx


def build_model(input_shape=(28, 28, 1), max_relu=4.0):
    inputs = keras.Input(shape=input_shape)
    x = keras.layers.Conv2D(8, (3, 3), padding="same")(inputs)
    outputs = keras.layers.ReLU(max_value=max_relu)(x)
    model = keras.Model(inputs=inputs, outputs=outputs)
    return model


model = build_model()

spec = (tf.TensorSpec((None, 28, 28, 1), tf.float32, name="input"),)

model_proto, _ = tf2onnx.convert.from_keras(
    model,
    input_signature=spec,
    opset=15,
    output_path="model.onnx",
)

This will generate the following onnx model which has Max(Min):

I added the adder rewrites mainly for generality, but I’m open to remove them if we feel they don’t add value.

[justinchuby]

I think the rules are reasonable. I will look at them more closely

[gramalingam]

LGTM ... had just a couple of minor comments/suggestions. Thanks!