[BUG] Pass "fuse_consecutive_squeezes" causes output mismatch

[BUG] Pass "fuse_consecutive_squeezes" causes output mismatch

Issue
Running the single pass fuse_consecutive_squeezes with onnxoptimizer 0.3.19 changes numerical outputs. A real model regresses only after this pass, while the original model runs correctly with the same oracle inputs.

Environment

• Ubuntu 20.04
• Python 3.10
• onnx==1.19.0
• onnxruntime==1.23.2
• onnxoptimizer==0.3.19 (latest)

Repro steps (run from this folder)

1. Download and unzip the attached archive, then cd into the extracted directory

fuse_consecutive_squeezes_repro.tar.gz

tar -xzvf fuse_consecutive_squeezes_repro.tar.gz
cd fuse_consecutive_squeezes_repro

2. Create a Python environment (Python 3.10) and install dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

3. Optimize a case with only fuse_consecutive_squeezes:

• python optimize_model.py --case ./case_02867_seed42954792
  (Writes model.opt.onnx next to model.onnx in the case directory.)

4. Differential test original vs optimized outputs using stored oracle inputs:

• python diff_test.py --case ./case_02867_seed42954792

Observed results

• case_02867_seed42954792: overall max_abs=4.883e-04, max_rel=6.431e-04; only one output tensor shows a small but reproducible mismatch, others match.

Expected
fuse_consecutive_squeezes should be semantics-preserving. Applying this pass alone should not change any output values. Please investigate why the optimized graph produces different outputs and whether the pass is incorrectly changing squeeze/unsqueeze ordering or shapes.

Differential Test Output Details

1. case_02867_seed42954792

Case: case_02867_seed42954792
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 26)
  output[1]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(4, 1)
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(4, 1)
  output[3]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(4, 4)
  output[4]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(34,)
  output[5]: max_abs=4.883e-04, max_rel=6.431e-04, shape=(4,)
  output[6]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(2, 1, 4)
  output[7]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(4, 34)
  output[8]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(34,)
  output[9]: max_abs=0.000e+00, max_rel=0.000e+00, shape=()
  output[10]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(34,)
  output[11]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 2, 17)
  output[12]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 2, 2)
  output[13]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 1, 3)
Overall: max_abs=4.883e-04, max_rel=6.431e-04

**Attachments**
- `README.md` (this document)
- `requirements.txt` (dependency versions)
- `optimize_model.py` (runs only `fuse_consecutive_squeezes` and saves `model.opt.onnx`)
- `diff_test.py` (runs original vs optimized with oracle inputs and reports max_abs/max_rel)
- `case_02867_seed42954792/` (contains `model.onnx` original model, `oracle.pkl` input data for both original and optimized models)

Start by running optimize_model.py for case_02867_seed42954792, then use diff_test.py with model.onnx and oracle.pkl to reproduce the mismatch from fuse_consecutive_squeezes alone. Trace the pass implementation and its squeeze/unsqueeze transformations, using the reported output[5] difference as the check. Done means the optimized model matches the original outputs for this case.