String upper slower in onnx than plain python

benchmark.py

import timeit
import numpy as np

def setup_code(model_name: str):
    return f"""
import string
import random
import onnxruntime as _ort
from onnxruntime_extensions import get_library_path as _lib_path
import numpy as np
import tensorflow as tf
from keras.layers import BatchNormalization, Identity, Layer

import onnx
from onnx import helper
import tf2onnx


def generate_random_string(string_len):
    return ''.join(random.choice(string.ascii_lowercase) for _ in range(string_len))


def generate_input(num_samples, num_features, string_len):
    return [[generate_random_string(string_len) for _ in range(num_features)] for _ in range(num_samples)]


so = _ort.SessionOptions()

so.register_custom_ops_library(_lib_path())
so.graph_optimization_level = _ort.GraphOptimizationLevel.ORT_ENABLE_ALL

session = _ort.InferenceSession("{model_name}.onnx", so)

input_data = {{
    'string_input': generate_input(1, 10, 8)
}}
"""

# Setup code
test_code = """
input_data["string_input"] = [[string.upper() for string in l] for l in input_data["string_input"]]
session.run(None, input_data)
"""

# Timeit statement
times = timeit.repeat(setup=setup_code("onnx_without_upper"), stmt=test_code, repeat=10, number=1000)
print(f"Execution time: {np.mean(times) * 1000} +/- {3 * np.std(times) * 1000} us")

test_code = """
session.run(None, input_data)
"""

# Timeit statement
times = timeit.repeat(setup=setup_code("onnx_with_upper"), stmt=test_code, repeat=10, number=1000)
print(f"Execution time: {np.mean(times) * 1000}  +/- {3 * np.std(times) * 1000} us")

onnx_with_upper.onnx

onnx_without_upper.onnx

requirements.txt

tensorflow==2.15.0
onnx
onnx_extensions