Profile the execution of a simple model

ONNX Runtime can profile the execution of the model. This example shows how to interpret the results.

import numpy
import onnx

import onnxruntime as rt
from onnxruntime.datasets import get_example


def change_ir_version(filename, ir_version=6):
    "onnxruntime==1.2.0 does not support opset <= 7 and ir_version > 6"
    with open(filename, "rb") as f:
        model = onnx.load(f)
    model.ir_version = 6
    if model.opset_import[0].version <= 7:
        model.opset_import[0].version = 11
    return model

Let’s load a very simple model and compute some prediction.

example1 = get_example("mul_1.onnx")
onnx_model = change_ir_version(example1)
onnx_model_str = onnx_model.SerializeToString()
sess = rt.InferenceSession(onnx_model_str, providers=rt.get_available_providers())
input_name = sess.get_inputs()[0].name

x = numpy.array([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]], dtype=numpy.float32)
res = sess.run(None, {input_name: x})
print(res)

[array([[ 1.,  4.],
       [ 9., 16.],
       [25., 36.]], dtype=float32)]

We need to enable to profiling before running the predictions.

options = rt.SessionOptions()
options.enable_profiling = True
sess_profile = rt.InferenceSession(onnx_model_str, options, providers=rt.get_available_providers())
input_name = sess.get_inputs()[0].name

x = numpy.array([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]], dtype=numpy.float32)

sess.run(None, {input_name: x})
prof_file = sess_profile.end_profiling()
print(prof_file)

onnxruntime_profile__2023-03-29_16-08-33.json

The results are stored un a file in JSON format. Let’s see what it contains.

import json

with open(prof_file, "r") as f:
    sess_time = json.load(f)
import pprint

pprint.pprint(sess_time)

[{'args': {},
  'cat': 'Session',
  'dur': 72,
  'name': 'model_loading_array',
  'ph': 'X',
  'pid': 3209769,
  'tid': 3209769,
  'ts': 2},
 {'args': {},
  'cat': 'Session',
  'dur': 307,
  'name': 'session_initialization',
  'ph': 'X',
  'pid': 3209769,
  'tid': 3209769,
  'ts': 94}]