Measuring performance of TfIdfVectorizer

The banchmark measures the performance of a TfIdfVectizer along two parameters, the vocabulary size, the batch size whether. It measures the benefit of using sparse implementation through the computation time and the memory peak.

A simple model

We start with a model including only one node TfIdfVectorizer. It only contains unigram. The model processes only sequences of 10 integers. The sparsity of the results is then 10 divided by the size of vocabulary.

import gc
import time
import itertools
from typing import Tuple
import numpy as np
import pandas
from onnx import ModelProto
from onnx.helper import make_attribute
from tqdm import tqdm
from onnxruntime import InferenceSession, SessionOptions
from onnx_extended.ext_test_case import measure_time, unit_test_going
from onnx_extended.memory_peak import start_spying_on
from onnx_extended.reference import CReferenceEvaluator
from onnx_extended.ortops.optim.cpu import get_ort_ext_libs
from onnx_extended.plotting.benchmark import vhistograms


def make_onnx(n_words: int) -> ModelProto:
    from skl2onnx.common.data_types import Int64TensorType, FloatTensorType
    from skl2onnx.algebra.onnx_ops import OnnxTfIdfVectorizer

    # from onnx_array_api.light_api import start
    # onx = (
    #     start(opset=19, opsets={"ai.onnx.ml": 3})
    #     .vin("X", elem_type=TensorProto.INT64)
    #     .ai.onnx.TfIdfVectorizer(
    #     ...
    #     )
    #     .rename(Y)
    #     .vout(elem_type=TensorProto.FLOAT)
    #     .to_onnx()
    # )
    onx = OnnxTfIdfVectorizer(
        "X",
        mode="TF",
        min_gram_length=1,
        max_gram_length=1,
        max_skip_count=0,
        ngram_counts=[0],
        ngram_indexes=np.arange(n_words).tolist(),
        pool_int64s=np.arange(n_words).tolist(),
        output_names=["Y"],
    ).to_onnx(inputs=[("X", Int64TensorType())], outputs=[("Y", FloatTensorType())])
    #     .rename(Y)
    #     .vout(elem_type=TensorProto.FLOAT)
    #     .to_onnx()
    # )
    return onx


onx = make_onnx(7)
ref = CReferenceEvaluator(onx)
got = ref.run(None, {"X": np.array([[0, 1], [2, 3]], dtype=np.int64)})
print(got)

[array([[1., 1., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 0., 0., 0.]], dtype=float32)]

It works as expected. Let’s now compare the execution with onnxruntime for different batch size and vocabulary size.

Benchmark

def make_sessions(
    onx: ModelProto,
) -> Tuple[InferenceSession, InferenceSession, InferenceSession]:
    # first: onnxruntime
    ref = InferenceSession(onx.SerializeToString(), providers=["CPUExecutionProvider"])

    # second: custom kernel equivalent to the onnxruntime implementation
    for node in onx.graph.node:
        if node.op_type == "TfIdfVectorizer":
            node.domain = "onnx_extended.ortops.optim.cpu"
            # new_add = make_attribute("sparse", 1)
            # node.attribute.append(new_add)

    d = onx.opset_import.add()
    d.domain = "onnx_extended.ortops.optim.cpu"
    d.version = 1

    r = get_ort_ext_libs()
    opts = SessionOptions()
    opts.register_custom_ops_library(r[0])
    cus = InferenceSession(
        onx.SerializeToString(), opts, providers=["CPUExecutionProvider"]
    )

    # third: with sparse
    for node in onx.graph.node:
        if node.op_type == "TfIdfVectorizer":
            new_add = make_attribute("sparse", 1)
            node.attribute.append(new_add)
    cussp = InferenceSession(
        onx.SerializeToString(), opts, providers=["CPUExecutionProvider"]
    )

    return ref, cus, cussp


if unit_test_going():
    vocabulary_sizes = [10, 20]
    batch_sizes = [5, 10]
else:
    vocabulary_sizes = [100, 1000, 5000, 10000]
    batch_sizes = [1, 10, 500, 1000, 2000]
confs = list(itertools.product(vocabulary_sizes, batch_sizes))

data = []
for voc_size, batch_size in tqdm(confs):
    onx = make_onnx(voc_size)
    ref, cus, sparse = make_sessions(onx)
    gc.collect()

    feeds = dict(
        X=(np.arange(batch_size * 10) % voc_size)
        .reshape((batch_size, -1))
        .astype(np.int64)
    )

    # sparse
    p = start_spying_on(delay=0.0001)
    sparse.run(None, feeds)
    obs = measure_time(
        lambda sparse=sparse, feeds=feeds: sparse.run(None, feeds), max_time=1
    )
    mem = p.stop()
    obs["peak"] = mem["cpu"].max_peak - mem["cpu"].begin
    obs["name"] = "sparse"
    obs.update(dict(voc_size=voc_size, batch_size=batch_size))
    data.append(obs)
    time.sleep(0.1)

    # reference
    p = start_spying_on(delay=0.0001)
    ref.run(None, feeds)
    obs = measure_time(lambda ref=ref, feeds=feeds: ref.run(None, feeds), max_time=1)
    mem = p.stop()
    obs["peak"] = mem["cpu"].max_peak - mem["cpu"].begin
    obs["name"] = "ref"
    obs.update(dict(voc_size=voc_size, batch_size=batch_size))
    data.append(obs)
    time.sleep(0.1)

    # custom
    p = start_spying_on(delay=0.0001)
    cus.run(None, feeds)
    obs = measure_time(lambda cus=cus, feeds=feeds: cus.run(None, feeds), max_time=1)
    mem = p.stop()
    obs["peak"] = mem["cpu"].max_peak - mem["cpu"].begin
    obs["name"] = "custom"
    obs.update(dict(voc_size=voc_size, batch_size=batch_size))
    data.append(obs)
    time.sleep(0.1)

    del sparse
    del cus
    del ref
    del feeds

df = pandas.DataFrame(data)
df["time"] = df["average"]
df.to_csv("plot_op_tfidfvectorizer_sparse.csv", index=False)
print(df.head())

  0%|          | 0/20 [00:00<?, ?it/s]
  5%|▌         | 1/20 [00:03<01:15,  3.98s/it]
 10%|█         | 2/20 [00:08<01:12,  4.04s/it]
 15%|█▌        | 3/20 [00:12<01:08,  4.02s/it]
 20%|██        | 4/20 [00:16<01:08,  4.29s/it]
 25%|██▌       | 5/20 [00:20<01:01,  4.11s/it]
 30%|███       | 6/20 [00:24<00:57,  4.14s/it]
 35%|███▌      | 7/20 [00:28<00:52,  4.02s/it]
 40%|████      | 8/20 [00:32<00:48,  4.05s/it]
 45%|████▌     | 9/20 [00:36<00:44,  4.02s/it]
 50%|█████     | 10/20 [00:41<00:41,  4.19s/it]
 55%|█████▌    | 11/20 [00:45<00:37,  4.15s/it]
 60%|██████    | 12/20 [00:49<00:32,  4.05s/it]
 65%|██████▌   | 13/20 [00:53<00:29,  4.15s/it]
 70%|███████   | 14/20 [00:57<00:25,  4.20s/it]
 75%|███████▌  | 15/20 [01:01<00:20,  4.12s/it]
 80%|████████  | 16/20 [01:05<00:16,  4.10s/it]
 85%|████████▌ | 17/20 [01:09<00:12,  4.09s/it]
 90%|█████████ | 18/20 [01:14<00:08,  4.22s/it]
 95%|█████████▌| 19/20 [01:18<00:04,  4.20s/it]
100%|██████████| 20/20 [01:22<00:00,  4.17s/it]
100%|██████████| 20/20 [01:22<00:00,  4.13s/it]
    average     deviation  min_exec  ...  voc_size  batch_size      time
0  0.000007  7.581871e-07  0.000007  ...       100           1  0.000007
1  0.000006  8.600416e-08  0.000006  ...       100           1  0.000006
2  0.000006  5.242533e-07  0.000005  ...       100           1  0.000006
3  0.000013  3.670098e-07  0.000012  ...       100          10  0.000013
4  0.000019  5.916206e-06  0.000015  ...       100          10  0.000019

[5 rows x 14 columns]

Processing time

piv = pandas.pivot_table(
    df, index=["voc_size", "name"], columns="batch_size", values="average"
)
print(piv)

batch_size           1         10        500       1000      2000
voc_size name
100      custom  0.000006  0.000006  0.000023  0.000054  0.000050
         ref     0.000006  0.000019  0.000027  0.000051  0.000076
         sparse  0.000007  0.000013  0.000317  0.000523  0.001015
1000     custom  0.000005  0.000007  0.000054  0.000085  0.000181
         ref     0.000006  0.000014  0.000077  0.000150  0.000291
         sparse  0.000006  0.000012  0.000299  0.000599  0.001170
5000     custom  0.000006  0.000010  0.000200  0.000389  0.000837
         ref     0.000008  0.000017  0.000286  0.000566  0.001150
         sparse  0.000006  0.000011  0.000295  0.000557  0.001005
10000    custom  0.000006  0.000014  0.000361  0.000788  0.002377
         ref     0.000010  0.000022  0.000572  0.001207  0.002468
         sparse  0.000006  0.000011  0.000263  0.000638  0.001057

Memory peak

It is always difficult to estimate. A second process is started to measure the physical memory peak during the execution every ms. The figures is the difference between this peak and the memory when the measurement began.

piv = pandas.pivot_table(
    df, index=["voc_size", "name"], columns="batch_size", values="peak"
)
print(piv / 2**20)

batch_size           1         10         500        1000       2000
voc_size name
100      custom  0.000000  0.000000   0.000000   0.000000   0.003906
         ref     0.000000  0.000000   0.011719   0.000000   0.000000
         sparse  0.019531  0.000000   0.035156   0.000000   0.000000
1000     custom  0.000000  0.000000   0.152344   0.000000   0.000000
         ref     0.000000  0.007812   0.007812   0.003906   0.011719
         sparse  0.000000  0.000000   0.000000   0.000000   0.000000
5000     custom  0.000000  0.000000   9.042969  20.640625  39.378906
         ref     0.000000  0.000000   0.015625   0.015625  38.640625
         sparse  0.000000  0.000000   0.000000   0.000000   0.000000
10000    custom  0.000000  0.000000  20.640625  39.378906  77.382812
         ref     0.000000  0.000000   0.015625  38.640625  77.378906
         sparse  0.000000  0.000000   0.000000   0.000000   0.000000

Graphs

ax = vhistograms(df)
fig = ax[0, 0].get_figure()
fig.savefig("plot_op_tfidfvectorizer_sparse.png")

Take away

Sparse works better when the sparsity is big enough and the batch size as well.