Hello. This time I compared the speed of loading images and converting them to tensors in PyTorch using various libraries.

Libraries Used

• opencv-python: 4.5.3.56
• Pillow(PIL): 8.3.2
• Pillow-SIMD: 7.0.0.post3
• torchvision.io(PIL): 0.10.1
• torchvision.io(accimage): 0.10.1
• scikit-image: 0.18.3

I’ll omit the installation methods for each.

Pillow-SIMD is a high-speed version of Pillow, but it hasn’t been updated recently and stopped at version 7.0.0. accimage is a library that can be specified as a backend for torchvision and can be installed with conda.

Experiment

The source code is as follows: Referenced from here.

import time
from contextlib import contextmanager

import cv2
from PIL import Image
import skimage.io
import torchvision
import torchvision.transforms.functional as TF
from torchvision.io import read_image


@contextmanager
def timer(name):
    t0 = time.time()
    yield
    print(f"[{name}] done in {(time.time() - t0)*1000:.03f} ms")


N_ITERS = 100
for FILENAME in ["./image.jpg","./image.png"]:
    print(FILENAME)
    with timer("PIL"):
        for i in range(N_ITERS):
            img = Image.open(FILENAME).convert("RGB")
            img = TF.to_tensor(img)

    with timer("OpenCV"):
        for i in range(N_ITERS):
            img = cv2.imread(FILENAME)[:, :, ::-1].copy()
            img = TF.to_tensor(img)

    with timer("default_torchvision"):
        for i in range(N_ITERS):
            img = read_image(FILENAME)

    with timer("accimage_torchvision"):
        torchvision.set_image_backend("accimage")
        for i in range(N_ITERS):
            img = read_image(FILENAME)

    with timer("skimage"):
        for i in range(N_ITERS):
            img = skimage.io.imread(FILENAME)
            img = TF.to_tensor(img)

I’ve also published the source on GitHub:

The commonly used mandrill image was used for testing.

Image Format	Library	Speed(ms)
jpg	opencv-python	673.799
	Pillow(PIL)	772.203
	Pillow-SIMD	648.018
	torchvision.io(PIL)	301.993
	torchvision.io(PIL-SIMD)	278.904
	torchvision.io(accimage)	274.502
	scikit-image	794.349
	scikit-image(PIL-SIMD)	545.362
png	opencv-python	1066.164
	Pillow(PIL)	1752.472
	Pillow-SIMD	1085.250
	torchvision.io(PIL)	877.775
	torchvision.io(PIL-SIMD)	867.422
	torchvision.io(accimage)	847.297
	scikit-image	1878.370
	scikit-image(PIL-SIMD)	1108.291

The fastest was torchvision.io.read_image. The acceleration with Pillow-SIMD also affects scikit-image, presumably because it’s used internally. However, torchvision didn’t show the effects of Pillow-SIMD.

While at it, I also compared image loading speed. Sometimes you might not convert to tensors. For torchvision.io.read_image which outputs tensors, I measured up to converting to ndarray with numpy().

Image Format	Library	Speed(ms)
jpg	opencv-python	457.877
	Pillow(PIL)	772.203
	Pillow-SIMD	648.018
	torchvision.io	273.161
	scikit-image	794.349
	scikit-image(PIL-SIMD)	545.362
png	opencv-python	962.926
	Pillow(PIL)	1752.472
	Pillow-SIMD	1085.250
	torchvision.io	795.085
	scikit-image	1878.370
	scikit-image(PIL-SIMD)	1108.291

Surprisingly, torchvision.io was also the fastest here.

Caution

Since Pillow-SIMD is based on an older Pillow, the following warning appears:

UserWarning: Your installed pillow version is < 7.1.0. Several security issues (CVE-2020-11538, CVE-2020-10379, CVE-2020-10994, CVE-2020-10177) have been fixed in pillow 7.1.0 or higher. We recommend to upgrade this library.

Since there appear to be security risks, it’s safer not to use it.