TILI: Turbo Image Loading Library

Introduction

Turbo Image Loading Library (TILI) is a collection of highly optimized library for Pytorch to accelerate the pre-processing of the image data for deep learning applications.

Highlights

Full accelerated pipeline from reading the disk to getting ready for inference
Flexibility through configurable pipeline and custom operators
Accelated by libjpeg-turbo (accimage), BackgroundGenerator (prefetch_generator) and pytorch with CUDA.

Benchmark

The benchmark run on 1K ImageNet test samples (imagenet_1k.zip) by Tesla P4 GPU with the simplest transforms (ImageDecoder, Resize, Normalize).

Image Size	256x256	512x512
torchvision (baseline)	9.7818 ms/p	12.5466 ms/p
+ libjpeg-turbo	4.3591 ms/p	5.6756 ms/p
w/o CUDA resize	3.7838 ms/p	4.8411 ms/p
w/ CUDA resize	4.0056 ms/p	4.0331 ms/p

Usage

Dependencies

$ conda install -c conda-forge accimage
$ pip install prefetch_generator

Examples

A simple example to use TILI for inference.

from labels import classname
import torch
import tili
from tili import transforms
from tili.pipeline import ImagePipline
from torchvision.models import mobilenet_v2
model = mobilenet_v2(pretrained=True)
model.eval()
if torch.cuda.is_available():
    model = model.cuda()

trans = transforms.Compose([
        transforms.ResizedCrop((224, 224)),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
    ])
pipe = ImagePipline(trans)
img = pipe('cat.jpg')
with torch.no_grad():
    y = model(img)
print("Predict:", classname[y.argmax().item()])

A simple example to use the other TILI transforms.

import numpy as np
import torch
import tili
from tili import transforms
from tili.pipeline import ImagePipline
trans = transforms.Compose([
        transforms.Resize((448, 448)),
        transforms.ResizedCrop((256, 256)),
        transforms.CenterCrop(224),
        transforms.Pad((16, 16, 16, 16), padding_mode='reflect'),
        transforms.HorizontalFlip(),
        transforms.VerticalFlip(),
        transforms.CvtColor('RGB2BGR'),
    ])
pipe = ImagePipline(trans)
img = pipe('cat.jpg')
npimg = img.squeeze_(0).data.numpy()
npimg = np.transpose(npimg, (1, 2, 0))
cv2.imshow('img', np.uint8(npimg*255))
cv2.waitKey()

Contact

Please contact ArlenCai (arlencai@tencent.com)

caibolun/TILI