Pixeltable YOLOX

pixeltable-yolox is a lightweight, Apache-licensed object detection library built on PyTorch. It is a fork of the MegVii YOLOX package originally authored by Zheng Ge et al, modernized for recent versions of Python and refactored to be more easily usable as a Python library.

pixeltable-yolox is still a work in progress! Some features of YoloX have not been ported yet.

​

Installation

pip install pixeltable-yolox

​

Inference

import requests
from PIL import Image
from yolox.models import Yolox, YoloxProcessor

url = "https://raw.githubusercontent.com/pixeltable/pixeltable-yolox/main/tests/data/000000000001.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model = Yolox.from_pretrained("yolox_s")
processor = YoloxProcessor("yolox_s")
tensor = processor([image])
output = model(tensor)
result = processor.postprocess([image], output)

# The labels are COCO category indices.
from yolox.data.datasets import COCO_CLASSES
print(COCO_CLASSES[result[0]['labels'][0]]) # Example: Print the first detected object's class name

This yields the following output:

[{'bboxes': [
   (272.36126708984375, 3.5648040771484375, 640.4871826171875, 223.2653350830078),
   (26.643890380859375, 118.68254089355469, 459.80706787109375, 315.089111328125),
   (259.41485595703125, 152.3223114013672, 295.37054443359375, 230.41783142089844)],
  'scores': [0.9417160943584335, 0.8170979975670818, 0.8095869439224117],
  'labels': [7, 2, 12]}]

The labels are COCO category indices.

from yolox.data.datasets import COCO_CLASSES

print(COCO_CLASSES[7])
# Output: 'truck'

​

Training

First unpack a COCO dataset into ./datasets/COCO:

COCO/
  annotations/
    instances_train2017.json
    instances_val2017.json
  train2017/
    # image files
  val2017/
    # image files

Then on the command line:

yolox train -c yolox-s -d 8 -b 64 --fp16 -o

For help:

yolox train -h

​

Background

​

Who are we and why are we doing this?

Pixeltable, Inc. is a venture-backed AI infrastructure startup. Our core product is Pixeltable, a database and orchestration system purpose-built for multimodal AI workloads.

Pixeltable integrates with numerous AI services and open source technologies. In the course of integrating with YOLOX, it became clear that there is a strong need in the vision community for a lightweight object detection library with an untainted open source license. It also became clear that while YOLOX is an excellent foundation, it would benefit greatly from code modernization and more regular updates.

We chose to build upon YOLOX both to simplify our own integration, and also to give something back to the community that will (hopefully) prove useful. The Pixeltable team has decades of collective experience in open source development. Our backgrounds include companies such as Google, Cloudera, Twitter, Amazon, and Airbnb, that have a strong commitment to open source development and collaboration.

​

Contributing

We welcome contributions from the community! If you’re interested in helping maintain and improve pixeltable-yolox, please stay tuned for our contributor’s guide which will be published after the initial release.