This is a ROS2 wrapper for the data-efficient transformer with an image classification head, DEIT. We utilize huggingface
and the transformers
for the source of the algorithm. The main idea is for this container to act as a standalone interface and node, removing the necessity to integrate separate packages and solve numerous dependency issues. We have built all different versions of the algorithm (i.e. base-patch16-224
, large-patch16-224
, etc.) for each ROS2 distribution.
Data-efficient Image Transformer (DeiT) is model that was pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224 (we also provide other resolutions in the model versions below).
- Install Docker and ensure the Docker daemon is running in the background.
- Run
docker pull shaderobotics/deit:${ROS2_DISTRO}-${MODEL_VERSION}
we support all ROS2 distributions along with all model versions found in the model version section below. - Follow the run commands in the usage section below
- Install Docker and ensure the Docker daemon is running in the background.
- Clone this repo with
git pull https://github.com/open-shade/deit.git
- Enter the repo with
cd deit
- To pick a specific model version, edit the
ALGO_VERSION
constant in/deit/deit.py
- Build the container with
docker build . -t [name]
. This will take a while. We have also provided associatedcloudbuild.sh
scripts to build on GCP all of the associated versions. - Follow the run commands in the usage section below.
base-patch16-224
tiny-patch16-224
base-distilled-patch16-224
small-patch16-224
tiny-distilled-patch16-224
base-distilled-patch16-384
small-distilled-patch16-224
More information about these versions can be found in the paper. tiny
, base
, large
, represent the number of weights stored (i.e. the size of the model). 384 is the resolution size of the image.
docker pull shaderobotics/deit:foxy-base-patch16-224
Run docker run -t --net=host shaderobotics/deit:${ROS_DISTRO}-${MODEL_VERSION}
. Your node should be running now. Then, by running ros2 topic list,
you should see all the possible pub and sub routes.
For more details explaining how to run Docker images, visit the official Docker documentation here. Also, additional information as to how ROS2 communicates between external environment or multiple docker containers, visit the official ROS2 (foxy) docs here.
Name | IO | Type | Use |
---|---|---|---|
deit/image_raw | sub | sensor_msgs.msg.Image | Takes the raw camera output to be processed |
deit/result | pub | String | Outputs the classification label from ImageNet 100 Classes as a string |
To test and ensure that this package is properly installed, replace the Dockerfile in the root of this repo with what exists in the demo folder. Installed in the demo image contains a camera stream emulator by klintan which directly pubs images to the DEIT node and processes it for you to observe the outputs.
To run this, run docker build . -t [name]
, then docker run --net=host -t [name]
. Observing the logs for this will show you what is occuring within the container. If you wish to enter the running container and preform other activities, run docker ps
, find the id of the running container, then run docker exec -it [containerId] /bin/bash