For learning computer vision systems in general, an app that detects objects(eventually, do instance segmentation instead) and translates the object names into different languages.
The model currently handles a video stream and returns the bounding box the primary object in the stream.
make serving-build
Is used to build the docker image. By default, the mobilenet_v3_large is used. Eager-mode pre-trained Pytorch models are defined in the model directory and can be expanded to add new backbone architectures to the object detection model. To use a different model, the image can be built with a MODEL_NAME arg:
make serving-build MODEL_NAME=resnet50
Before installing the app dependencies via make, pyenv and poetry have to be installed on your system. Running:
make venv-activate
make install
will install the app's dependencies.
NOTE: the Dockerfile in the root directory packages the step above, but on MacOS the webcam will not work inside a container.
- Bring up the
torchservecontainer:
make serving-start
to bring up the model service at http://localhost:8080.
- Start the app:
poetry run streamlit run app/main.py