Both models will be downloaded using the vit_h weights.
docker build -t sam-builder -f Dockerfile-build .
docker run -d --name sam-builder1 sam-builder
docker cp sam-builder1:/home/model-store ./
We copy these to model-store and use this locally by both the GPU and the CPU Torchserve containers.
you can delete the container once models are copied
docker rm -f sam-builder1
With the GPU, inference time should be about 1.8 seconds or less depending on the GPU. On an older 1080 Ti Pascal GPU, inference time is 1.67 seconds without compilation.
docker build -t sam-gpu -f Dockerfile-gpu .
bash start_serve_encode_gpu.sh
docker build -t sam-cpu -f Dockerfile-cpu .
bash start_serve_decode_cpu.sh
The CPU service is served on 7080 by default. 8080 for the GPU service by default.
curl http://127.0.0.1:7080/predictions/sam_vit_h_encode -T ./data/sample-img-fox.jpg
All tests in tests/ cover the functionality of the decoder. Logic in decode.py is run in pytest fixtures within conftest.py when outputs need to be shared by different tests in test_decode.py.
To start running the tests, make sure you have the test models. You should have the same models used during inference, including
(test.py3.10) (base) rave@rave-desktop:~/segment-anything-services/tests/models$ tree
.
├── sam_vit_h_4b8939.pth
└── sam_vit_h_decode.onnx
You can get both by unzipping the .mar archives copied to model-store from the sam-builder container you started in the previous step. Then, move the .onnx and the .pth files to ./tests/models/
unzip model-store/sam_vit_h_decode.mar -d ./sam_decode_mar
unzip model-store/sam_vit_h_encode.mar -d ./sam_encode_mar
cp ./sam_decode_mar/sam_vit_h_decode.onnx ./tests/models/
cp ./sam_decode_mar/sam_vit_h_4b8939.pth ./tests/models/
Install the testing environment with hatch: pip install hatch
Then, create the environment. I tested with Python 3.10, Python 3.11 does not work because of an onnxruntime version issue.
hatch -e test.py3.10 shell
Then, run tests with pytest
pytest tests
If you have access, download from the devseed s3:
aws s3 sync s3://segment-anything/model-weights/ model-weights
otherwise, get checkpoints from the original repo: https://github.com/facebookresearch/segment-anything/tree/main#model-checkpoints
This step takes a long time presumably because the uncompiled weights are massive. Packaging the ONNX model is faster in the later steps.
mkdir -p model_store_encode
torch-model-archiver --model-name sam_vit_h_encode --version 1.0.0 --serialized-file model-weights/sam_vit_h_4b8939.pth --handler handler_encode.py
mv sam_vit_h_encode.mar model_store_encode/sam_vit_h_encode.mar
mkdir -p models
python scripts/export_onnx_model.py --checkpoint model-weights/sam_vit_h_4b8939.pth --model-type vit_h --output models/sam_vit_h_decode.onnx
We'll put this in the model_store_decode directory, to keep the onnx model files distinct from the torchserve .mar model archives. model_store/ is created automatically by Torchserve in the container, which is why we're make a local folder here called "model_store_decode".
mkdir -p model_store_decode
torch-model-archiver --model-name sam_vit_h_decode --version 1.0.0 --serialized-file models/sam_vit_h_decode.onnx --handler handler_decode.py
mv sam_vit_h_decode.mar model_store_decode/sam_vit_h_decode.mar
Use this container to test the model in a GPU enabled jupyter notebook server with geospatial and pytorch dependencies installed.
docker build -t sam-dev -f Dockerfile-dev .
You can run test_endpoint.ipynb to then use the two running services you started above. The dependencies are minimal for this notebook, install them on your own or you can run them in the jupyter server below.
This is a GPU enabled container that is set up with SAM and some other dependencies we commonly use. You can use it to try out SAM model in a notebook environment. Remove the --gpus arg if you don't have a GPU.
docker run -it --rm \
-v $HOME/.aws:/root/.aws \
-v "$(pwd)":/segment-anything-services \
-p 8888:8888 \
-e AWS_PROFILE=devseed \
--gpus all sam-dev
- Install the dependencies to deploy from npm
npm install -g @openaddresses/deploy
- Set up your credentials by navigating to the AWS SSO login page, and selecting "Command Line Access" to copy the temporary credentials. paste these to
~/.aws/credentialsand rename the AWS profile (ask Ryan for this.) touch ~/.deployrc.jsonand fill it with
{
"profilename": {
"region": "us-east-1"
}
}
- run deploy init to generate a
.deployrc.jsonconfig for the repo. both the local and global config are needed to deploy. Fill in args when prompted based on the profile name - commit and make a PR with any changes. wait for all github actions to complete so that the model archives and docker images are built
deploy update prodto deploy changes to prod
Q: What GPUs was this tested with?
A: The encoder and decoder are tested locally with a 1080 Ti (Pascal) and 3090 (Ampere). In production, the encoder runs on a p3.2xlarge (Tesla V100). Older Kepler GPUs such as K80s are not tested. See this issue for guidance on adapting the docker images to work with Kepler series GPUs.
Q: Why two services?
A: We're exploring cost effective ways to run image encoding in a separate, on-demand way from the CPU decoder. Eventually we'd like to remove the need for the CPU torserve on the backend and run the decoding in the browser.
Q: Can I contribute or ask questions?
A: This is currently more of a "working in the open" type repo that we'd like to share with others, rather than a maintained project. But feel free to open an issue if you have an idea. Please understand if we don't respond or are slow to respond.
We use hatch to build the sam-serve package. This wheel file is built from source when building the cpu or gpu docker images. We also use hatch to publish the package to PYPI. We don't have CI CD yet, so if you are interested in contributing, increment the version in your PR and notify the maintainers @rbavery or @rub21 and we will publish the package on PR merge.
The main commands are
hatch build to make the whl file
hatch publish to publish a release
hatch can be installed with pip or pipx
pip install hatch
The model and code is licensed under the Apache 2.0 license.
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., ... Girshick, R. (2023). Segment Anything. arXiv:2304.02643. https://github.com/facebookresearch/segment-anything
The scripts/export_onnx_model.ipynb and notebooks/sam_onnx_model_example_fox.ipynb are from the original repo.