reproduce step

1. clone this repository and set the model files.

  git clone --recurse-submodules https://github.com/newgrit1004/EfficientSAM.git
  cd EfficientSAM
  unzip weights/efficient_sam_vits.pt.zip
  mv efficient_sam_vits.pt ./weights

2. build the docker image

# check the cuda version in Dockerfile. Modify the base image depending on your environment.
docker compose -f docker-compose-build.yml build

docker compose up -d
docker exec -it tfjs_test /bin/bash -c "cd /workspace && /bin/bash"

# inside the container
python export_to_onnx.py # generated onnx files are in "./weights" folder.
python export_onnx_to_tensorflow.py # generated tensorflow files are in "./saved_model" folder.

# if python export_onnx_to_tensorflow.py is not executed,
# install the onnx_tf manually inside the container.
docker exec -it tfjs_test /bin/bash -c "cd /workspace && /bin/bash"
root@b9fb8b01ab27:/workspace# cd onnx-tensorflow/
root@b9fb8b01ab27:/workspace/onnx-tensorflow# pip install -e .
root@b9fb8b01ab27:/workspace/onnx-tensorflow# cd ../
root@b9fb8b01ab27:/workspace# python export_to_onnx.py
root@b9fb8b01ab27:/workspace# python export_onnx_to_tensorflow.py

3. compare torch model and tensorflow model result

• See the jupyter notebook "compare_tf_torch_result.ipynb" file.
• Run the jupyter notebook in local(tensorflow 2.16.1 required)

python3 -m venv .venv
. .venv/bin/activate
pip3 install -r requirements.txt

# Then run the jupyter notebook cells in order.

4. Convert the tensorflow model into tfjs model

docker compose up -d
docker exec -it tfjs_test /bin/bash -c "cd /workspace && /bin/bash"

# inside the container
pip install tensorflow==2.16.1
tensorflowjs_converter \
      --input_format tf_saved_model \
      --output_format tfjs_graph_model \
      saved_model \
      tfjs_model # generated tfjs_model files are in "./tfjs_model" folder.

5. See the result of tensorflow.js model inference

Right click on index.html file then click "open with live server."

See the result on console.