jainal09/densecap_gpu repository overview

⁠Usage:

1. Pull the image

docker pull jainal09/densecap_gpu:latest

1. Fork the repo - https://github.com/qassemoquab/stnbhwd⁠

2. Head to https://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/⁠ and find your gpu sm version.

3. Open the CMakeLists.txt file in edit mode in the forked repo of stnbhwd.

4. Edit line no. 55

LIST(APPEND CUDA_NVCC_FLAGS "-arch=sm_20")

to your sm version.

For Example for my nvidia tesla k80 gpu sm version is sm_37.

LIST(APPEND CUDA_NVCC_FLAGS "-arch=sm_37")

Commit and push the code.

1. Open the file stnbhwd-scm-1.rockspec in edit mode in forked repo of stnbhwd.

2. Change the source - url to the url of the forked repository.

For example:

source = {
   url = "git://github.com/qassemoquab/stnbhwd.git",
}

to

source = {
   url = "git://github.com/jainal09/stnbhwd.git",
}

1. In Github click on raw button in file stnbhwd-scm-1.rockspec and copy the raw url

2. Run the image and access the shell by:

docker run --gpus all -it -p 7000:7000 jainal09/densecap_gpu:latest

1. run the following command:

luarocks install [********the raw url of stnbhwd-scm-1.rockspec of your forked repo ********]

⁠Inference and testing the build

1. Download an image

wget https://raw.githubusercontent.com/jcjohnson/densecap/master/imgs/elephant.jpg

1. Run the model on this image

th run_model.lua -input_image elephant.jpg

1. If the above step ran successful without any error than you can find caption results at densecap/vis/data/results.json

Also you can run:

cd vis
python -m SimpleHTTPServer 7000

And head to localhost:7000 in browser to view the results.

⁠Bonus Rest Api Server

I have also added a rest api server for inference on the pre trained model!

1. Start the server

uvicorn main:app --reload --host 0.0.0.0 --port 7000

1. Head to localhost:7000/docs to upload an image and fetch its captions. (Only works on single image files)

2. Curl Command

curl -X POST "http://localhost:7000/uploadfile/" -H  "accept: application/json" -H  "Content-Type: multipart/form-data" -F "[email protected];type=image/jpeg"