Giphy открыл исходный код своей модели глубокого обучения для обнаружения звезд

искусственный интеллект глубокое обучение Информация
Giphy открыл исходный код своей модели глубокого обучения для обнаружения звезд

GIPHY's Open Source Celebrity Detection Deep Learning Model and Code

About

GIPHY is proud to release our custom machine learning model that is able to discern over 2,300 celebrity faces with 98% accuracy. The model was trained to identify the most popular celebs on GIPHY, and can identify and make predictions for multiple faces across a sequence of images, like GIFs and videos.

Этот проект был разработан командой GIPHY R&D с целью создания модели глубокого обучения, которая могла бы аннотировать наш самый популярный контент так же, или, надеюсь, лучше, чем аналогичные модели, предлагаемые крупными технологическими компаниями.Мы очень гордимся наших результатов и опубликовали эту модель и обучающий код для широкой публики в надежде, что другие смогут развить нашу работу, интегрировать модель в свои собственные проекты или, возможно, извлечь уроки из нашего подхода.

Read more about the project on the GIPHY engineering blog.

You can play with the model on the demo page, и мы предоставили3D projection of all our celebrity class embeddings along with a list of all celebrities available with the model.

Thank you!

The GIPHY R&D Team

Nick Hasty @jnhasty, Ihor Kroosh @tilast, Dmitry Voitekh @dvoitekh, Dmytro Korduban @dkorduban

Try it out!

Follow the instructions in the examples directory to download the model and test it on your own GIFs and videos.

Prerequisites

  1. Python 3.6 or higher

  2. For Linux: libsm, libxext, libxrender

Training & Transfer Learning Experimentation Pipeline

Preliminary steps:

  1. Create a work directory to store results of experiments (it's not mandatory to locate this directory within the project). Example is provided here.

  2. Inside a work directory create an experiment directory. It's name must match the name of the related experiment python file (e.g. example_experiment directory for example_experiment.py file).

  3. Create a directory face_recognition inside work directory, which must contain weights for MTCNN model (3 files with names det1.npy, det2.npy, and det3.npy that can be copied from Giphy pretrained resources archive).

  4. Create a file labels.csv inside the experiment directory. It must be of the following structure (see example here):

    Label,Index
Person1,0
Person2,1
Person3,3
...
скопировать код

  5. Create directory raw_dataset inside the experiment directory. It's a dataset of uncropped images. It must be of the following structure:

    - raw_dataset
  - Person1
    image1.jpg
    image2.jpg
    image3.jpg
    ...
  + Person2
  + Person3
  ...
скопировать код

So the overall structure of the work directory is as follows:

- workdir
  - example_experiment
    - raw_dataset
      - Person1
        image1.jpg
        image2.jpg
        image3.jpg
        ...
      + Person2
      + Person3
      ...
    labels.csv
  - face_detection
    det1.npy
    det2.npy
    det3.npy
скопировать код

After that, you need to choose where you going to run training: on CPU or on GPU. According to this decision you need to change requirements_cpu.txt to requirements_gpu.txt in setup.py, or leave it as is. Also, please, mind changing tensorflow version if necessary.

Using Python 3.6 Package

  1. Create a virtual environment to localize dependencies (optional):

    virtualenv.pypa.io/en/latest/

    pip install --upgrade virtualenv
virtualenv -p python3 venv
activate
source ./venv/bin/activate
скопировать код

  2. Install the package:

    pip install -e .
скопировать код

  3. Create your experiment (see experiments directory for example).

  4. cp .env.example .env and fill the missing values in the .env file if needed.

  5. Run it from the top-level directory like:

    python experiments/example_experiment.py
скопировать код

Using Docker

  1. Install nvidia-docker2

  2. Modify your /etc/docker/daemon.json file:

    {
    "default-runtime": "nvidia",
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}
скопировать код

  3. Create your experiment (see experiments directory for examples). After that, specify name of this file in Dockerfile in CMD section.

  4. cp .env.example .env and fill the missing values, also ensure that you changed data directory and tensorboard port in docker-compose.yml file.

  5. Before running any commands you need to explicitly activate the following env variables in your shell, since they are required during container launch:

    1. TENSORBOARD_PORT

    2. WORKDIR

    3. LOCAL_WORKDIR

  6. Run via docker-compose:

    docker-compose up --build
скопировать код

    Or via plain Docker commands, for example:

    docker build -t celebrity-detection-model-train .
docker run --rm --volume $LOCAL_WORKDIR:$WORKDIR --env-file .env --runtime=nvidia --shm-size 8G -p $TENSORBOARD_PORT:$TENSORBOARD_PORT celebrity-detection-model-train
скопировать код
Категории