train

Bonnetal! Training (Python)

Description

• Contains Python inferfaces to all tasks in order to allow for:
  • Training of included/novel architectures on included/novel datasets.
  • Evaluate the accuracy on such datasets.
  • Generate nice paper images 😄
  • Generate inference-ready models that can be picked up by our deployment interface

---

Tasks

• Full-image classification.
• Semantic Segmentation.
• Instance Segmentation.
• Object Detection.
• CNN Keypoint/Feature Extraction
• Counting

---

Dependencies

These are the requirements for running the entirety of the training interface (all tasks). If your system is still not running after installing all these dependencies, have a look at the docker files we include. We put an effort into making sure the pipeline works there, and constantly update the installed dependencies, but if we forget one, please submit a PR to this readme to help the rest.

System dependencies

First you need to install the nvidia driver and CUDA, so have fun!

• CUDA Installation guide: link

• Then you can do the other dependencies:

  $ sudo apt-get update 
  $ sudo apt-get install -yqq  build-essential ninja-build \
    python3-dev python3-pip apt-utils curl git cmake unzip autoconf autogen \
    libtool mlocate zlib1g-dev python3-numpy python3-wheel wget \
    software-properties-common openjdk-8-jdk libpng-dev  \
    libxft-dev ffmpeg
  $ sudo updatedb

Python dependencies

• Then install the Python packages needed:

  $ sudo pip3 install -U pip
  $ sudo pip3 install numpy==1.14.0 \
    onnx==1.5.0 \
    torchvision==0.2.2.post3 \
    opencv_python==3.4.0.12 \
    scipy==0.19.1 \
    Pillow==6.0.0 \
    genpy==2016.1.3 \
    scikit_learn==0.20.3 \
    tensorflow==1.13.1 \
    PyYAML==5.1
  $ export PYTHONPATH=/usr/local/lib/python3.6/dist-packages/cv2/:$PYTHONPATH # Needed if you have ROS installed (you may want to put it in your .bashrc)

Pytorch

• From source: When building the Bonnetal C++ deployment interface, we had some problems between the provided libtorch binaries and the Dual ABI introduced in GCC 5.1, so to avoid installing twice with 2 different methods, we recommend building PyTorch from source here too (the version on my GitHub is in sync with the last one supported by our framework).

  $ git clone --recursive https://github.com/tano297/pytorch
  $ cd pytorch
  $ sudo python3 setup.py install

• (optionally) From pip: Link

TensorRT (optional)

In order to test if the models will work with TensorRT during inference with the C++ libraries, the Python interface also provides a way to infer using TensorRT with an homologous pipeline. If you want to use this interface you need to:

• Install TensorRT: Link.
• Install PyCUDA: Link, Help!.