By Paul-Emmanuel SOTIR [email protected]
This project is under Open Source MIT License, see ./LICENSE or more details.

WIP: This Project is still under active development, and at an early stage of development

DeepCV is a Kedro PyTorch project which aims to simplify the implementation of simple vision tasks. DeepCV allows you to easily define vision processing pipelines by leveraging recent DeepLearning algorithms along with the usual OpenCV tools.

Some of DeepCV's main features are:

• Custom lossless Image and video compression codec using learned arithmetic encoder policies to minimize image and video sizes (which means faster streaming, faster reading from disk, lower storage size)
• Feature extraction and matching using lightweight CNNs, improving the reliability and reproducibility of image processing pipelines compared to implementations that rely on classical feature extractors such as SIFT, ORB, ...
• The deepcv.meta python module contains various utilities to make it easier to define models, train models with ignite, search hyperparameters with NNI, follow and visualize training experiments with MLFlow/kedro/TensorboardX..., preprocess and augment data, schedule learning rate(s) with One Cycle policy, perform meta-learning thanks to various tools like HyperparametersEmbedding and as well as meta deep-learning abstractions (e.g. Experiment, DatasetStats, Task, Hyperparameters, HyperparameterSpace) stored for each experiments in a 'metadataset', ...
• deepcv.meta.base_module.DeepcvModule A base class for easier DeepCV model definition: model sub-modules (NN blocks or layers) can be defined in a simple and generic manner in ./conf/base/parameters.yml and a shared image embedding block of a few convolution layers can allow learning to be transferred between any DeepCV image models by sharing, training, forking and/or merging these shared weights.
• [./conf/base/parameters.yml] can also specify data augmentation and preprocessing recipes.
• ...

git clone ...
...
# You can then run tests scripts to verify successfull installation of DeepCV:
python tests/test1.py

TODO: ...

import deepcv

def main():
    pass

if __name__ == "__main__":
    main()

TODO: more in depth description

Alternatively, if you need documentation from a specific branch or updated with your contributions, you can build sphinx documentation by following these instructions:

git clone ...

TODO: bla bla

Interesting third party projects which could be integrated into this project

Eventually create submodule(s) for the following github projects under third_party directory (see https://git-scm.com/book/fr/v2/Utilitaires-Git-Sous-modules + script to update submodule to latest release commit?):

• ImageNetV2
• Detectron2
• Apex (https://github.com/NVIDIA/apex): needs to be installed manually along with PyProf (optional)
• Use pytorch-OpCounter for pytorch model memory/FLOPs profiling: https://github.com/Lyken17/pytorch-OpCounter
• SinGAN pytorch implementation: https://github.com/tamarott/SinGAN
• https://github.com/MegviiDetection/video_analyst
• https://github.com/rwightman/pytorch-image-models

External dependencies TODO: Remove this section kedro/mlflow/ignite/pytorch/tensorboard//NNI/Apex/Scikit-learn/Numpy/pandas/Jupyter/.../Python/Conda/CUDA + DeepCV with ffmpeg(+ faster/hardware-accelerated video h264/VP9/AV1 decompression lib?) + DeepCV docker image with or without GPU acceleration + keep in mind portability (e.g. to android, ARM, jetson, ...)

Kedro project generated using Kedro 0.15.7

Take a look at the documentation to get started.

In order to get the best out of the template:

• Please don't remove any lines from the .gitignore file provided
• Make sure your results can be reproduced by following a data engineering convention, e.g. the one we suggest here
• Don't commit any data to your repository
• Don't commit any credentials or local configuration to your repository
• Keep all credentials or local configuration in conf/local/

Dependencies should be declared in src/requirements.txt for pip installation and src/environment.yml for conda installation.

kedro install

You can run your Kedro project with:

kedro run

Have a look at the file src/tests/test_run.py for instructions on how to write your tests. You can run your tests with the following command:

kedro test

To configure the coverage threshold, please have a look at the file .coveragerc.

In order to use notebooks in your Kedro project, you need to install Jupyter:

pip install jupyter

For using Jupyter Lab, you need to install it:

pip install jupyterlab

After installing Jupyter, you can start a local notebook server:

kedro jupyter notebook

You can also start Jupyter Lab:

kedro jupyter lab

And if you want to run an IPython session:

kedro ipython

Running Jupyter or IPython this way provides the following variables in scope: proj_dir, proj_name, conf, io, parameters and startup_error.

Once you are happy with a notebook, you may want to move your code over into the Kedro project structure for the next stage in your development. This is done through a mixture of cell tagging and Kedro CLI commands.

By adding the node tag to a cell and running the command below, the cell's source code will be copied over to a Python file within src/<package_name>/nodes/.

kedro jupyter convert <filepath_to_my_notebook>

Note: The name of the Python file matches the name of the original notebook.

Alternatively, you may want to transform all your notebooks in one go. To this end, you can run the following command to convert all notebook files found in the project root directory and under any of its sub-folders.

kedro jupyter convert --all

In order to automatically strip out all output cell contents before committing to git, you can run kedro activate-nbstripout. This will add a hook in .git/config which will run nbstripout before anything is committed to git.

Note: Your output cells will be left intact locally.

In order to package the project's Python code in .egg and / or a .wheel file, you can run:

kedro package

After running that, you can find the two packages in src/dist/.

To build API docs for your code using Sphinx, run:

kedro build-docs

See your documentation by opening docs/build/html/index.html.

To generate or update the dependency requirements for your project, run:

kedro build-reqs

This will copy the contents of src/requirements.txt into a new file src/requirements.in which will be used as the source for pip-compile. You can see the output of the resolution by opening src/requirements.txt.

After this, if you'd like to update your project requirements, please update src/requirements.in and re-run kedro build-reqs.