v0.2.0: Improved models, a demo app and a vision API

This release brings a big performance leap to classification models while providing easier contribution and deployment tools.

Note: pyrovision 0.2.0 requires PyTorch 1.11 and torchvision 0.12 or newer.

Highlights

🎁 Gradio Demo

In order to showcase pyrovision added value, a short Gradio demo app was added to the repo!
You can try it out live on HF Spaces over 👉 here

💻 API boilerplate

Thanks to the great FastAPI by Sebastián Ramírez (@tiangolo), we have added a small template to deploy your own vision API 😁

Check out in the folder ./api how to get yours running!

🤗 HF Hub integration

Contributions are important in open source projects and we're happy to announce that our model checkpoints are now available on HF Hub. See it as a github for data files (checkpoints for instance) 👍

You can load a model from the hub with two lines:

from pyrovision.models.utils import model_from_hf_hub
model = model_from_hf_hub("pyronear/rexnet1_0x")

If you upload your model as well, you can load it from the hub by changing "pyronear/rexnet1_0x" into "hf_user/model_repo"

🔥 OpenFire reloaded

You might have noticed that our first version of OpenFire started to have a lot of URLs failing. We've scaled up our queries to retrieve public images and checked them manually to produce a new updated version of ~7000 train images and ~800 validation images. The dataset was used to train the new model checkpoints!

The dataset is also available on HF datasets: https://huggingface.co/datasets/pyronear/openfire

Breaking changes

✝️ Deprecated modules

The following modules and features were deprecated:

• pyrovision.nn
• pyrovision.datasets.wildfire & pyrovision.datasets.video_utils
• pyrovision.models.densenet & pyrovision.models.ssresnet

Full changelog

Breaking Changes 🛠

• remove dataset creation related modules by @MateoLostanlen in #136
• refactor: Refactored package for incoming v0.2 by @frgfm in #138

New Features 🚀

• chore: add workflow to publish docker image by @MateoLostanlen in #120
• feat: Added codecarbon integration by @frgfm in #140
• feat: Added new classification checkpoints by @frgfm in #150
• feat: Added a Gradio demo by @frgfm in #151
• feat: Added minimal API template by @frgfm in #152
• feat: Updated OpenFire extract by @frgfm in #153
• feat: Added image prefetching option to Openfire by @frgfm in #158
• feat: Improved all model checkpoints and added HF hub loading function by @frgfm in #159
• ci: Added release note template file by @frgfm in #160

Bug Fixes 🐛

• increase OpenFire download threshold by @MateoLostanlen in #119
• chore: Updated conda meta by @frgfm in #126
• chore: Fixed documentation build and added CI jobs for doc sanity check by @frgfm in #130
• fix: Fixed import for torch 1.10 release by @MateoLostanlen in #129
• Update README.md by @fe51 in #134
• ignore version file by @MateoLostanlen in #137
• docs: Fixed URL in CONTRIBUTING by @frgfm in #141
• docs: Fixed author entry in pyproject by @frgfm in #144
• fix: Fixed OpenFire num_samples & image verification by @frgfm in #156
• chore: Bumped pylocron version by @frgfm in #157
• [ci: Fixed release publish job by @frgfm in https://github.com//pull/161

Improvements

• docs: add inference exemple in readme by @MateoLostanlen in #116
• docs: remove email adress by @MateoLostanlen in #122
• chore: Updated CI jobs by @frgfm in #128
• docs: Updated README & CONTRIBUTING by @frgfm in #127
• refactor: Refactored setup.py by @frgfm in #133
• chore: Updated license from AGPLv3 to Apache 2 by @frgfm in #132
• ci: Updated python version for builds by @MateoLostanlen in #135
• ci: Updated funding from OpenCollective to Github by @frgfm in #142
• style: Improved code style and documentation by @frgfm in #143
• chore: Updates version specifiers & conda recipe by @frgfm in #145
• docs: Updates documentation and README by @frgfm in #146
• ci: Updated header verification, env collection and docker job by @frgfm in #147
• feat: Updated augmentations for training and disabled codecarbon logger by @frgfm in #149
• feat: Speeds up OpenFire image verification & improves READMEs by @frgfm in #155

New Contributors

• @fe51 made their first contribution in #134

Full Changelog: v0.1.2...v0.2.0

Bug fix release

This release expands compatibility to latest releases of pytorch and torchvision.

Note: pyrovision 0.1.2 requires PyTorch 1.8 and torchvision 0.9 or newer.

Breaking changes

• Reorganized the references folder (#104)

New features

• Added Dockerfile for ubuntu environment (#108)
• Added conda build dependencies (#112)
• Added our OpenFire V2 in release attachment for open_firev2 (#121)

Bug fixes

• Fixed mobilenet imports for latest torchvision release (#115)

WildFire dataset and additional pretrained models

This release adds a video dataset and multiple pretrained models.

Note: pyrovision 0.1.1 requires PyTorch 1.2 and torchvision 0.4.0 or newer.

Highlights

Datasets

Wildfire detection visual datasets
New

• Added WildFireDataset (#46, #47, #77) and FrameExtractor (#48, #63)
• Added subsampled variant of WildFireDataset (#66)

Improvements

• Cleaned codebase (#76, #97)
• Improved constructor flexibility of OpenFire (#34)

Models

Common model architectures with pretrained parameters for wildfire detection
New

• Added pretrained SubsamplerResNet (#69)
• Added pretrained ReXNet (#86)

Improvements

• Replaced Flatten with torch version (#102)

References

Scripts for training models
New

• Added training script for WildFireDataset (#64)

Improvements

• Updated training script of WildFire (#101)

Test

Verification of the package well-being before release
Improvements

• Updated OpenFire download tolerance (#100)
• Cleaned out fixtures (#105)

Documentation

Online resources for potential users
New

• Added WildFireDataset to documentation (#54)

Improvements

• Updated CONTRIBUTING (#71)
• Updated README (#75, #87)

Fixes

• Fixed README (#51, #110)
• Fixed requirements (#106)
• Removed utils module (#107)

Others

Other tools and implementations
New

• Switched from CircleCI to Github Workflow (#68, #73)
• Added markdown URL checker (#72)
• Updated license (#78, #95)

Improvements

• Fixed header encoding (#55)
• Improved CI trigger event (#74)
• Updated setup.py (#67)
• Renamed package (#84)
• Moved env collection script outside of lib (#96, #98, #110)

Fixes

• Fixed version attribute (#57)
• Fixed codecov integration (#83)
• Fixed doc deployment (#85)

Dataset and models for binary classification

This release adds a binary classification dataset and several model architectures for wildfire detection.

Note: pyronear 0.1.0 requires PyTorch 1.2 and torchvision 0.4.0 or newer

Release 0.1.0

Datasets

Wildfire detection visual datasets
New

• Add OpenFire dataset for image classification (#11)(#15)(#18)(#21)(#32)(#39)

Models

Common model architectures with pretrained parameters for wildfire detection
New

• Add pretrained ResNet (#37)(#41)
• Add pretrained DenseNet (#41)
• Add pretrained MobileNetV2 (#41)

Documentation

Documentation for users, contributors and developers of pyronear

• Add docs for pyronear.datasets (#11)
• Add docs for pyronear.models (#37)
• Add docs for pyronear.nn (#41)
• Add docs for pyronear.utils (#7)
• Add contribution guidelines (#1)(#8)(#29)

Tests

Verifications of the package well-being before release

• Add test for pyronear.datasets (#11)(#16)(#39)
• Add test for pyronear.models (#37)(#41)
• Add test for pyronear.nn (#41)
• Add test for pyronear.utils(#17)

Reference scripts

Reference training scripts used to obtain state dictionaries of pretrained models

• Add training script with pytorch and torchvision dependencies (#23)(#30)(#32)(#41)
• Add training script with fastai dependency (#23)(#30)(#32)(#41)

Checkpoint for the first Pyroneal Model

This first model is a rexnet1_0x with a cutom head using AdaptiveConcatPool2d