v0.5.0

⚠️ Breaking Changes ⚠️

• ⬆️ [email protected] which drops support for TensorFlow 1.13 (#87) @lgeiger

🎉 Features

• Add names to all models (#88) @koenhelwegen

📖 Documentation

• Fix markdown syntax in DoReFa docstring (#76) @lgeiger

👷‍♂️ Internal Improvements

• Run unit tests in parallel using pytest-xdist (#83) @lgeiger
• Test against TensorFlow 1.15.0 stable (#81) @lgeiger
• Use GitHub actions to run unittests and auto-release (#80) @lgeiger

⬆️ Dependencies

• Bump larq to 0.8.1 (#90) @leonoverweel
• Bump scipy from 1.3.2 to 1.3.3 (#89) @dependabot-preview
• Bump scipy from 1.3.1 to 1.3.2 (#85) @dependabot-preview
• Upgrade black (#84) @lgeiger
• Bump pillow from 6.2.0 to 6.2.1 (#82) @dependabot-preview
• Bump pillow from 6.1.0 to 6.2.0 (#78) @dependabot-preview
• Bump tensorflow==2.0.0 (#77) @lgeiger

v0.4.2

🎉 Features

• Add DoReFa-net (#75)

🐛 Bugs

• Correctly pass top argument to decode_predictions (#73)

👷‍♀ Internal Improvements

• [ImgBot] Optimize images (#74)

DoReFa-Net

Pretrained weights for DoReFa-Net with 1 bit weights and 2 bit activations

v0.4.1

🐛 Bugs

• Bi-Real Net: Fix layer bug and point to updated pretrained weights (#69, #72)

👷 Internal Improvements

• ⬆️ [email protected] (#70)
• ⬆️ [email protected] (#71)

Bi-Real Net

Pretrained weights for corrected Bi-Real net architecture (the previous version contained one additional layer by accident)

v0.4.0

🎉 Features

• Made compatible with Zookeeper >= 0.5 (#62)
• Add ReseNetE18 and BinaryDenseNet{28,37,45} (#60)
• Add BinaryDenseNet37Dilated (#66)

🐛 Bugs

• Update weight-download url to point to correct repo. (#63, #64)

📖 Documentation

• Use ImageNet dataset version 5.0.0 (#53)
• Remove irrelevant docstrings (#58)
• Cleanup docs build (#67)

👷 Internal Improvements

• Bump scipy from 1.3.0 to 1.3.1 (#56)

Binary DenseNet

Pretrained weights for BinaryDenseNet

ResNetE18

Pretrained weights for ResNetE18

v0.3.0

🎉 Features

• Use improved preprocessing for pretrained models (#52):
  • All models are retrained from scratch and now exceed the accuracies claimed by the respective papers: https://larq.dev/models/#available-models
  • All models include plots of the training process in the docs: https://larq.dev/models/api/

📖 Documentation

• Add missing numpy import (#51)

XNOR-Net

Pretrained weights for XNOR-Net