Keras-retinanet

Changes since last release:

- Fix VGG imagenet download.
- Add numpy as dependency.
- Convert generators to Keras `Sequence`s.
- Float16 support.
- Expose learning rate parameter.
- Add validation loss as optional step.

Changes since last release

- Evaluation uses progressbar
- Correct initialization of weights for classification submodel
- Fix issue with evaluating when there are gaps in classes
- Add configuration (currently only for anchor settings)
- Refactor how annotation are generated in the generators
- Use CPU to convert model
- Update to keras 2.2.4
- Add NCHW support

Credits to
adreo00
borakrc
yecharlie
ddowling
enricoliscio
hgaiser
baek-jinoo
de-vri-es
penguinmenac3
Morten Back Nielsen
relh
vcarpani

Changes since last release

- Optimizations for generators
- Improved documentation.
- OID Challenge 2018 support.
- Keras version bumped to 2.2.0.
- Add option for class specific filtering (NMS).
- Add flake8 for code testing.
- Merged COCO and non-COCO evaluation scripts.
- Correct image preprocessing for MobileNet and DenseNet.

Credits to:
apacha
hgaiser
de-vri-es
lvaleriu
cclauss
HolyGuacamole
leonardvandriel
PhilippMarquardt
vcarpani

Changes since last release

- Implement DenseNet, VGG backbones.
- Add option to freeze backbone layers.
- Add logging of evaluation to tensorboard.
- Add pretty colors for 80 classes.
- Fix batch_size > 1 issues.
- Refactor model outputs (should hopefully stay like this now).
- Simplified training by splitting into "training model" and "inference model".
- Add structure for backbone specific functions (such as `load_model`).
- Encode regression as x1/y1/x2/y2 offsets (increases mAP to 0.350, previously 0.345).
- Use `nearest` upsampling method.

Credits to:
vidosits
cgratie
DiegoAgher
eduramiba
GuillaumeErhard
Muhannes
hgaiser
iver56
jjiunlin
srslynow
de-vri-es
Ori226
pedroconceicao
pderian
rodrigo2019
lvaleriu
yhenon

Changes since last release

- Corrected FPN architecture as per paper.
- Set default image size to minimum of 800px.
- Change NMS to perform per-class NMS.
- Small correction for bbox transform.
- Add OID data generator.
- Change default NMS threshold to 0.5.
- Add MobileNet backbone.
- Add tensorboard callback.
- Add tool for debugging datasets.
- Improve speed of data augmentation methods.
- Add ability to resume training.
- Add evaluation tool for custom datasets (only computes mAP at the moment).
- Add `skip_mismatch` to weights loading, allows transfer learning from pretrained COCO model.

Credits to:
awilliamson
hgaiser
de-vri-es
mxvs
wassname
mkocabas
lvaleriu