Tensornets

This is a small release that resolves installation issues on Windows.

Areas of improvement

- Reproducibility improvements.
- Add missing relu in DenseNet (the top-5 accuracies increased by 2.58-3.2%).
- Revise the zero paddings of `pool1` in ResNets (the top-5 accuracies increased by 0.6-1.16%).
- Unit tests / CI improvements.
- Disable tests if only README has been changed.
- New APIs: MS COCO utils.

API changes

- Add MS COCO utils.
- Add training data loader for VOC.
- Add training codes for YOLOv2.

Breaking changes

None.

Note that the following descriptions include changes for the 0.3.3. Additionally, The 0.3.2 doesn't exist due to PyPI conflicts.

Areas of improvement

- Bug fixes.
- New APIs: image classification and object detection models.
- Unit tests / CI improvements.
- Documentation improvements.
- Float16 supporting improvements (see 14).
- Python 3 compatibility improvements.

API changes

- Add image classification models (`Darknet19`, `MobileNet35v2`, ..., `MobileNet140v2`, and `PNASNetlarge`).
- Add references object detection models (`YOLOv3VOC` and `YOLOv3COCO`).
- Add operations (`local_flatten` and `upsample`).

Breaking changes

None.

This is a small release that improves dependency importing and resolves PyPI conflicts.

Areas of improvement

- Bug fixes.
- New APIs: object detection models and PASCAL VOC utils.
- Unit tests / CI improvements.
- Documentation improvements.

API changes

- Add generic object detection models (`YOLOv2`, `TinyYOLOv2`, and `FasterRCNN`).
- Add references object detection models (`YOLOv2VOC`, `TinyYOLOv2VOC`, `FasterRCNN_ZF_VOC`, and `FasterRCNN_VGG16_VOC`).
- Add PASCAL VOC utils.
- Add operations (`stack`: `tf.stack`, `gather`: `tf.gather`, `srn`: a spatial local response normalization)
- Revise `nets.pretrained` to perform each model's `pretrained`.

Breaking changes

None.