All notable changes to this project will be documented in this file.
New major release. Added support for Keras 3
Added convNeXt 3D models
Padding fixes for mobilenetv2
, inceptionresnetv2
and inceptionv3
. It's needed for correct work with segmentation models.
- Added EfficientNet and EfficientNet v2 models (with converted imagenet weights)
- Fixed error in converter which skip bias for imagenet weights
include_top
parameter now False by default. It was made because imageNet weights now only available forinclude_top = False
. If useinclude_top=True
andclasses=<N>
, you must useweights=None
.- New converted imagenet weights available
- Added parameter
stride_size
to control how strides/poolings are made. By default it's equal to2
. Now it's possible to set individual stride for each stage of model. For example:
stride_size = [
(1, 2, 1),
(2, 2, 4),
(2, 2, 4),
(2, 4, 2),
(2, 2, 2),
]
Here each tuple control individual stride/pooling. While each tuple control stride for each dimension.
Strides doesn't affect model structure and you can use imagenet
weights with such modified models.
- For some models (resnet, resnext, senet, densenet, vgg16, vgg19) it's possible to increase number of blocks using
repetition
parameter. It can be useful if you need to add more poolings and layers.imagenet
weights won't work for modified models. - Minimum TF version bumped to 2.8.0