EfficientNet v0.1.7

Changelog

• Support for Tensorflow 2.0

EfficientNet v0.1.6.1

Changelog

Bugfixes

• Removes a redundant buggy reshape in the optimize_coefficients function which might break if using user-defined cost functions.

EfficientNet v0.1.6

Changelog

Improvements

• Addition of the optimize module which includes functions to compute valid candidates of hyperparameters, either via the function given in the paper or via a user-defined cost function.

EfficientNet v0.1.5

Changelog

Improvements

• Testing of models to ensure that they are built properly, load weights and make reasonable predictions.

Bugfixes

• Major bug where only a single EfficientNet model could be built (and subsequent models would throw shape mismatch errors in Add()) is now fixed.

EfficientNet v0.1.4

Changelog

New Weights

• Weights for the B4 and B5 models are now available, and have been ported.
• The code for the models have been updated to default load those weights.

EfficientNet v0.1.3

Changelog

Bugfix

• Fix weight loading issues due to incorrect parameter parsing in round_filters(...)

EfficientNet v0.1.2

Changelog

Bugfixes

• Fixed the incorrect DropConnect implementation.
• Fixed preprocess_input function due to backend not being supplied.

EfficientNet v0.1.0

Initial release

EfficientNets ImageNet Weights

Imagenet Weights

Weights for following model configurations:

• EfficientNetB0
• EfficientNetB1
• EfficientNetB2
• EfficientNetB3
• EfficientNetB4
• EfficientNetB5

Weights for B6-B7 will be ported once available from the Tensorflow repository