Moving to tf.keras

Hi. First of all thanks for your repo! Have you considered moving to tf.keras instead of using separate keras? It would make your repo compatible with new TF 2.0 out of the box while still being compatible with older versions of TF as well.

[qubvel]

Hi @bonlime @lc0 @NiklasDL @chAwater @srikrishna0130 @northanapon @Tyler-D

I have prepared a lib update to work together with keras and tf.keras and need your help! If someone have a time to test it and give a feedback it would be great (autotests have been passed)!

Some words about update:

1. API almost the same (moved get_preprocessing to root)

import segmentation_models as  sm

model = sm.Unet('resnet18')
preprocess_input = sm.get_preprocessing('resnet18')

1. Choose your framework: a) If you have Keras installed -> lib will load with Keras, if not lib will load with tf.keras b) You can setup env var SM_FRAMEWORK="keras"/SM_FRAMEWORK="tf.keras" before import segmentation_models c) You can change framework in runtime

>>> import segmentation_models as sm
Segmentation models: using `keras` framework.
>>> sm.framework() 
'keras'
>>> sm.set_framework('tf.keras')
>>> sm.framework() 
'tf.keras'

1. I have removed almost all custom objects, so now trained models (except efficientnets) can be loaded without importing segmentation moldes lib.
2. You also can change image_data_format (may be useful to convert models (#69 ))

keras.set_image_data_format('channels_first`)

Installation

Now you can install it

$ pip install -U git+https://github.com/qubvel/segmentation_models@feature/tf.keras \
                 git+https://github.com/qubvel/classification_models@feature/tf.keras \
                 git+https://github.com/qubvel/efficientnet

Roll back:

pip install -U segmentation-models==0.2.1

Waiting for your feedback, thank you!

[qubvel]

Hi @bonlime Yes, I have such plans, I think it should be done in a way of keras-applications with possibility to define your framework (keras or tf.keras). I did not still move to tf.keras by myself, so cannot guarantee this process will be fast.

Next week I am going to start adopting classification-models to make this possible. Any advices, bug reports, testing, PRs from community are appreciated.