Keras data format converter

Generates equal keras models with the desired data format

Requirements

tensorflow >= 2.0

API

convert_channels_first_to_last(model: keras.Model, inputs_to_transpose: List[str] = None, verbose: bool = False) -> keras.Model

convert_channels_last_to_first(model: tf.keras.Model, inputs_to_transpose: List[str] = None, verbose: bool = False) \ -> tf.keras.Model

model: Keras model to convert

inputs_to_transpose: list of input names that need to be transposed due tothe data foramt changing

verbose: detailed output

Getting started

from tensorflow import keras
from keras_data_format_converter import convert_channels_last_to_first

# Load Keras model
keras_model = keras.models.load_model("my_image_model")

# Call the converter (image_input is an input that needs to be transposed, can be different for your model)
converted_model = convert_channels_last_to_first(keras_model, ["image_input"])

Supported Layers with Special handling

• Normalization layers
• Permute
• Reshape
• Concatenate
• Dot
• MultiHeadAttention
• TFOpLambda (Inserted by the Functional API construction whenever users call a supported TF symbol on KerasTensors, see here at Tensorflow repo for more info)

Unsupported Layers due to lack of data_format property

• Cropping1D
• Upsampling1D
• Zeropadding1D
• All layers in tensorflow.keras.preprocessing

How to deploy

• Create a new release version on GitHub
• Update parameters in setup.py (usually version and download_url)
• Run python setup.py sdist in root directory
• Run pip install twine
• Run twine upload dist/*

License

This software is covered by MIT License.