How to make model use optimized Convolution2DTransposeBias Layer

Kevin_Bondzio · March 27, 2023, 12:43am

Hello,

some of the mediapipe models use optimized layers such as Convolution2DTransposeBias.
If I create a model myself, is there a way I can make the tflite converter to use this optimized layer?

Relevant github issue:

github.com/google/mediapipe

Why use custom op Convolution2DTransposeBias in palm_detection.tflite?

opened 06:34AM - 14 Nov 19 UTC

closed 06:11PM - 10 Dec 19 UTC

apivovarov

solution:hands

I'm not sure why mediapipe uses custom op `Convolution2DTransposeBias` in `palm_…detection.tflite`. Can it be replaced with built-In ops `TRANSPOSE_CONV ` and `ADD`? `TRANSPOSE_CONV ` op was added to TFLite in version [1.9](https://github.com/tensorflow/tensorflow/blob/r1.9/tensorflow/contrib/lite/schema/schema.fbs) I tried to create Keras model with `Conv2DTranspose` layer and I set `use_bias=True` and `bias_initializer='random_uniform'`. https://keras.io/layers/convolutional/#conv2dtranspose ``` data = keras.layers.Input(shape=(32, 32, 3)) x = keras.layers.Conv2D(filters=10, kernel_size=(3, 3), strides=(2, 2), padding='same')(data) x2 = keras.layers.Conv2DTranspose(filters=10, kernel_size=(3, 3), strides=(2, 2), padding='same', use_bias=True, bias_initializer='random_uniform')(x) ``` That Keras model can be converted to TFLite model. Keras `Conv2DTranspose` op was mapped to two built-in TFLite operators `TRANSPOSE_CONV` and `ADD` Resulting TFLite model visualization: (download html and open it): https://www.dropbox.com/s/au39pvmw3fquhm1/conv2d_transpose_bias.tflite.htm?dl=0

Thanks in advance!

chunduriv · April 27, 2023, 6:59am

@Kevin_Bondzio,

Welcome to the Tensorflow Forum!

We can use custom operators to allow conversion of an unsupported Tensorflow operator in Tensorflow Lite. If the flatbuffer file(.tflite) needs to be edited we can make use of schema and flatc command which helps us to convert a .tflite model to an editable .json.

The edited .json can be converted back to .tflite model.

Thank you!