convert_keras_model() does not work as expected for BinaryDenseNet37 Dilated and XNORNet

[ZhanqiuHu]

I tried using python 3.6 + LCE 0.6.2 and python 3.7/3.8 + LCE 0.7.0 to run the following code, and the tflite file generated has unexpected sizes:

For python 3.6 + LCE 0.6.2:
XNOR tflite: 88.9 MB
BinaryDenseNet37 tflite: 25.6 MB

For python 3.7/3.8 + LCE 0.7.0:
XNOR tflite: 235.2 MB
BinaryDenseNet37 tflite: 5.4 MB (this looks normal)

Do you know what is causing this and what will be a solution? Thanks a lot!

import tensorflow as tf
import larq_zoo as lqz
import larq as lq


input_tensor = tf.keras.layers.Input(shape=(224, 224, 3))
# model = lqz.literature.BinaryDenseNet37Dilated(input_tensor=input_tensor, weights="imagenet")
model = lqz.literature.XNORNet(input_tensor=input_tensor, weights="imagenet")

lq.models.summary(model, print_fn=None, include_macs=True)

import os
path = os.path.join(os.getcwd(), './tflite_models')
if not os.path.exists(path):
    os.makedirs(path)
with open(os.path.join(path,name+'.tflite'), 'wb') as flatbuffer_file:
    flatbuffer_bytes = lce.convert_keras_model(model)
    flatbuffer_file.write(flatbuffer_bytes)

[lgeiger]

It looks like dilated convolutions weren't properly converted in LCE 0.6.2 which was based on the TensorFlow 2.5 converter, this seems to fixed in LCE 0.7.

With respect to XNOR-Net:
It looks like there was a regression in 0.7 or in the underlying TensorFlow converter which leads to incorrect fusion of the xnor weights quantizer in one of the layers:

I would need to take a closer look at this conversion issue next week, for now I'd recommend sticking with 0.6.2 for converting XNORNet.