<< Back
Table of Contents:
InceptionResNetv2
is the second model architecture used as a CNN backbone. After a few trained models using Xception
model for this task, we wanted to test a bigger and deeper model to see if there's going to be any performance impact. The InceptionResNetv2
is only slightly slower than the Xception
but offers much deeper model (so also an ability to learn more complex and more multi-dimentional dependeces) and about twice as many parameters. This yielded our best model so far, the model_0004_inceptionresnetv2_v3
.
This model architecture has been used in the following model lines:
model_0004_inceptionresnetv2
model_0005_inceptionresnetv2
model_0006_inceptionresnetv2
model_0007_inceptionresnetv2
model_0008_irv2_data_td
model_0009_irv2_cr_tl
model_0010_irv2_tcb
and we are currently replacing it with the Regnets
(TBA).
We're using Keras implementation of the InceptionResNetv2
model without the head and with randomly initialized parameters (with an ability to use Imagenet
to initialize parameters):
Import:
from tensorflow.keras.applications import InceptionResNetv2
Usage:
cnn_backbone = InceptionResNetv2(weights="imagenet" if settings['CNN_USE_PRETRAINED_WEIGHTS'] else None, include_top=False, input_shape=model_input['shape'])
The paper describing the model: https://arxiv.org/pdf/1602.07261.pdf