tensorflow-batchnorm-folding 1.0.9

Folds BN layers in tf keras models.

Batch-Normalization Folding

In this repository, we propose an implementation of the batch-normalization folding algorithm from IJCAI 2022. Batch-Normalization Folding implements the batch normalization layer by folding it into a appropriate layer. The original batch-normalization layers are removed without changing the predictive function defined by the neural network. The simpliest scenario is an application for a fully-connected layer followed by a batch-normalization layer, we get

x \mapsto \gamma \frac{Ax + b - \mu}{\sigma + \epsilon} + \beta = \gamma \frac{A}{\sigma +\epsilon} x + \frac{b - \mu}{\sigma + \epsilon} + \beta

Thus the two layers can be expressed as a single fully-connected layer at inference without any change in the predictive function.

Use

This repository is available as a pip package (use pip install tensorflow-batchnorm-folding). This implementation is compatible with tf.keras.Model instances. It was tested with the following models

• ResNet 50
• MobileNet V2
• MobileNet V3
• EfficentNet B0

To run a simple test:

from batch_normalization_folding.folder import fold_batchnormalization_layers
import tensorflow as tf
mod=tf.keras.applications.efficientnet.EfficientNetB0()
folded_model,output_str=fold_batchnormalization_layers(mod,verbose=True)

The output_str is either the ratio num_layers_folded/num_layers_not_folded or 'failed' to state a failure in the process.

Parameters

The function fold_batchnormalization_layers can be called with multiple parameters.

Parameter	Option
model	Model to fold.
folding_mechanism	ban-off uses BaN-OFF als folding mechanism (recommended) simple uses a folding mechanism that only folds neighboring layers. Default: ban-off
verbose	True prints additional information during the folding process. False Disables the printing of additional information. Default: False

Layer folding

This Python implementation supports folding into many types of layers, supported is the folding of batch-normalization parameters into:

• Dense
• Conv1D
• Conv2D
• DepthwiseConv2D

To Do

• unit test on all keras applciations models
• check package installement
• deal with Concatenate layers

Cite

@inproceedings{yvinec2022fold,
  title={To Fold or Not to Fold: a Necessary and Sufficient Condition on Batch-Normalization Layers Folding},
  author={Yvinec, Edouard and Dapogny, Arnaud and Bailly, Kevin},
  journal={IJCAI},
  year={2022}
}

Performance on Base Models

+------------------------------------+
|             ResNet 50              |
+------------------------------------+
| BN layers folded         |    53   |
| BN layers not folded     |    0    |
+------------------------------------+
|          EfficientNet B0           |
+------------------------------------+
| BN layers folded         |    49   |
| BN layers not folded     |    0    |
+------------------------------------+
|            MobileNet V2            |
+------------------------------------+
| BN layers folded         |    52   |
| BN layers not folded     |    0    |
+------------------------------------+
|            MobileNet V3            |
+------------------------------------+
| BN layers folded         |    34   |
| BN layers not folded     |    0    |
+------------------------------------+
|        Inception ResNet V2         |
+------------------------------------+
| BN layers folded         |   204   |
| BN layers not folded     |    0    |
+------------------------------------+
|            Inception V3            |
+------------------------------------+
| BN layers folded         |    94   |
| BN layers not folded     |    0    |
+------------------------------------+
|               NASNet               |
+------------------------------------+
| BN layers folded         |    28   |
| BN layers not folded     |   164   |
+------------------------------------+
|            DenseNet 121            |
+------------------------------------+
| BN layers folded         |    59   |
| BN layers not folded     |    62   |
+------------------------------------+