Experimental setup

In this report, I am going to lament on my experiments with the EvoNorm layers proposed in Evolving Normalization-Activation Layers. In the paper, the authors attempt to unify the normalization layers and activation functions into a single computation graph. The authors claim -

Several of these layers enjoy the property of being independent from the batch statistics.

I used Colab to perform my experiments. The authors tested the EvoNorm layers on MobileNetV2, ResNets, MnasNet, and EfficientNets. I decided to try out some quick experiments on a Mini Inception architecture as shown in this blog post. I trained them on the CIFAR10 dataset.

👉 GitHub repo to reproduce results.

Adam + BN-ReLU + No Data Augmentation

SGD + BN-ReLU + No Data Augmentation

SGD params:

opt = tf.keras.optimizers.SGD(lr=1e-2, momentum=0.9, decay=1e-2 / EPOCHS)

SGD + BN-ReLU + Data Augmentation

EvoNorm B0 + No Data Augmentation

EvoNorm B0 + Data Augmentation

EvoNorm S0 + No Data Augmentation + Groups8

EvoNorm S0 + No Data Augmentation + Groups16

EvoNorm S0 + No Data Augmentation + Groups32

Observations on EvoNom S0 layers without data augmentation

Hyperparameter sweep on EvoNorm S0 layers without data augmentation

EvoNorm S0 + Data Augmentation + Groups8

Final remarks

As we saw for this quick experimental setup, EvoNorm layers fail to match the performance of BN-ReLU. But this should not be treated as a foregone conclusion. I encourage you to try the EvoNorm layers out in your own experiments and let me know via Twitter (@RisingSayak) what you find.

👉 Colab notebook to reproduce results.

Acknowledgement

Hanxiao Liu (first author of the paper) for helping me correct the implementation.