Experiments with the ideas presented in Training BatchNorm and Only BatchNorm: On the c of Random Features in CNNs by Frankle et al. In this paper, Frankle et al. explore the expressiveness of the random features in CNNs by starting with the following experimental setup:

• They first set all the layers of a CNN to trainable=False.
• Before they kickstart model training, they also set the Batch Norm layers to be trainable.

This simple experimental setup led to some pretty amazing discoveries on the expressive power of the randomly initialized layers in a CNN. So, the authors further explore the question - what if we only train the Batch Norm layers and lead this setup to a potential optimum? Their findings were pretty intriguing.

CIFAR10

ResNet20 (Thanks to the Keras Idiomatic Programmer repo)

• CIFAR10_Subset.ipynb: Runs experiments on a GPU with a subset of the CIFAR10 dataset.
• CIFAR10_Full.ipynb: Runs experiments on a GPU with the full CIFAR10 dataset.
• CIFAR10_Full_TPU.ipynb: Runs experiments on a TPU with the full CIFAR10 dataset.
• CIFAR10_Full_TPU_Different_LR_Schedules.ipynb: Runs experiments on a TPU with the full CIFAR10 dataset but with different learning rate schedules.
• All_Layers_Frozen.ipynb: As the name suggests this notebook shows what happens when all the layers of a CNN is made non-trainable.
• Varying_Batch_Sizes.ipynb: Runs experiments with varying batch sizes (only batch norm layer as trainable).
• Visualization.ipynb: Visualizes the learned convolution filters of the networks.
• Visualization_II.ipynb: Almost same as Visualization.ipynb with a bit different visualization plots.

Below is the output of the first trained convolution layer (all the layers were trained from scratch in this case)

Below is the output of the first trained convolution layer (this time only the Batch Norm layers were trained)

More results can be found here: https://app.wandb.ai/sayakpaul/training-bn-only. A more detailed report can be found here.

I trained both the variants of the networks for 75 epochs. Naturally, the one that contains only the BN layers as trainable ones would take longer to converge because of the number of parameters. But that can be used as a proxy to alleviate the problems of huge model size.

Although the notebooks are available as Colab-ready I trained all of them on a pre-configured AI Platform Notebook to make the experiments more reproducible. Thanks to the ML-GDE program program for the GCP Credits.