Lottery_Ticket_Hypothesis-TensorFlow_2 VS Neural_Network_Pruning

I have already implemented "The Lottery Ticket Hypothesis" by Frankle et al. using TensorFlow 2. You can refer to the code here. Here, a binary mask (0, 1) is used for element-wise multiplication to keep the number of pruned parameters constant because by default, when you apply gradient descent algorithm, then using the weight update rule, all of the weights are updated.

Some of my recent experiments in GitHub can be referred: Lottery Ticket Hypothesis implementation and Neural Network Pruning.

I am using TensorFlow version: 2.3.0 and Python3. I am experimenting in Quantizing a pruned and trained Conv-2 CNN model. The model architecture is: conv -> conv -> max pool -> dense -> dense -> output for CIFAR-10. You can see the Jupyter-notebook here.

The code can be accessed here

here

True, removing he_normal initialization does increase the accuracy. For most of my previous experiments I have usually used the kernel initialization as mentioned in the different author's paper(s). Therefore for ResNet, I thought of using Kaiming He initialization as he is the author of the research paper. However, the default kernel initialization in TF2 is 'glorot_uniform' which leads to 60.04% val_accuracy.

Used Global, Absolute Magnitude Weight, Unstructured and Iterative pruning using ResNet-50 with Transfer Learning on CIFAR-10 dataset. Surprisingly, a sparsity of 99.078% has been achieved. The code can be referred here.

Some of my recent experiments in GitHub can be referred: Lottery Ticket Hypothesis implementation and Neural Network Pruning.

I have coded "Global, unstructured & iterative" pruning using ResNet-18 trained from scratch on CIFAR-10 dataset in PyTorch. You can refer to the code here. Let me know your comments/thoughts.

I looked into torch.nn.utils.prune module but it doesn't present an end-to-end example and the code that I came up with doesn't seem to work.