kumasento / taylor-pruning Goto Github PK

We need to normalise the criterion values across layers (Section 2.3)

Since we're using mask to prune, it is necessary to export the model to evaluate the correctness.

When selecting channels for pruning, we don't avoid removing all channels of a single layer. The possible solution is to filter the first num_channels_to_prune by some restrictive conditions, e.g., they should not contain all channels of any layer.

We currently don't support ResNet-50 due to its:

• Batchnorm modules in between
• Residual connections

This feature would be helpful to have an overview of how this framework works.

Section 2.4 should be implemented as well.

Instead of using a fixed number of channels for each iteration, it might be better to set a value that is related to the FLOPS.