zeyuanyin / network-pruning

Network Pruning

Network Pruning

We implentment weight pruning (unstrctured pruning) in two senerios:

• global pruning (forward + backward)
• gredient pruning (only backward)

Channel pruning is not considered (also named strctured pruning, neuron pruning, connection pruning) and the implementation is in /.old (not maintained, not recommended, just FYI).

Unofficial Implementation of DWP

This repo only contains the code for network pruning. For the attack code, there are a lot of open-source framework you can use, and the only modification is to apply the pruning methods on the surrogate model. Line 198-203 in test.py shows how to use the pruning functions.

from global_pruning import global_prune
model = global_prune(model, p_prune = args.p_prune, p_bern = args.p_bern)

from gredient_pruning import gredient_prune
model = gredient_prune(model, p_prune = args.p_prune, p_bern = args.p_bern)

Implemention Details

3 steps in weight pruning

• collect all weights (solo tensor element, one value).
• sort all weights (in ascending order) by absolute value.
• set the smallest top-k weights to zero.

Global pruning

modify the module.weight directly and the original weight is saved in module.weight_orig, which can be restored by module.weight = nn.Parameter(module.weight_orig.clone()).

Gradient pruning

with torch.autograd.graph.saved_tensors_hooks(self.pack, self.unpack) is used to register the hooks on nn.Conv2d to modify the unpacked inputs (weights) during unpacking process from computing graph in gradient computation.

Hyperparameters

• p_prune: pruning ratio, the percentage of weights to be pruning candidates.
• p_bern: bernoulli ratio, the percentage of candidates to be pruned.
• pruning ratio = p_prune * p_bern, where p_bern usually is 1.

ImageNet Test Results

Test script:

python test.py \
    --data /path/to/imagenet/ \
    --prune_type global \
    --p_prune 0.1 \
    --p_bern 1.0

Global pruning (forward + backward)

p_bern set to default 1.0, which means all pruning candidates are pruned.

network	p_prune	pruning ratio	acc@1
resnet18	0	0%	69.76%
resnet18	0.1	10%	67.99%
resnet18	0.2	20%	59.36%
resnet18	0.3	30%	49.99%
resnet18	0.4	40%	35.35%
resnet18	0.5	50%	17.35%
---------	---------	---------	---------
resnet50	0	0%	76.15%
resnet50	0.1	10%	76.14%
resnet50	0.2	20%	76.06%
resnet50	0.3	30%	75.66%
resnet50	0.4	40%	75.05%
resnet50	0.5	50%	73.26%
resnet50	0.6	60%	64.75%
resnet50	0.7	70%	25.43%

Large model (resnet50) is more robust to pruning, containing more redundant weights.

Gradient pruning (only backward)

same result as no pruning since the forward pass is not pruned. It may be useful in sparse (dynamic) backpropagation in training.

Reference

https://pytorch.org/tutorials/intermediate/pruning_tutorial.html#remove-pruning-re-parametrization

Wang, Hung-Jui, Yu-Yu Wu, and Shang-Tse Chen. "Enhancing Targeted Attack Transferability via Diversified Weight Pruning." arXiv preprint arXiv:2208.08677 (2022).

Han, Song, et al. "Learning both weights and connections for efficient neural network." Advances in neural information processing systems 28 (2015).

network-pruning's Issues

question about DWP

Can DWP be used to non-targeted attacks?