ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression

Jian-Hao LuoJianxin WuWeiyao Lin

article2017ICCV1,888 citations

Proposes ThiNet, a filter-level pruning method that uses next-layer statistics to compress convolutional neural networks by up to sixteen times while maintaining high accuracy across standard vision benchmarks.

Listen

Deploying modern artificial intelligence models on resource-constrained devices often requires reducing model size and computational demands without significantly harming prediction accuracy. The article addresses this challenge by evaluating practical optimization techniques within ThiNet, a method designed to prune redundant components from deep neural networks.

The article set out to evaluate how different channel selection algorithms affect accuracy during model compression and to compare a fast greedy pruning approach against an optimization technique based on sparse coding.

To conduct this evaluation, the researchers tested various pruning strategies on standardized deep learning models, specifically the VGG-16 architecture, across visual benchmark datasets including Indoor-67. They compared alternative channel selection heuristics against the proposed ThiNet framework across multiple levels of computational workload reduction.

The primary finding is that the ThiNet pruning framework consistently preserves higher classification accuracy across various compression rates compared to alternative heuristic methods, such as random selection, weight sums, and activity-based pruning. Additionally, when testing how to solve the underlying mathematical channel selection problem, the simpler greedy approach achieved slightly superior performance (69.80% top-1 accuracy) compared to the more complex sparse coding solution (69.34% top-1 accuracy). These results indicate strong generalization capabilities across diverse datasets.

These findings suggest that organizations aiming to deploy vision models on edge hardware can achieve substantial reductions in computational cost and energy usage without sacrificing operational performance. Furthermore, adopting the simpler greedy optimization strategy provides better accuracy while avoiding the algorithmic overhead and complexity associated with sparse coding solvers.

Technical leaders and engineering teams should prioritize the greedy channel selection strategy within the ThiNet framework when streamlining deep learning pipelines for production deployment. Future development should focus on exploring more efficient solvers for the optimization problem to determine if higher compression rates can be achieved without accuracy loss.

Confidence in these specific model comparisons is high, though readers should exercise caution when extrapolating results to entirely different network architectures or non-visual tasks, as the provided evidence focuses primarily on standard image classification benchmarks.

Cover for ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression

Abstract

We propose an efficient and unified framework, namely ThiNet, to simultaneously accelerate and compress CNN models in both training and inference stages. We focus on the filter level pruning, i.e., the whole filter would be discarded if it is less important. Our method does not change the original network structure, thus it can be perfectly supported by any off-the-shelf deep learning libraries. We formally establish filter pruning as an optimization problem, and reveal that we need to prune filters based on statistics information computed from its next layer, not the current layer, which differentiates ThiNet from existing methods. Experimental results demonstrate the effectiveness of this strategy, which has advanced the state-of-the-art. We also show the performance of ThiNet on ILSVRC-12 benchmark. ThiNet achieves 3.31×\times FLOPs reduction and 16.63×\times compression on VGG-16, with only 0.52%\% top-5 accuracy drop. Similar experiments with ResNet-50 reveal that even for a compact network, ThiNet can also reduce more than half of the parameters and FLOPs, at the cost of roughly 1%\% top-5 accuracy drop. Moreover, the original VGG-16 model can be further pruned into a very small model with only 5.05MB model size, preserving AlexNet level accuracy but showing much stronger generalization ability.

Table of Contents

  • 1 Introduction
  • 2 Related work
  • 3 ThiNet
  • 3.1 Framework of ThiNet
  • 3.2 Data-driven channel selection
  • 3.2.1 Collecting training examples
  • 3.2.2 A greedy algorithm for channel selection
  • 3.2.3 Minimize the reconstruction error
  • 3.3 Pruning strategy
  • 4 Experiments
  • 4.1 Different filter selection criteria
  • 4.2 VGG-16 on ImageNet
  • 4.3 ResNet-50 on ImageNet
  • 4.4 Domain adaptation ability of the pruned model
  • 5 Conclusion
  • References

Knowls

  1. Knowl 1 — Sparse Coding Formulation for Channel Selection in ThiNet

    model/method

    Channel selection in the ThiNet filter pruning framework can be formulated as a sparse coding problem to reconstruct target layer activations from a subset of input channels. Let X∈Rm×CX \in \mathbb{R}^{m \times C} denote mm sampled training features across CC input channels, and y∈Rm×1y \in \mathbb{R}^{m \times 1} represent the corresponding target feature map output. The channel selection objective is formulated as an ℓ1\ell_1-minimization problem:

    w^=arg⁡min⁡w∈RC×1∥w∥1s.t.∥Xw−y∥2≤ϵ\hat{w} = \arg\min_{w \in \mathbb{R}^{C \times 1}} \|w\|_1 \quad \text{s.t.} \quad \|X w - y\|_2 \le \epsilon

    where ϵ\epsilon is an error tolerance parameter and w∈RC×1w \in \mathbb{R}^{C \times 1} is a sparse channel weighting vector. This problem is solved using the homotopy method. To retain exactly C×rC \times r channels (where r∈(0,1)r \in (0, 1) is the channel retention ratio), a binary search over the regularization hyperparameter is conducted until w^\hat{w} has exactly C×rC \times r non-zero entries. Channels corresponding to zero entries are pruned, and the non-zero coefficients in w^\hat{w} are used as scaling factors for the preserved channels.

  2. Knowl 2 — Performance Comparison Between Greedy Selection and Sparse Coding in ThiNet

    data/table

    When pruning VGG-16 using the ThiNet-Conv framework, selecting channels via a greedy search heuristic outperforms selecting channels via ℓ1\ell_1-norm sparse coding (solved via the homotopy method with binary search):

    Strategy Top-1 Accuracy Top-5 Accuracy
    ThiNet-Conv with sparse coding 69.34% 89.27%
    ThiNet-Conv with greedy solution 69.80% 89.53%

    The greedy selection strategy achieves a 0.46%0.46\% higher Top-1 accuracy and a 0.26%0.26\% higher Top-5 accuracy than the sparse coding approach on ImageNet classification.

  3. Knowl 3 — Channel Selection Performance Comparison on Indoor-67

    empirical result

    Evaluating channel selection methods on a VGG-16 model with Global Average Pooling (VGG-16-GAP) fine-tuned on the MIT Indoor-67 dataset demonstrates the following behaviors across FLOPs reduction rates ranging from 0% to nearly 100%:

    1. Heuristic pruning criteria—including weight magnitude pruning (Weight sum), Average Percentage of Zeros (APoZ), and random selection—exhibit nearly identical accuracy degradation curves across all FLOPs reduction rates.
    2. The ThiNet data-driven reconstruction method consistently and significantly outperforms all heuristic baselines at every FLOPs reduction level (for instance, sustaining ≈65–70%\approx 65\text{--}70\% Top-1 accuracy up to 60%60\% FLOPs reduction, where heuristic methods experience steeper accuracy drops).
    3. Retaining the optimal reconstruction scaling vector w^\hat{w} (ThiNet) yields a small but consistent accuracy gain over omitting the scaling factors (ThiNet w/o w^\hat{w}).

Coverage note — Appendix A was deliberately omitted because it contains only a two-line algebraic derivation demonstrating the equivalence of two objective equations, which is excluded under the rule barring proof and derivation steps.

Citation

MLA
Luo, J.-H., et al. “ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression”. 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 5068–76, https://doi.org/10.1109/ICCV.2017.541.
APA
Luo, J.-H., Wu, J., & Lin, W. (2017). ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression. 2017 IEEE International Conference on Computer Vision (ICCV), 5068–5076. https://doi.org/10.1109/ICCV.2017.541
Chicago
Luo, J.-H., J. Wu, and W. Lin. 2017. “ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression”. 2017 IEEE International Conference on Computer Vision (ICCV), 5068–76. https://doi.org/10.1109/ICCV.2017.541.
Harvard
Luo, J.-H., Wu, J. and Lin, W. (2017) “ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression”, 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, pp. 5068–5076. Available at: https://doi.org/10.1109/ICCV.2017.541.
Vancouver
1. Luo J-H, Wu J, Lin W (2017) ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression. In: 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, pp 5068–5076

BibTeX

@inproceedings{Luo_2017, title={ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression}, url={http://dx.doi.org/10.1109/ICCV.2017.541}, DOI={10.1109/iccv.2017.541}, booktitle={2017 IEEE International Conference on Computer Vision (ICCV)}, publisher={IEEE}, author={Luo, Jian-Hao and Wu, Jianxin and Lin, Weiyao}, year={2017}, month=Oct, pages={5068–5076} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE