Back

Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning

Sixing YuArya MazaheriAli Jannesari

article2022ICML54 citations

Proposes a multi-stage graph embedding framework that captures hierarchical neural network topologies to automatically learn optimal, FLOP-constrained pruning policies via reinforcement learning across diverse architectures.

Listen

Deploying modern deep neural networks to resource-constrained edge devices, such as smartphones, autonomous vehicles, and robotics, is challenging due to the heavy computing power and memory demands of these models. While network pruning compresses models by eliminating unnecessary channels and parameters, conventional methods typically rely on manual heuristics or local layer-by-layer analyses. These traditional practices overlook global structural patterns and architectural changes that occur during pruning. Furthermore, existing automated reinforcement learning approaches often encounter reward imbalances that drive models toward minimal, ineffective compression ratios.

The article demonstrates an automated, topology-aware network compression framework that models neural architectures as hierarchical graphs and identifies optimal pruning strategies using reinforcement learning. The approach aims to drastically reduce computational operations—measured in floating-point operations—while preserving or improving model accuracy.

To achieve this, the article introduces a multi-stage graph neural network that decomposes complex models into simplified hierarchical graphs based on recurring structural patterns. It pairs this embedding method with a reinforcement learning agent guided by proximal policy optimization within a continuous compression environment. The environment systematically prunes model layers until a targeted computation budget is reached, using the final accuracy as the primary learning signal. The framework was evaluated across multiple standard image classification benchmarks, including CIFAR-10, CIFAR-100, and ImageNet, spanning diverse architectures such as standard residual networks, VGG-16, and mobile-optimized architectures like MobileNet and ShuffleNet.

The evaluation yielded several key results. First, the proposed approach achieved competitive or superior compression and accuracy compared to existing state-of-the-art methods; for example, reducing computation by 52% in ResNet-110 and 54% in ResNet-56 resulted in net accuracy gains of 0.63% and 0.10%, respectively, by mitigating over-fitting. Second, heavily compressed models demonstrated rapid recoverability, regaining most pruned accuracy within a single fine-tuning epoch and stabilizing in fewer than 50 epochs. Third, the learned structural policies proved transferable, enabling graph encoders pre-trained on one network to prune related architectures 1.12 times faster per round while matching native search accuracy. Fourth, practical deployment tests on graphics processing units showed notable resource reductions, such as cutting VGG-16 memory usage by 141 megabytes and delivering a 1.38-times inference speedup.

These findings indicate that automated, graph-based compression can significantly reduce model latency and memory footprints, lowering operational hardware costs and accelerating edge deployment timelines. The results challenge standard human heuristics, proving that deeper network layers are not uniformly critical and that middle or residual layers often contain substantial redundancy. Consequently, engineering teams can replace manual trial-and-error tuning with scalable, automated pruning workflows.

Organizations deploying computer vision models to resource-limited devices should adopt automated, graph-aware pruning to compress deep architectures. Engineering teams should also reuse pre-trained graph policies across related network families to minimize search computation. Before broad production rollout, teams must conduct on-device latency and throughput testing on target edge hardware, as high-end server GPUs do not fully showcase the latency gains achieved on specialized mobile processors.

Confidence in these findings is high across standard vision architectures and datasets. However, a key limitation is that hardware evaluation was conducted primarily on a single desktop GPU rather than embedded edge platforms. Readers should exercise caution when extrapolating specific inference speedups to custom edge hardware, and additional validation is warranted before applying the technique beyond convolutional image models.

Iowa State UniversityTechnische Universität Darmstadtgraph embeddingsgraph neural networksnetwork topologytopology-aware network pruningmulti-stage graph embeddingreinforcement learning
arXiv: 2102.03214yusx-swapp/GNN-RL-Model-Compression

No sufficiently relevant recommendations were found.

Cover for Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning

Abstract

Model compression is an essential technique for deploying deep neural networks (DNNs) on power and memory-constrained resources. However, existing model-compression methods often rely on human expertise and focus on parameters’ local importance, ignoring the rich topology information within DNNs. In this paper, we propose a novel multi-stage graph embedding technique based on graph neural networks (GNNs) to identify DNN topologies and use reinforcement learning (RL) to find a suitable compression policy. We performed resource-constrained (i.e., FLOPs) channel pruning and compared our approach with state-of-the-art model compression methods. We evaluated our method on various models from typical to mobile-friendly networks, such as ResNet family, VGG-16, MobileNet-v1/v2, and ShuffleNet. Results show that our method can achieve higher compression ratios with a minimal fine-tuning cost yet yields outstanding and competitive performance. The code is open-sourced at https://github.com/yusx-swapp/GNN-RL-Model-Compression.

Topics

graph embeddingsgraph neural networksnetwork topologyreinforcement learningtopologyresource-constrainednetwork architecturecompression algorithmstuning methodhigh compression ratio

Keywords

topology-aware network pruningmulti-stage graph embeddingreinforcement learningneural network pruningchannel pruningmulti-stage GNNhierarchical graph representationproximal policy optimizationneural architecture searchactor-critic network

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 3. Approach
  • 3.1. Hierarchical Graph Representation
  • 3.2. Multi-stage GNN
  • 3.3. Network Pruning Using Reinforcement Learning and m-GNN
  • 4. Experimental Results
  • 4.1. Implementation Details
  • 4.2. Comparisons with State-of-the-art
  • 4.3. Ablation Study
  • 4.4. Inference Acceleration and Memory Saving
  • 5. Conclusion
  • Acknowledgement
  • References

Knowls

  1. Knowl 1 — DNN-Graph turns constrained pruning into a topology-updating search episode

    model/method

    GNN-RL prunes a supplied deep neural network (DNN) iteratively under a user-specified model-size constraint, using FLOPs as the constraint in the experiments. At each step, the DNN-Graph environment checks the current model size. If the constraint is not yet met, it converts the current compressed network into a hierarchical computational graph, the graph encoder uses that graph to produce pruning ratios, and a channel pruner applies those ratios to the network. Because pruning changes the network topology, the environment rebuilds the graph after every pruning step. Once the model meets the size constraint, the episode ends and the pruned model’s validation accuracy supplies the reinforcement-learning reward. This stopping rule lets the agent search toward the target size without putting model size into the reward or requiring a separate heuristic to adjust pruning ratios.

  2. Knowl 2 — DNNs are represented as hierarchical graphs of operations and motifs

    definition

    A DNN is represented as an ll-level hierarchical computational graph. At level tt, a graph GitG_i^t has nodes VitV_i^t representing hidden states, directed edges EitE_i^t representing operations, and a collection Gt−1G^{t-1} of lower-level computational graphs that define the operations available at that level. The bottom level contains selected primitive operations; repeated subgraphs built from those primitives can be extracted and treated as compound operations, or motifs, at the next level. Thus an edge at a higher level can stand for an entire lower-level computation rather than a single primitive operation. For convolutional-network pruning, the authors use a two-level hierarchy: the first level represents convolution operations and motifs, and the second represents the network assembled from those motifs.

  3. Knowl 3 — m-GNN embeds motifs in stages using edge features and graph pooling

    model/method

    The multi-stage GNN (m-GNN) embeds a hierarchical DNN graph from the bottom level upward. At each level, a graph encoder maps each graph to an embedding, and embeddings of lower-level graphs serve as edge features at the next level. Graphs at the same hierarchy level share GNN parameters, so a recurring motif need only be embedded once. The embedding of the top-level graph is the representation of the DNN.

    Unlike standard GCN message passing, m-GNN incorporates the edge embedding through an element-wise product. For a node ii with neighbor set NiN_i, a message from neighbor jj along an edge corresponding to lower-level graph kk is combined as

    hit+1=∑j∈Ni1ciWt(hjt⊙ekt−1),h_i^{t+1}=\sum_{j\in N_i}\frac{1}{c_i}W^t\left(h_j^t\odot e_k^{t-1}\right),

    where hjth_j^t and hit+1h_i^{t+1} are node feature vectors, ekt−1e_k^{t-1} is the embedding of the lower-level graph represented by that edge, WtW^t is a learned weight matrix, cic_i is a normalization coefficient, and ⊙\odot denotes element-wise multiplication. The authors select this operation so that edge type modulates how much neighbor information is passed; they also note that element-wise multiplication is associative across hierarchy stages.

    To form a graph-level embedding, m-GNN uses learnable pooling weights over node embeddings. Pruned nodes are represented by zero vectors, preserving the feature-array size when pruning changes the number of nodes. GNN parameters are shared among graphs at a level, while each computational graph has its own pooling parameters.

  4. Knowl 4 — The pruning agent predicts continuous layer ratios and is rewarded only for accuracy

    model/method

    The GNN-RL agent receives the current hierarchical graph as its state. A graph encoder produces a graph embedding, and a multilayer perceptron (MLP) maps that embedding to one pruning ratio per hidden layer. The action is a vector A=[a1,…,aN]TA=[a_1,\ldots,a_N]^T, where NN is the number of pruning layers and each ratio aia_i is constrained to [0,1)[0,1) by clipping. The reward is the negative Top-1 validation error of the compressed DNN; model size is excluded from the reward because the DNN-Graph episode terminates automatically when the size constraint is met.

    The policy is trained with proximal policy optimization (PPO), with m-GNN included in the actor-critic network and updated end-to-end. The PPO update uses a clipped probability-ratio objective; the clipping parameter used in the experiments is 0.20.2. In the reported policy comparison, PPO achieved better performance and faster convergence than DDPG, and the authors report convergence within half a GPU-hour for a CIFAR-10 experiment on an NVIDIA V100.

  5. Knowl 5 — GNN-RL achieves competitive accuracy at substantial FLOPs reductions across benchmarks

    data/table

    The following results report the paper’s Top-1 accuracy comparisons on CIFAR-10/100 and ImageNet. FLOPs reduction is the reported reduction from the unpruned model; the accuracy gap is the reported percentage-point difference from the original model. The selected comparison rows show that GNN-RL often obtains a larger reduction or higher accuracy than the listed reference method, but not in every comparison: for example, EagleEye reports higher ResNet-56 accuracy at a smaller reduction, and the ImageNet MobileNet-v2 baselines report higher accuracy at smaller reductions.

    Could not parse LaTeX table

    The paper reports that GNN-RL’s pruned ResNet-110 and ResNet-56 models exceed their original test accuracy by 0.630.63 and 0.100.10 percentage points, respectively. The authors suggest overfitting as a possible explanation for these gains, noting that the original models reached 100% training accuracy.

  6. Knowl 6 — Experimental configuration combines two-level graphs, PPO training, and dataset-specific fine-tuning

    experimental setup

    The experiments use two-level graph representations with primitive operation types for k×kk\times k convolution, depth-wise convolution, point-wise convolution, and k×kk\times k max-pooling. Graph node and edge features are initialized from a standard distribution with feature size 20. The actor and critic each use a graph encoder with hidden and embedding sizes of 50 and an MLP with hidden size 200; the critic MLP has one output. The RL agent is trained with Adam at learning rate 3×10−43\times10^{-4} and β=(0.9,0.999)\beta=(0.9,0.999), updating its policy every 100 search episodes for 20 epochs per update. The discount factor is 0.990.99, action standard deviation is 0.50.5, and PPO clipping parameter is 0.20.2.

    The evaluation datasets are CIFAR-10, CIFAR-100, and ImageNet. The reward is computed using 5,000 sampled test images as a validation set for CIFAR-10/100 and 10,000 sampled test images for ImageNet. For ResNets, residual-connected layers share a pruning ratio to avoid feature-map mismatches. For MobileNet-v1/v2, pruning is restricted to linear expansion layers and point-wise filters within blocks, with shared ratios for residual-connected MobileNet-v2 expansion layers. ShuffleNet blocks are pruned together, and each block’s expansion/output layer is preserved.

    After pruning, the authors fine-tune for 100 epochs on CIFAR-10 and 150 epochs on CIFAR-100 and ImageNet, using SGD with learning rate 5×10−35\times10^{-3}, batch size 512, weight decay 5×10−45\times10^{-4}, and cosine learning-rate decay. For ImageNet MobileNet-v1/v2, they additionally fine-tune for one epoch before reward evaluation so the agent receives a more informative accuracy signal.

  7. Knowl 7 — Hierarchical representation greatly reduces graph size while improving node-classification accuracy

    empirical result

    On the ResNet-110 computational graph, the authors compared a plain graph processed by a GCN with a hierarchical graph processed by m-GNN. The hierarchical representation reduced the graph from 394,412 nodes and 581,332 edges to 12,460 nodes and 20,221 edges. In the associated node-classification experiment, accuracy increased from 83.50% with the plain graph and GCN to 84.20% with the hierarchical graph and m-GNN. The comparison supports the authors’ claim that motif-based hierarchy can reduce graph-processing scale while retaining useful structural information.

  8. Knowl 8 — A graph encoder learned on ResNet-56 transfers to ResNet-44 pruning

    empirical result

    For a topology-transfer experiment, the authors first trained GNN-RL on ResNet-56 and then reused its graph encoder when pruning ResNet-44. During transfer, the graph encoder’s gradients were disabled and only the MLP mapping graph embeddings to pruning actions was updated. The transferred agent showed a learning curve similar to direct pruning search on ResNet-44 and made each search round 1.12 times faster. It produced a ResNet-44 with 51% FLOPs reduction and 93.23% accuracy, compared with 93.10% accuracy for the original model. The result supports transfer of the learned policy between networks with similar topology.

  9. Knowl 9 — Pruned models reduce measured GPU latency and memory, with gains varying by architecture

    data/table

    The authors measured inference on an NVIDIA GTX 1080 Ti using PyTorch 1.7.1 and batch size 32. The FLOPs column below gives the remaining FLOPs as a percentage of the original model. Every listed pruned model used less measured GPU memory and had lower latency than its unpruned counterpart, although the latency gains were modest for some mobile-oriented models. For VGG-16, pruning to 20% of original FLOPs reduced memory from 528 MB to 387 MB and latency from 0.11 ms to 0.08 ms; the paper notes that pruning convolutional layers also reduces feature-map memory, despite the model’s dense layer contributing relatively little FLOPs. The authors caution that measurements on their resource-rich GPU may not reveal the full advantages of mobile-friendly networks on mobile hardware.

    Could not parse LaTeX table

  10. Knowl 10 — ResNet pruning allocations reveal nonuniform redundancy and an accuracy-dependent FLOPs limit

    empirical result

    For ResNet-110, the authors report that test accuracy was highest at a 66% FLOPs reduction; reductions beyond 66% caused test accuracy to drop sharply. The learned layer-wise pruning policies were nonuniform: GNN-RL assigned relatively high pruning to residual-connection layers and also pruned substantially in middle layers 45–65 and deep layers 90–109 of ResNet-110. This allocation differs from handcrafted policies that assume deeper layers should generally be pruned less, and the authors interpret it as evidence that such fixed rules do not generalize reliably across network structures. The corresponding reported policies include a 59% FLOPs-reduced ResNet-110 and a 49% FLOPs-reduced ResNet-56.

Coverage note — The detailed fine-tuning recovery curves and the full PPO-versus-DDPG learning curves are not reproduced; the paper reports their qualitative findings, but the plotted curves do not provide a complete set of numerical measurements suitable for standalone knowls.

References

  1. 1.Anwar, S., Hwang, K., and Sung, W. Structured pruning of deep convolutional neural networks. Proc. of the J. Emerg. Technol. Comput. Syst., 13(3), February 2017. ISSN 1550-4832.
  2. 2.Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J. Distributed optimization and statistical learning via the alternating direction method of multipliers. Found. Trends Mach. Learn., 3(1):1–122, January 2011. ISSN 1935-8237.
  3. 3.Chatzianastasis, M., Dasoulas, G., Siolas, G., and Vazirgiannis, M. Graph-based neural architecture search with operation embeddings. In Proc. of the IEEE/CVF International Conference on Computer Vision (CVPR), pp. 393–402, 2021.
  4. 4.Chen, J., Chen, S., and Pan, S. J. Storage efficient and dynamic flexible runtime channel pruning via deep reinforcement learning. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M. F., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 14747–14758. Curran Associates, Inc., 2020.
  5. 5.Chen, T., Ji, B., Ding, T., Fang, B., Wang, G., Zhu, Z., Liang, L., Shi, Y., Yi, S., and Tu, X. Only train once: A one-shot neural network training and pruning framework. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, volume 34, pp. 19637–19651. Curran Associates, Inc., 2021.
  6. 6.Chin, T.-W., Ding, R., Zhang, C., and Marculescu, D. Towards efficient model compression via learned global ranking. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
  7. 7.Dudziak, L., Chau, T., Abdelfattah, M. S., Lee, R., Kim, H., and Lane, N. D. BRP-NAS: Prediction-based NAS using gcns, 2021.
  8. 8.Gao, S., Huang, F., Cai, W., and Huang, H. Network Pruning via Performance Maximization. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9270–9280, June 2021.
  9. 9.Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W., and Keutzer, K. A survey of quantization methods for efficient neural network inference, 2021.
  10. 10.Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In Proc. of International conference on machine learning, pp. 1263–1272. PMLR, 2017.
  11. 11.Guo, S., Wang, Y., Li, Q., and Yan, J. Dmcp: Differentiable markov channel pruning for neural networks. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020.
  12. 12.Guo, Y., Yao, A., and Chen, Y. Dynamic network surgery for efficient dnns. In Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., and Garnett, R. (eds.), Proc. of the Advances in Neural Information Processing Systems, volume 29, pp. 1379–1387. Curran Associates, Inc., 2016.
  13. 13.Guo, Y., Zheng, Y., Tan, M., Chen, Q., Chen, J., Zhao, P., and Huang, J. NAT: Neural architecture transformer for accurate and compact architectures. In Proc. of the Advances in Neural Information Processing Systems, volume 32, pp. 737–748. Curran Associates, Inc., 2019.
  14. 14.Han, S., Mao, H., and Dally, W. J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In Proc. of International Conference on Learning Representations (ICLR), 2016.
  15. 15.He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proc. of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.
  16. 16.He, Y., Zhang, X., and Sun, J. Channel pruning for accelerating very deep neural networks. In Proc. of the IEEE International Conference on Computer Vision, pp. 1389–1397, 2017.
  17. 17.He, Y., Kang, G., Dong, X., Fu, Y., and Yang, Y. Soft filter pruning for accelerating deep convolutional neural networks. In International Joint Conference on Artificial Intelligence (IJCAI), pp. 2234–2240, 2018a.
  18. 18.He, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., and Han, S. AMC: AutoML for model compression and acceleration on mobile devices. In Proc. of the European Conference on Computer Vision (ECCV), pp. 784–800, 2018b.
  19. 19.He, Y., Liu, P., Wang, Z., Hu, Z., and Yang, Y. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
  20. 20.Hinton, G., Vinyals, O., and Dean, J. Distilling the knowledge in a neural network. In Proc. of NIPS Deep Learning and Representation Learning Workshop, 2015.
  21. 21.Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al. Searching for mobilenetv3. In Proc. of the IEEE/CVF International Conference on Computer Vision, pp. 1314–1324, 2019.
  22. 22.Huang, G., Liu, S., Van der Maaten, L., and Weinberger, K. Q. Condensenet: An efficient densenet using learned group convolutions. In Proc. of the IEEE conference on computer vision and pattern recognition, pp. 2752–2761, 2018.
  23. 23.Jian-Hao Luo, J. W. Autopruner: An end-to-end trainable filter pruning method for efficient deep model inference. Pattern Recognition, 2020.
  24. 24.Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In Proc. of the International Conference on Learning Representations (ICLR), 2017.
  25. 25.Krizhevsky, A. and Hinton, G. Learning multiple layers of features from tiny images. 2009.
  26. 26.Lai, C.-I. J., Zhang, Y., Liu, A. H., Chang, S., Liao, Y.-L., Chuang, Y.-S., Qian, K., Khurana, S., Cox, D., and Glass, J. Parp: Prune, adjust and re-prune for self-supervised speech recognition. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, volume 34, pp. 21256–21272. Curran Associates, Inc., 2021.
  27. 27.Li, B., Wu, B., Su, J., and Wang, G. Eagleeye: Fast subnet evaluation for efficient neural network pruning. In Vedaldi, A., Bischof, H., Brox, T., and Frahm, J.-M. (eds.), Proc. of the Computer Vision – ECCV 2020, pp. 639–654, Cham, 2020a. Springer International Publishing. ISBN 978-3-030-58536-5.
  28. 28.Li, H., Kadav, A., Durdanovic, I., Samet, H., and Graf, H. P. Pruning filters for efficient convnets, 2016.
  29. 29.Li, Y., Gu, S., Mayer, C., Gool, L. V., and Timofte, R. Group sparsity: The hinge between filter pruning and decomposition for network compression. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8018–8027, 2020b.
  30. 30.Li, Y., Gu, S., Zhang, K., Van Gool, L., and Timofte, R. DHP: Differentiable meta pruning via hypernetworks, 2020c.
  31. 31.Li, Y., Hao, C., Li, P., Xiong, J., and Chen, D. Generic neural architecture search via regression. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, volume 34, pp. 20476–20490. Curran Associates, Inc., 2021a.
  32. 32.Li, Z., Yuan, G., Niu, W., Zhao, P., Li, Y., Cai, Y., Shen, X., Zhan, Z., Kong, Z., Jin, Q., et al. Npas: A compiler-aware framework of unified network pruning and architecture search for beyond real-time mobile acceleration. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14255–14266, 2021b.
  33. 33.Liben-Nowell, D. and Kleinberg, J. The link-prediction problem for social networks. Journal of the American Society for Information Science and Technology, 58(7):1019–1031, 2007.
  34. 34.Liebenwein, L., Baykal, C., Lang, H., Feldman, D., and Rus, D. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020.
  35. 35.Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. Continuous control with deep reinforcement learning. In Proc. of the ICLR (Poster), 2016.
  36. 36.Lin, J., Rao, Y., Lu, J., and Zhou, J. Runtime neural pruning. In Proc. of the Advances in Neural Information Processing Systems, pp. 2181–2191, 2017.
  37. 37.Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., and Shao, L. Hrank: Filter pruning using high-rank feature map. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1529–1538, 2020.
  38. 38.Liu, N., Ma, X., Xu, Z., Wang, Y., Tang, J., and Ye, J. AutoCompress: An automatic dnn structured pruning framework for ultra-high compression rates. In Proc. of the Artificial Intelligence Conference (AAAI), pp. 4876–4883, 2020.
  39. 39.Liu, Y., Liu, L., Lin, C., Dong, Z., and Wang, W. Learnable Motion Coherence for Correspondence Pruning. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3237–3246, June 2021.
  40. 40.Liu, Z., Mu, H., Zhang, X., Guo, Z., Yang, X., Cheng, K.-T., and Sun, J. Metapruning: Meta learning for automatic neural network channel pruning. In Proc. of the IEEE International Conference on Computer Vision, pp. 3296–3305, 2019.
  41. 41.Ma, N., Zhang, X., Zheng, H.-T., and Sun, J. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proc. of the European conference on computer vision (ECCV), pp. 116–131, 2018.
  42. 42.Mehta, S., Hajishirzi, H., and Rastegari, M. Dicenet: Dimension-wise convolutions for efficient networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.
  43. 43.Ning, X., Zhao, T., Li, W., Lei, P., Wang, Y., and Yang, H. DSA: More efficient budgeted pruning via differentiable sparsity allocation. In Proc. of 16th European Computer Vision Conference, pp. 592–607. Springer, 2020a.
  44. 44.Ning, X., Zheng, Y., Zhao, T., Wang, Y., and Yang, H. A generic graph-based neural architecture encoding scheme for predictor-based nas. In Proc. of European Conference on Computer Vision, pp. 189–204. Springer, 2020b.
  45. 45.Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. doi: 10.1007/s11263-015-0816-y.
  46. 46.Schlichtkrull, M., Kipf, T. N., Bloem, P., van den Berg, R., Titov, I., and Welling, M. Modeling relational data with graph convolutional networks. In Gangemi, A., Navigli, R., Vidal, M.-E., Hitzler, P., Troncy, R., Hollink, L., Tordai, A., and Alam, M. (eds.), The Semantic Web, pp. 593–607, Cham, 2018. Springer International Publishing. ISBN 978-3-319-93417-4.
  47. 47.Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms, 2017.
  48. 48.Shi, H., Pi, R., Xu, H., Li, Z., Kwok, J. T., and Zhang, T. Bridging the gap between sample-based and one-shot neural architecture search with BONAS, 2019.
  49. 49.Tan, M. and Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. of the International Conference on Machine Learning, pp. 6105–6114. PMLR, 2019.
  50. 50.Tang, Y., Wang, Y., Xu, Y., Tao, D., XU, C., Xu, C., and Xu, C. Scop: Scientific control for reliable neural network pruning. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M. F., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 10936–10947. Curran Associates, Inc., 2020.
  51. 51.Wang, H., Zhang, Q., Wang, Y., and Hu, R. Structured probabilistic pruning for deep convolutional neural network acceleration. British Machine Vision Conference, 2017.
  52. 52.Wang, T., Wang, K., Cai, H., Lin, J., Liu, Z., Wang, H., Lin, Y., and Han, S. Apq: Joint search for network architecture, pruning and quantization policy. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2075–2084, 2020.
  53. 53.Wang, Z., Li, C., and Wang, X. Convolutional neural network pruning with structural redundancy reduction. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14913–14922, June 2021.
  54. 54.Yang, T.-J., Howard, A., Chen, B., Zhang, X., Go, A., Sandler, M., Sze, V., and Adam, H. Netadapt: Platform-aware neural network adaptation for mobile applications. In The European Conference on Computer Vision (ECCV), September 2018.
  55. 55.Yao, L., Pi, R., Xu, H., Zhang, W., Li, Z., and Zhang, T. Joint-detnas: Upgrade your detector with nas, pruning and dynamic distillation. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10170–10179, 2021.
  56. 56.Ye, M., Gong, C., Nie, L., Zhou, D., Klivans, A., and Liu, Q. Good subnetworks provably exist: Pruning via greedy forward selection. In III, H. D. and Singh, A. (eds.), Proc. of the 37th International Conference on Machine Learning, volume 119 of Proc. of Machine Learning Research, pp. 10820–10830. PMLR, 13–18 Jul 2020.
  57. 57.Ying, R., You, J., Morris, C., Ren, X., Hamilton, W. L., and Leskovec, J. Hierarchical graph representation learning with differentiable pooling. In Proc. of the 32nd International Conference on Neural Information Processing Systems, NIPS’18, pp. 4805–4815, Red Hook, NY, USA, 2018. Curran Associates Inc.
  58. 58.Yu, J. and Huang, T. Autoslim: Towards one-shot architecture search for channel numbers, 2019.
  59. 59.Yu, S., Mazaheri, A., and Jannesari, A. Auto graph encoder-decoder for neural network pruning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6362–6372, October 2021.
  60. 60.Zhang, T., Ye, S., Zhang, K., Tang, J., Wen, W., Fardad, M., and Wang, Y. A systematic DNN weight pruning framework using alternating direction method of multipliers. ECCV, 2018a.
  61. 61.Zhang, X., Zhou, X., Lin, M., and Sun, J. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proc. of the IEEE conference on computer vision and pattern recognition, pp. 6848–6856, 2018b.
  62. 62.Zhuang, T., Zhang, Z., Huang, Y., Zeng, X., Shuang, K., and Li, X. Neuron-level structured pruning using polarization regularizer. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M. F., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 9865–9877. Curran Associates, Inc., 2020.

Citation

MLA
Yu, S., et al. “Topology-Aware Network Pruning Using Multi-stage Graph Embedding and Reinforcement Learning”. International Conference on Machine Learning, vol. 162, 2022, pp. 25656–67, https://proceedings.mlr.press/v162/yu22e.html.
APA
Yu, S., Mazaheri, A., & Jannesari, A. (2022). Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning. International Conference on Machine Learning, 162, 25656–25667. https://proceedings.mlr.press/v162/yu22e.html
Chicago
Yu, S., A. Mazaheri, and A. Jannesari. 2022. “Topology-Aware Network Pruning Using Multi-stage Graph Embedding and Reinforcement Learning”. International Conference on Machine Learning 162: 25656–67. https://proceedings.mlr.press/v162/yu22e.html.
Harvard
Yu, S., Mazaheri, A. and Jannesari, A. (2022) “Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning”, International Conference on Machine Learning. PMLR, pp. 25656–25667. Available at: https://proceedings.mlr.press/v162/yu22e.html.
Vancouver
1. Yu S, Mazaheri A, Jannesari A (2022) Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning. In: International Conference on Machine Learning. PMLR, pp 25656–25667

BibTeX

@InProceedings{pmlr-v162-yu22e,
  title = 	 {Topology-Aware Network Pruning using Multi-stage Graph Embedding and Reinforcement Learning},
  author =       {Yu, Sixing and Mazaheri, Arya and Jannesari, Ali},
  booktitle = 	 {Proceedings of the 39th International Conference on Machine Learning},
  pages = 	 {25656--25667},
  year = 	 {2022},
  editor = 	 {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
  volume = 	 {162},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {17--23 Jul},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v162/yu22e/yu22e.pdf},
  url = 	 {https://proceedings.mlr.press/v162/yu22e.html},
  abstract = 	 {Model compression is an essential technique for deploying deep neural networks (DNNs) on power and memory-constrained resources. However, existing model-compression methods often rely on human expertise and focus on parameters’ local importance, ignoring the rich topology information within DNNs. In this paper, we propose a novel multi-stage graph embedding technique based on graph neural networks (GNNs) to identify DNN topologies and use reinforcement learning (RL) to find a suitable compression policy. We performed resource-constrained (i.e., FLOPs) channel pruning and compared our approach with state-of-the-art model compression methods. We evaluated our method on various models from typical to mobile-friendly networks, such as ResNet family, VGG-16, MobileNet-v1/v2, and ShuffleNet. Results show that our method can achieve higher compression ratios with a minimal fine-tuning cost yet yields outstanding and competitive performance.}
}
Metadata:DOI registry

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: https://creativecommons.org/licenses/by/4.0/