Pith. sign in

REVIEW 4 major objections 5 minor 68 references

QuarterMap: Efficient Post-Training Token Pruning for Visual State Space Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A training-free pruning scheme that discards three of every four pixels before VMamba's four-directional scan and restores them with nearest-neighbor upsampling delivers up to 1.11x higher throughput with under a percentage point of top-1…

desk verdict A clean, training-free token pruning module for VMamba that gives modest speedups, but the paper never proves it beats simply using a smaller VMamba checkpoint. read the letter →

arxiv 2507.09514 v1 pith:VCP2PEUL submitted 2025-07-13 cs.CV cs.AI

Tien-Yu Chi , Hung-Yueh Chiang , Diana Marculescu , Kai-Chiang Wu This is my paper
classification cs.CVcs.AI
keywords visualstatespacemodelsVMambaactivationpruningtokenpost-trainingefficiencynearest-neighborupsamplingthroughputoptimizationmedicalimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a large share of VMamba's runtime cost is spent scanning spatially redundant activations, and that a training-free pruning scheme can remove that redundancy. QuarterMap keeps one pixel out of every four in the spatial grid before the four-directional cross-scan, runs the selective scan on the quarter-size map, and restores the original resolution with nearest-neighbor upsampling after cross-merge. On ImageNet-1K this yields up to a 1.11x throughput gain on VMamba-B with a 0.86% top-1 accuracy drop, and on the medical model MedMamba it raises throughput by 1.21x with no accuracy loss. Because the method changes no weights and needs no retraining, it is positioned as a plug-and-play deployment-time optimization for SSM-based vision backbones.

What carries the argument

The load-bearing mechanism is the pairing of a deterministic spatial downsampling before the scan with nearest-neighbor upsampling after it. QuarterMap is a function that, for a chosen interval m and retention count n, retains every n-th element in both spatial dimensions, reducing an H x W feature map to about n/m of its pixels; the paper's chosen setting is m=2, n=1, giving one quarter of the pixels. Pruning happens before the cross-scan, so the savings come from shorter input sequences to the SSM recurrence and to the selective-scan linear projections, while nearest-neighbor upsampling after cross-merge restores the original shape with minimal overhead (about 0.2 ms versus 9.7 ms for token merging). The method is applied only to selected blocks, controlled by an interval k=3, skipping early layers that encode fundamental features.

What would settle it

Run QuarterMap with the same pruning budget but with random pixel selection instead of the regular grid and compare ImageNet-1K accuracy; if random pruning matches the grid's accuracy, the spatial-redundancy assumption is not doing the work. Alternatively, evaluate on a high-frequency, texture-heavy subset of ImageNet where adjacent-pixel similarity fails; a top-1 drop well beyond the reported 0.86% would mark the boundary of the method.

Watch

Extended reading notes

Core claim

The central discovery is that VMamba's four-directional scan carries enough spatial redundancy that three of every four spatial positions can be discarded before the scan and synthesized afterward by nearest-neighbor copying, with accuracy nearly intact. QuarterMap applies this as a deterministic grid pruning: with interval m=2 and retention n=1, it keeps every other element in each spatial dimension (one quarter of the pixels), applies the cross-scan, selective scan, and cross-merge to the reduced map, then upsamples back to the original size. The paper reports a 1.11x speedup on VMamba-B on ImageNet-1K with a 0.86% top-1 drop, better accuracy-throughput trade-offs than token merging, and consistent gains on VMamba-T/S, on ADE20K segmentation, and on MedMamba across four MedMNIST tasks. The paper also shows the method is specific to VMamba-like 4D-scan SSMs: applying the same pruning to CNNs or 1D-scan SSMs such as ViM causes substantially larger accuracy losses.

Load-bearing premise

The method assumes, without proof, that adjacent spatial positions in VMamba's feature maps carry similar information, so discarding three of every four pixels before the scan and replacing them with nearest-neighbor copies after it costs little accuracy.

Editorial extensions

If this is right

  • QuarterMap achieves up to 1.11x throughput on VMamba-B with a 0.86% top-1 accuracy drop on ImageNet-1K, and up to 1.16x with a 1.30% drop when applied every two blocks.
  • On MedMamba-T, QuarterMap raises throughput by 1.21x across four MedMNIST tasks with no change in classification accuracy, including class-wise accuracy on BloodMNIST.
  • On ADE20K semantic segmentation with VMamba-UperNet, QuarterMap incurs a 0.76% drop in all-pixel accuracy and a 1.75% drop in mIoU for the base model.
  • Because the method changes no weights and needs no training data, it can be applied to already-deployed models and combined with other techniques such as quantization.
  • Token merging methods like ToMe are suboptimal for VMamba because merge and unmerge operations add more overhead than the scan savings; QuarterMap avoids this.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same pruning pattern could plausibly extend to other SSM vision backbones that use multi-directional scanning, such as PlainMamba's 4D variant, though the paper's results there show larger drops than on VMamba.
  • The fixed grid pattern suggests a testable extension: adapt the pruning interval per layer based on a cheap redundancy estimate, which could push the Pareto frontier beyond the uniform k=3 schedule.
  • The nearest-neighbor reconstruction assumption implies that QuarterMap's accuracy should degrade most on high-frequency detail; a dataset or evaluation emphasizing fine texture could reveal where the method's bound breaks.
  • Because QuarterMap is orthogonal to weight quantization, combining both could compound deployment gains; the paper notes this as a direction but does not measure it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes QuarterMap, a post-training activation-pruning method for VMamba-style visual state space models. Before the four-directional cross-scan, QuarterMap retains one out of every two spatial positions along each dimension (m=2, n=1) in selected blocks (every k=3 blocks), reducing the scanned sequence to one quarter of its original size; after cross-merge it restores full resolution with nearest-neighbor upsampling. The authors report ImageNet-1K top-1 accuracy and throughput for VMamba-T/S/B with and without QuarterMap, ADE20K segmentation with UperNet, MedMNIST classification with MedMamba, and an architecture-family comparison; they also ablate block-selection interval, pruning interval, retained-element count, and upsampling method.

Significance. If its claims held, QuarterMap would be a simple, training-free, architecture-aware plug-in for fixed VMamba deployments, and the paper has the virtue of reporting a broad set of experiments, including a direct comparison with ToMe's merge overhead and ablations over its main hyperparameters. However, the reported evidence does not currently establish the central deployment advantage: on the paper's own numbers, a pruned VMamba-B operating point is Pareto-dominated by the unmodified VMamba-S checkpoint, and the transferability argument in Appendix D.1 is contradicted by the ADE20K results. The accuracy-drop headline in the abstract also does not match the tables. The idea is interesting and easy to test, but the claims need to be re-benchmarked against the relevant baselines and revised.

major comments (4)
  1. [Section 3.1, Table 1; Appendix D.1, Table 4] The central efficiency claim is not compared against the most relevant post-training alternative, namely selecting an existing smaller pretrained VMamba checkpoint. In Table 4, QM-VMamba-B with k=3 reaches 83.02% top-1 at 654 img/s, while unmodified VMamba-S reaches 83.64% at 811 img/s; QM-VMamba-B with k=2 reaches 82.58% at 682 img/s, while unmodified VMamba-T reaches 82.60% at 1548 img/s. These are strict Pareto dominations. Appendix D.1 acknowledges that VMamba-S has slightly higher accuracy and throughput but justifies this with an untested transferability argument, and Table 5 shows the opposite: QM-Base gets 49.21 mIoU on ADE20K, below the unmodified Small baseline's 50.6 mIoU. The paper should either add a direct comparison with standard smaller checkpoints or provide a downstream task where the pruned larger model preserves an advantage; otherwise the deployment-time benefit reduces to 'a pruned large model is faster than the same large model,' which is a much weaker claim.
  2. [Abstract and Section 3.1] The headline claim of 'less than 0.9% accuracy drop' is not supported by the reported results. Table 4 shows that with k=3, the top-1 drops are 1.10% for VMamba-T and 1.22% for VMamba-S, and only VMamba-B achieves 0.86%. The abstract should be qualified to name the specific configuration (VMamba-B) or should state the actual range of drops across variants. Similarly, 'up to 11% speedup' is inconsistent with the k=2 rows in Table 4, which report 1.14x to 1.16x speedups; the text should either report the full range or explain why only the k=3 point is being highlighted.
  3. [Appendix E] The hyperparameters (k=3, m=2, n=1) are chosen from ablations on the same ImageNet-1K validation set used to report final accuracy, and only single-run measurements are reported without error bars or repeated trials. This makes the 'optimal trade-off' and 'consistently improves' statements difficult to evaluate; part of the reported accuracy at the selected configuration is the result of selection on the evaluation set. Please either use a separate tuning split, report mean and standard deviation over multiple runs, or add a sentence acknowledging this limitation and show that nearby hyperparameter choices give similar behavior.
  4. [Section 2 and Section 3.4] The method rests on the assumption that adjacent spatial positions in VMamba's feature maps carry similar information, so that discarding three of four pixels before the scan and replacing them with nearest-neighbor copies after cross-merge costs little accuracy. The only supporting evidence is qualitative attention-map and effective-receptive-field visualization (Figure 4). The paper should provide a quantitative test of this assumption, for example by measuring the correlation of adjacent activations across layers, or by an ablation that compares nearest-neighbor upsampling with a more expressive upsampling at the same computational budget. Such an experiment would also clarify whether the accuracy loss in Table 5 comes from the pruning or from the upsampling stage.
minor comments (5)
  1. [Throughout] There are several typos and formatting errors: 'BoodMNIST' in Section 3.2, 'accuarcy' in Section 2, 'overheadd' in Tables 1 and 4, 'Restuls' in Appendix D.4, 'lineard' in Section B.2, and 'and and' in Section E.2.
  2. [Section 1] The sentence 'We prove that conventional methods, such as token merging, are suboptimal for VMamba' overstates what is shown; the paper gives an empirical latency comparison, not a proof. Please rephrase as 'we show empirically'.
  3. [Tables 1 and 4] Table 1 and Table 4 contain exactly the same VMamba-B rows; please consolidate to avoid redundant reporting and use one table in the main text and a reference in the appendix.
  4. [Abstract] The phrase 'less than 0.9% accuracy drop' should name the exact model and configuration, because Table 4 shows larger drops on VMamba-S and VMamba-T.
  5. [Section E.2 and Figure 6] The notation 'n continuous pixels' is unclear; since the paper uses m=2, n=1, please clarify that the retained positions form a regular sub-sampling pattern rather than a contiguous block when n > 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: QuarterMap is an empirical post-training transform whose reported accuracy and throughput are measured against external benchmarks, not derived from its own assumptions by construction.

full rationale

The paper makes no mathematical derivation claim: QuarterMap is an explicitly empirical post-training transform (uniform spatial subsampling before the SS2D scan, nearest-neighbor upsampling after cross-merge). The claimed accuracy drops and throughput gains are measured on external benchmarks (ImageNet-1K, ADE20K, MedMNIST), not derived from the method's assumptions by construction. The key assumption, that adjacent spatial positions carry similar information, is stated as a hypothesis in Section 2 and is tested empirically rather than used to force results. Hyperparameters k=3, m=2, n=1 are selected through ablations in Appendix E; this is standard tuning on the validation set and does not make the reported operating point a fitted quantity masquerading as a prediction. No load-bearing self-citations appear: the cited VMamba, Mamba, and MedMamba works are external architecture baselines, and no uniqueness theorem is imported from the authors' prior work. Appendix D.1's admission that VMamba-S (83.64%, 811 img/s) dominates QM-VMamba-B (83.02%, 654 img/s) is a legitimate efficiency-claim concern, but it is an omitted-baseline/comparison issue, not circularity: the paper's accuracy-throughput numbers are not equivalent to its inputs by construction. Overall, the derivation chain is self-contained and empirically grounded, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The only parameters the method introduces are the block selection interval, pruning interval, and retained element count, all selected by ablation on the same benchmarks. The key domain assumption is spatial redundancy in VMamba; no new theoretical entities or mechanisms are introduced.

free parameters (3)
  • Block selection interval k = 3
    Chosen based on Pareto analysis of accuracy-throughput trade-off on ImageNet-1K (Appendix E); controls how frequently pruning is applied.
  • Pruning interval m = 2
    Selected via ablation on ImageNet-1K (Appendix E.2); m=2 gives the claimed accuracy-latency trade-off.
  • Retained elements n = 1
    Selected via ablation (Appendix E.2); retaining one pixel per interval is used.
assumptions (3)
  • domain assumption Adjacent spatial positions in VMamba feature maps carry similar information, so 2x subsampling before scan and nearest-neighbor upsampling after merge preserves accuracy.
    Stated in Section 2 (Pruning Stage and Upsampling Stage) as the justification for the method; no formal proof is given.
  • domain assumption The VMamba SS2D cross-scan aggregates neighboring information across four directions, making uniform subsampling before scanning lossless for task-relevant content.
    Implicit in Section 2 and supported only by ERF/attention visualizations, not by a derivation.
  • domain assumption Throughput measurements on a single A100 GPU with batch size 128 reflect deployment performance.
    Appendix C.3 gives the hardware but no repeated runs or measurement protocol are specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QuarterMap: Efficient Post-Training Token Pruning for Visual State Space Models." pith.science (2026). https://pith.science/paper/VCP2PEUL

@misc{pith2026250709514,
  author       = {Pith},
  title        = {Pith review of: QuarterMap: Efficient Post-Training Token Pruning for Visual State Space Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCP2PEUL}},
  note         = {Machine review of arXiv:2507.09514}
}
read the original abstract

State space models (SSMs) reduce the quadratic complexity of transformers by leveraging linear recurrence. Recently, VMamba has emerged as a strong SSM-based vision backbone, yet remains bottlenecked by spatial redundancy in its four-directional scan. We propose QuarterMap, a post-training activation pruning method that removes redundant spatial activations before scanning and restores dimensions via nearest-neighbor upsampling. Our method improves throughput without retraining. On ImageNet-1K, QuarterMap achieves up to 11% speedup on VMamba with less than 0.9% accuracy drop, and yields similar gains on ADE20K segmentation. Beyond VMamba, we validate QuarterMap on MedMamba, a domain-specific model that shares the same four-directional scanning structure, where it consistently improves throughput while preserving accuracy across multiple medical imaging tasks. Compared to token merging methods like ToMe, QuarterMap is tailored for SSMs and avoids costly merge-unmerge operations. Our method offers a plug-and-play tool for deployment-time efficiency without compromising transferability.

Figures

Figures reproduced from arXiv: 2507.09514 by the authors.

Figure 1
Figure 1. The accuracy-throughput trade-off when applying Quar￾terMap and Token Merging (ToMe). Our method demonstrates that QuarterMap not only increases throughput but also maintains comparable accuracy. In contrast, ToMe experiences a drop in throughput due to the overhead of merge and unmerge operations. shick, 2015; Redmon et al., 2016). However, CNNs exhibit limitations in capturing long-range dependencies. Vision Trans… view at source ↗
Figure 2
Figure 2. Illustration of the VMamba model architecture (left) and the proposed QuarterMap applied to the SS2D mechanism (right). The top flow shows the original cross-scan and cross-merge operations, while the bottom applies QuarterMap, pruning activations before scan and restoring spatial dimensions via nearest-neighbor upsampling. This reduces spatial redundancy and improves runtime efficiency without retraining. (Liu et a… view at source ↗
Figure 3
Figure 3. Illustration of the pruning stage, where m represents the interval size and n indicates the number of elements retained in the spatial dimensions. Pruning Stage Within each selected block, QuarterMap per￾forms a downsampling operation on the spatial dimensions of x before cross-scan. For a specified interval m, the func￾tion T retains every n elements in both the H and W dimen￾sions, as depicted in [PITH_FULL_IMAGE… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of (a) attention maps and (b) effective re￾ceptive fields before and after applying QuarterMap. The visu￾alizations highlight the differences introduced by QuarterMap, demonstrating its selective pruning of redundant information while preserving the model’s …
Figure 5
Figure 5. Figure 5: Pareto-optimal analysis from different block selection interval k. D.4. Extended Restuls on Other Architectures [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Ablation studies on feature map pruning in QuarterMap on ImageNet-1K classification. drop, highlighting the critical role of early blocks in encoding fundamental low-level features essential for downstream tasks. This finding aligns with prior computer vision studies e…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 37 canonical work pages

  1. [1]

    E., and Cohan, A

    Beltagy, I., Peters, M. E., and Cohan, A. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020

    arXiv 2004

  2. [2]

    and Hoffman, J

    Bolya, D. and Hoffman, J. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 4599--4603, 2023

    work page 2023

  3. [3]

    Token merging: Your vit but faster, 2023

    Bolya, D., Fu, C.-Y., Dai, X., Zhang, P., Feichtenhofer, C., and Hoffman, J. Token merging: Your vit but faster, 2023. URL https://arxiv.org/abs/2210.09461

    arXiv 2023

  4. [4]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 248--255. Ieee, 2009

    2009

  5. [5]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021

    2021

  7. [7]

    v., Williams, C., Winn, J., and Zisserman, A

    Everingham, M., Gool, L. v., Williams, C., Winn, J., and Zisserman, A. The PASCAL Visual Object Classes Challenge 2008 , 2008. URL http://host.robots.ox.ac.uk/pascal/VOC/voc2008

    work page 2008

  8. [8]

    and Carbin, M

    Frankle, J. and Carbin, M. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018

    arXiv 2018

Show all 68 references
  1. [9]

    F., Powell, J

    Franklin, G. F., Powell, J. D., Emami-Naeini, A., and Powell, J. D. Feedback control of dynamic systems, volume 4. Prentice hall Upper Saddle River, 2002

    2002

  2. [10]

    Y., Dao, T., Saab, K

    Fu, D. Y., Dao, T., Saab, K. K., Thomas, A. W., Rudra, A., and Re, C. Hungry hungry hippos: Towards language modeling with state space models. In ICLR, 2022

    2022

  3. [11]

    Fast r-cnn

    Girshick, R. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pp.\ 1440--1448, 2015

    2015

  4. [12]

    Rich feature hierarchies for accurate object detection and semantic segmentation

    Girshick, R., Donahue, J., Darrell, T., and Malik, J. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 580--587, 2014

    2014

  5. [13]

    and Dao, T

    Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  6. [14]

    Hippo: Recurrent memory with optimal polynomial projections

    Gu, A., Dao, T., Ermon, S., Rudra, A., and R \'e , C. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 33: 0 1474--1487, 2020

    2020

  7. [15]

    Efficiently modeling long sequences with structured state spaces

    Gu, A., Goel, K., and Re, C. Efficiently modeling long sequences with structured state spaces. In ICLR, 2021

    2021

  8. [16]

    Han, S., Mao, H., and Dally, W. J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015

    2015 arXiv

  9. [17]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  10. [18]

    Channel pruning for accelerating very deep neural networks

    He, Y., Zhang, X., and Sun, J. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision, pp.\ 1389--1397, 2017

    2017

  11. [19]

    Soft filter pruning for accelerating deep convolutional neural networks

    He, Y., Kang, G., Dong, X., Fu, Y., and Yang, Y. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018

    2018 arXiv

  12. [20]

    T., Baumann, S

    Hu, V. T., Baumann, S. A., Gui, M., Grebenkova, O., Ma, P., Fischer, J., and Ommer, B. Zigma: Zigzag mamba diffusion model. arXiv preprint arXiv:2403.13802, 2024

    2024 arXiv

  13. [21]

    Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. NeurIPS, pp.\ 1106--1114, 2012

    2012

  14. [22]

    Kálmán, R. E. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82 0 (1): 0 35--45, 1960

    1960

  15. [23]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In ICML, pp.\ 1207--1216, 2000

    2000

  16. [24]

    Optimal Brain Damage

    LeCun, Y., Denker, J., and Solla, S. Optimal Brain Damage . In Advances in Neural Information Processing Systems , volume 2. Morgan-Kaufmann, 1989. URL https://proceedings.neurips.cc/paper_files/paper/1989/hash/6c9882bbac1c7093bd25041881277658-Abstract.html

    1989

  17. [25]

    Li, H., Kadav, A., Durdanovic, I., Samet, H., and Graf, H. P. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016

    2016 arXiv

  18. [26]

    Videomamba: State space model for efficient video understanding

    Li, K., Li, X., Wang, Y., He, Y., Wang, Y., Wang, L., and Qiao, Y. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024

    2024 arXiv

  19. [27]

    Videomamba: State space model for efficient video understanding

    Li, K., Li, X., Wang, Y., He, Y., Wang, Y., Wang, L., and Qiao, Y. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pp.\ 237--255. Springer, 2025

    2025

  20. [28]

    Supervised masked knowledge distillation for few-shot transformers

    Lin, H., Han, G., Ma, J., Huang, S., Lin, X., and Chang, S.-F. Supervised masked knowledge distillation for few-shot transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19649--19659, 2023

    2023

  21. [29]

    Hrank: Filter pruning using high-rank feature map

    Lin, M., Ji, R., Wang, Y., Zhang, Y., Zhang, B., Tian, Y., and Shao, L. Hrank: Filter pruning using high-rank feature map. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1529--1538, 2020

    2020

  22. [30]

    Fq-vit: Post-training quantization for fully quantized vision transformer

    Lin, Y., Zhang, T., Sun, P., Li, Z., and Zhou, S. Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824, 2021

    2021 arXiv

  23. [31]

    Efficientvit: Memory efficient vision transformer with cascaded group attention

    Liu, X., Peng, H., Zheng, N., Yang, Y., Hu, H., and Yuan, Y. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 14420--14430, 2023

    2023

  24. [32]

    VMamba : Visual State Space Model , April 2024

    Liu, Y., Tian, Y., Zhao, Y., Yu, H., Xie, L., Wang, Y., Ye, Q., and Liu, Y. VMamba : Visual State Space Model , April 2024. URL http://arxiv.org/abs/2401.10166. arXiv:2401.10166 [cs]

    2024 arXiv

  25. [33]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pp.\ 10012--10022, 2021 a

    2021

  26. [34]

    Post-training quantization for vision transformer

    Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., and Gao, W. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34: 0 28092--28103, 2021 b

    2021

  27. [35]

    Swin transformer v2: Scaling up capacity and resolution

    Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, pp.\ 12009--12019, 2022

    2022

  28. [36]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Ma, J., Li, F., and Wang, B. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024

    2024 arXiv

  29. [37]

    S4nd: Modeling images and videos as multidimensional signals with state spaces

    Nguyen, E., Goel, K., Gu, A., Downs, G., Shah, P., Dao, T., Baccus, S., and R \'e , C. S4nd: Modeling images and videos as multidimensional signals with state spaces. Advances in neural information processing systems, 35: 0 2846--2861, 2022

    2022

  30. [38]

    Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024

    Pei, X., Huang, T., and Xu, C. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. URL https://arxiv.org/abs/2403.09977

    2024 arXiv

  31. [39]

    K., et al

    Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Cao, H., Cheng, X., Chung, M., Grella, M., GV, K. K., et al. RWKV: reinventing rnns for the transformer era. In EMNLP, pp.\ 14048--14077, 2023

    2023

  32. [40]

    Dynamicvit: Efficient vision transformers with dynamic token sparsification

    Rao, Y., Zhao, W., Liu, B., Lu, J., Zhou, J., and Hsieh, C.-J. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34: 0 13937--13949, 2021

    2021

  33. [41]

    Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks, 2023

    Rao, Y., Liu, Z., Zhao, W., Zhou, J., and Lu, J. Dynamic spatial sparsification for efficient vision transformers and convolutional neural networks, 2023. URL https://arxiv.org/abs/2207.01580

    2023 arXiv

  34. [42]

    You only look once: Unified, real-time object detection

    Redmon, J., Divvala, S., Girshick, R., and Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 779--788, 2016

    2016

  35. [43]

    Sengupta, S., Harris, M., Zhang, Y., and Owens, J. D. Scan primitives for gpu computing. In Proceedings of the 22nd ACM SIGGRAPH/EUROGRAPHICS Symposium on Graphics Hardware, GH '07, pp.\ 97–106, Goslar, DEU, 2007. Eurographics Association. ISBN 9781595936257

    2007

  36. [44]

    R., Zhao, W., Wang, K., and You, Y

    Shi, M., Zhou, Y., Yu, R., Li, Z., Liang, Z., Zhao, X., Peng, X., Vedantam, S. R., Zhao, W., Wang, K., and You, Y. Faster vision mamba is rebuilt in minutes via merged token re-training, 2025. URL https://arxiv.org/abs/2412.12496

    2025 arXiv

  37. [45]

    and Zisserman, A

    Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015

    2015

  38. [46]

    T., Warrington, A., and Linderman, S

    Smith, J. T., Warrington, A., and Linderman, S. Simplified state space layers for sequence modeling. In ICLR, 2022

    2022

  39. [47]

    Patch slimming for efficient vision transformers

    Tang, Y., Han, K., Wang, Y., Xu, C., Guo, J., Xu, C., and Tao, D. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12165--12174, 2022

    2022

  40. [48]

    Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting

    Tang, Y., Dong, P., Tang, Z., Chu, X., and Liang, J. Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 5663--5673, 2024

    2024

  41. [49]

    Dim: Diffusion mamba for efficient high-resolution image synthesis

    Teng, Y., Wu, Y., Shi, H., Ning, X., Dai, G., Wang, Y., Li, Z., and Liu, X. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024

    2024 arXiv

  42. [50]

    Training data-efficient image transformers & distillation through attention

    Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., and J \' e gou, H. Training data-efficient image transformers & distillation through attention. In ICML, pp.\ 10347--10357, 2021

    2021

  43. [51]

    Z., Khabsa, M., Fang, H., and Ma, H

    Wang, S., Li, B. Z., Khabsa, M., Fang, H., and Ma, H. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020

    2006 arXiv

  44. [52]

    Mamba-unet: Unet-like pure visual mamba for medical image segmentation

    Wang, Z., Zheng, J.-Q., Zhang, Y., Cui, G., and Li, L. Mamba-unet: Unet-like pure visual mamba for medical image segmentation. arXiv preprint arXiv:2402.05079, 2024

    2024 arXiv

  45. [53]

    L., Gugger, S., Drame, M., Lhoest, Q., and Rush, A

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., Davison, J., Shleifer, S., von Platen, P., Ma, C., Jernite, Y., Plu, J., Xu, C., Scao, T. L., Gugger, S., Drame, M., Lhoest, Q., and Rush, A. M. Huggingface's tr...

    2020 arXiv

  46. [54]

    S., and Xie, S

    Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I. S., and Xie, S. Convnext v2: Co-designing and scaling convnets with masked autoencoders, 2023. URL https://arxiv.org/abs/2301.00808

    2023 arXiv

  47. [55]

    Unified perceptual parsing for scene understanding

    Xiao, T., Liu, Y., Zhou, B., Jiang, Y., and Sun, J. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pp.\ 418--434, 2018

    2018

  48. [56]

    Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation

    Xing, Z., Ye, T., Yang, Y., Liu, G., and Zhu, L. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp.\ 578--588. Springer, 2024

    2024

  49. [57]

    Yang, C., Chen, Z., Espinosa, M., Ericsson, L., Wang, Z., Liu, J., and Crowley, E. J. Plainmamba: Improving non-hierarchical mamba in visual recognition. In 35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024 . BMVA, 2024. URL https://pape...

    2024

  50. [58]

    Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis

    Yang, J., Shi, R., and Ni, B. Medmnist classification decathlon: A lightweight automl benchmark for medical image analysis. In IEEE 18th International Symposium on Biomedical Imaging (ISBI), pp.\ 191--195, 2021

    2021

  51. [59]

    Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification

    Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., and Ni, B. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10 0 (1): 0 41, 2023

    2023

  52. [60]

    Adavit: Adaptive tokens for efficient vision transformer, 2022

    Yin, H., Vahdat, A., Alvarez, J., Mallya, A., Kautz, J., and Molchanov, P. Adavit: Adaptive tokens for efficient vision transformer, 2022. URL https://arxiv.org/abs/2112.07658

    2022 arXiv

  53. [61]

    I., Han, X., Gao, M., Lin, C.-Y., and Davis, L

    Yu, R., Li, A., Chen, C.-F., Lai, J.-H., Morariu, V. I., Han, X., Gao, M., Lin, C.-Y., and Davis, L. S. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 9194--9203, 2018

    2018

  54. [62]

    and Li, Z

    Yue, Y. and Li, Z. Medmamba: Vision mamba for medical image classification, 2024. URL https://arxiv.org/abs/2403.03849

    2024 arXiv

  55. [63]

    Exploring token pruning in vision state space models

    Zhan, Z., Kong, Z., Gong, Y., Wu, Y., Meng, Z., Zheng, H., Shen, X., Ioannidis, S., Niu, W., Zhao, P., et al. Exploring token pruning in vision state space models. arXiv preprint arXiv:2409.18962, 2024

    2024 arXiv

  56. [64]

    Scene parsing through ade20k dataset

    Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., and Torralba, A. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 633--641, 2017

    2017

  57. [65]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Zhu, L., Liao, B., Zhang, Q., Wang, X., Liu, W., and Wang, X. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024

    2024

  58. [66]

    Vision transformer pruning

    Zhu, M., Tang, Y., and Han, K. Vision transformer pruning. arXiv preprint arXiv:2104.08500, 2021

    2021 arXiv

  59. [67]

    Discrimination-aware channel pruning for deep neural networks

    Zhuang, Z., Tan, M., Zhuang, B., Liu, J., Guo, Y., Wu, Q., Huang, J., and Zhu, J. Discrimination-aware channel pruning for deep neural networks. Advances in neural information processing systems, 31, 2018

    2018

  60. [68]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.