Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Compress Any Segment Anything Model (SAM)

T0 review · 3 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 The paper claims that any Segment Anything Model can be compressed to roughly a fifth of its size by storing each pair of weights as one integer index on a dense trajectory, with under 1% segmentation loss and no fine-tuning.

desk verdict Birkhoff is a legitimate extension of Hyper-Compression to SAMs with strong external results, but the 60-second and inference-acceleration claims are overstated and need fixing. read the letter →

arxiv 2507.08765 v1 pith:G6WBBX2M submitted 2025-07-11 cs.CV cs.AI

Juntong Fan , Zhiwei Hao , Jianqiang Shen , Shang-Ling Jui , Yi Zhang , Jing-Xiao Liao , Feng-Lei Fan This is my paper
classification cs.CVcs.AI
keywords SegmentAnythingModelcompressiondata-freeHyper-CompressiondensetrajectoryHyperLinearpost-trainingquantizationzero-shotsegmentation
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

Birkhoff claims that the weight matrices of Segment Anything Models can be lossy-compressed without any data or fine-tuning by exploiting a density property of irrational rotations: a single integer index along a one-dimensional trajectory stands in for a pair of adjacent weights. On 18 SAM-family models evaluated on COCO, LVIS, and SA-1B, the paper reports compression ratios mostly above 4.3x and up to 5.17x, with post-compression changes in mIoU and mAP almost always within 1%, and with the compression itself finished in under a minute for most models. The practical stake is that a segmentation foundation model trained at great data cost can be shrunk to roughly a fifth of its storage size while retaining its original zero-shot behavior, which matters for deployment on phones, vehicles, and clinical settings. A fused linear operator called HyperLinear is what makes the compressed model still infer almost as fast as the original.

What carries the argument

The carrying object is the dense torus trajectory generated by the fractional-part map $\tau(z)=z-\lfloor z\rfloor$ with frequency vector $a=[1/(\pi+1),1/(\pi+2)]$, or its adaptive variant $a=[(l/U),l]/\sqrt{(l/U)^2+l^2}$. Because $a_1,a_2$ are irrationally related, the set $\{\tau(\theta\cdot a):\theta\in\{0,\ldots,U\}\}$ samples a curve that is dense in the unit square, so each adjacent weight pair $w^{(g)}$ is encoded by the single integer $\theta^*_g$ that best reproduces it (Eqs. 7-15). Algorithm 1 searches over the box size $l$, codebook size $U$, and category count $M$ to minimize mean absolute error, and the HyperLinear operator fuses decompression with a block matrix product (Eq. 23) so that one memory access decodes two parameters during inference.

What would settle it

Run Birkhoff's Algorithm 1 on a SAM variant with outlier-heavy or strongly non-uniform linear-layer weights, for instance a model with weights far outside the chosen box before scaling, using the default hyperparameters of Table IV, and measure the mIoU change on SA-1B. If a mean absolute error near 0.002 is accompanied by a performance drop larger than 1%, the finite-codebook coverage assumption is falsified; similarly, finding any model in the 18-model family whose post-compression score drop exceeds 1% under the reported settings would contradict the core claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a SAM layer's weight matrix can be replaced by a much smaller integer matrix $\Theta$ plus a few auxiliary scalars, with no data and no retraining, because each $1\times 2$ block of adjacent weights is reproduced to within a small mean absolute error by evaluating a one-parameter fractional-part trajectory at an integer index. On 18 SAM-family models tested on COCO, LVIS, and SA-1B, the paper reports compression ratios above 4.3x for most models (5.17x for SAM2-B) and post-compression mIoU/mAP changes almost always inside 1%, with the whole compression finished in under 60 seconds for most models. The paper also claims that its fused HyperLinear operator keeps inference essentially at the original model's speed, and that on COCO box-prompt segmentation the data-free method matches or beats fine-tuning-based and data-free quantization baselines at comparable ratios.

Load-bearing premise

For every layer of every SAM variant, each pair of adjacent weights must be close enough to one of the finitely many sampled points on the trajectory that the integer-index approximation keeps the accumulated error beneath the level that changes segmentation scores by about 1%.

Editorial extensions

If this is right

  • Compression without retraining sidesteps catastrophic forgetting and the need for the original training data, which is often unavailable for domain-specific SAM variants.
  • At 4-5x compression, storage footprint drops from hundreds of megabytes or gigabytes to tens or hundreds of megabytes, making large ViT-H models more feasible on mobile and edge devices.
  • Most models compress in under 60 seconds, so a cloud-based model factory could batch-compress many SAM variants on demand.
  • Because HyperLinear fuses decompression with matrix multiplication, post-compression inference stays close to the original speed, with the largest models losing only a few percent of speed.
  • On COCO box-prompt segmentation, the data-free Birkhoff result at 4-5x ratio matches or beats INT6 quantization baselines, even ones that use fine-tuning.

Reading between the lines

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

  • Editorial extension: the mechanism is not SAM-specific; any network whose weights are stored as dense matrices could be compressed the same way, so the strongest untested implication is that the same fidelity guarantee may transfer to other transformer-heavy vision and language models.
  • Editorial extension: the compression ratio is currently tied to the group size of two weights per integer index, so pushing substantially beyond 5x would require trajectories dense in higher dimensions, entropy coding of the stored indices, or a relaxation of the sub-1% performance budget.
  • Editorial extension: the paper's link between MAE and downstream mIoU is empirical; a rigorous bound connecting codebook size $U$, MAE, and final segmentation accuracy would let a practitioner verify the 1% guarantee before deployment rather than after.
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

3 major / 5 minor

Summary. The paper proposes Birkhoff, a data-free compression method for Segment Anything Models and their variants. The method builds on Hyper-Compression, which replaces each pair of weights by a single integer index along a dense one-dimensional trajectory, and introduces a fused CUDA operator, HyperLinear, that combines decompression with matrix multiplication to preserve inference speed. The authors report experiments on 18 SAM variants across COCO, LVIS, and SA-1B, claiming compression ratios of 3.3x to 5.17x, compression times mostly under one minute, post-compression performance drops mostly within 1%, and inference slowdowns of only a few percent. The code is open-sourced.

Significance. If the empirical claims hold, Birkhoff is a practically valuable data-free alternative to quantization, pruning, and distillation for SAM-family models. The evaluation is broad (18 models, three external datasets) and uses external mIoU/mAP metrics, so the central results are not circularly validated. The open-sourced code, the absence of fine-tuning data requirements, and the dedicated HyperLinear operator for inference acceleration are concrete strengths. The main weaknesses are that the universal claims in the abstract and introduction are stronger than the reported measurements, and the finite-codebook approximation used in the actual algorithm lacks a supporting error bound.

major comments (3)
  1. [Abstract and Section I vs. Section IV-B1/Figure 6] The abstract states that 'the compression is finished within 60 seconds for all models' and the introduction repeats 'no more than 60s in all cases.' This is directly contradicted by Section IV-B1 and Figure 6, which report 95.33 seconds for SAM-H, 96.54 seconds for SAM-HQ-H, and 95.40 seconds for MobileSAMv2(ViT-H); the text itself acknowledges that these ViT-H models take 'approximately 90 seconds.' Since fast compression is a core advertised advantage of the method, this overclaim must be corrected by restating the bound as applying to most models or by reporting the ViT-H times as a separate class.
  2. [Section III-A, Eq. (8), and Algorithm 1] Equation (8) asserts that for every weight pair and every epsilon there exists a continuous theta* satisfying the approximation, but Algorithm 1 searches only over integers 0..U, with U chosen from a small candidate set per model. The paper provides no bound connecting U, the finite trajectory sampling spacing, per-layer reconstruction MAE, and downstream segmentation accuracy. This matters because a layer whose weights are not near the sampled trajectory could push the performance drop above the claimed 1%. Please supply either a worst-case MAE bound as a function of U or an empirical sensitivity analysis reporting per-layer reconstruction error, the worst-case layer, and mIoU/mAP degradation as U varies.
  3. [Section III-A, Figure 6, and Table IV] The compression ratio is not precisely defined. The method stores one integer per pair of float weights, which by itself yields at most a 4x ratio if the integer is stored as int16 and 8x if stored as uint8, yet the paper reports ratios of 4.31x to 5.17x. The manuscript must specify the bit-width and packing format of the stored integer tensor and auxiliary parameters, and state how the reported ratio is computed. This is also necessary for a fair comparison with the INT6 quantization ratios in Table VI.
minor comments (5)
  1. [Section IV-B1] The text says 'For instance, SAM-B achieves a 5.17x ratio in 20.83 seconds,' but Figure 6 reports 5.17x and 20.83 seconds for SAM2-B, while SAM-B has 4.31x and 15.81 seconds; this appears to be a typo and should be corrected so the headline example matches the abstract.
  2. [Throughout] There are several typos: 'cetegories' in Section III-A.2, 'ompress' in Section IV-B1, and 'Supplymentary' in Section IV-B2. These should be fixed.
  3. [Eq. (23)] The symbol S is used both for the scaling factor in the decompression formula and for the block size in Eqs. (17) and (22); please rename one of them to avoid ambiguity.
  4. [References] Reference [10] lists the arXiv identifier as '2023.09602', which does not appear to be valid; please verify and correct the identifier.
  5. [Section IV-B2] Detailed per-model results are deferred to a 'Supplementary' that is not included with the manuscript; please include at least one table in the main text with per-model mIoU/mAP before and after compression so that the 'within 1% in most cases' claim can be checked directly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the compression method is re-derived in place, hyperparameters are fit by MAE in the usual lossy-compression sense, and the central claims are validated on external benchmarks; the 60-second timing overclaim is an internal accuracy error, not a circular derivation.

full rationale

Birkhoff's derivation chain is not circular. The Hyper-Compression mechanism is fully described in Sections II-C and III-A of this paper: pairs of weights are encoded by an irrational-rotation trajectory theta -> tau(theta*a), and the existence of a dense trajectory is supported by ergodic theory [31], an external mathematical source, rather than by the authors' own prior work. Although the method is inherited from the authors' own [13], the present paper re-derives the relevant equations (Eqs. 4-15) and Algorithm 1 in place; [13] functions as an attribution, not as the load-bearing justification for the compression guarantee. The hyperparameters (l, U, M) are selected by minimizing decompression MAE on the model's own weights (Algorithm 1, Section III-C1), which is ordinary rate-distortion fitting in lossy compression, and the paper does not rename this fit as a prediction. The claimed validation ('less than 1% drop', 'compression ratio 5.17x') is measured on held-out external benchmarks (COCO, LVIS, SA-1B) against the original models, so the central claim has independent empirical content. The abstract's 'compression is finished within 60 seconds for all models' is contradicted by the paper's own Figure 6, which reports 95.33s, 96.54s, and 95.40s for the three ViT-H models; this is a factual/consistency defect in the reported results, not a circularity of the derivation. No fitted parameter is relabeled as a prediction, no uniqueness claim is imported from the authors' prior work, and no ansatz is hidden behind a citation. Therefore no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on a few assumed mathematical facts and empirical design choices. The most important are the density of the trajectory (only strictly true for continuous theta), the finite-codebook coverage for real SAM weights, and MAE as a proxy for segmentation quality. The hyperparameters l, U, and M are fitted to the model weights during compression. No new physical or conceptual entities are introduced; HyperLinear is a software operator.

free parameters (3)
  • l (box side length) = 0.1 for all models (chosen from alternatives)
    Determines the side of the square box into which parameter pairs are scaled; l=0.1 chosen after the analysis in Figure 3; affects MAE and compression fidelity.
  • U (trajectory upper bound) = per-model candidate sets, e.g., 1600 for SAM-B, up to 40000 for sensitive models
    Integer bound for theta; larger U gives a denser codebook and lower error but lowers the compression ratio. Selected per model by minimizing MAE in Algorithm 1.
  • M (number of scaling categories) = [1,2,3] for all models
    Number of outlier scaling categories; hand-selected and small in range, used in Eq. (11)-(12).
assumptions (4)
  • standard math The sequence tau(theta*a) with a=[1/(pi+1), 1/(pi+2)] is dense in [0,1]^2 because a1 and a2 are irrationally independent (Sec III-A).
    Invoked to justify representing any 2D weight pair by a scalar theta; the rational independence of the specific a values is asserted, not proved, and density holds for continuous theta, not for the finite U used in the algorithm.
  • domain assumption After centering and scaling into a box of side l, most or all weight pairs lie close enough to the discrete trajectory that the nearest of U+1 codebook points yields MAE around 0.001 (Eq. 8, Eq. 14, Figure 3).
    This is the load-bearing empirical premise of the whole compression scheme; it is validated only on the 18 tested models, and no theoretical coverage bound is given.
  • domain assumption Minimizing mean absolute error between original and decompressed weights is a sufficient proxy for preserving segmentation mIoU/mAP (Algorithm 1, Sec IV-A2).
    The hyperparameters U, l, M are selected to minimize MAE, not to optimize downstream metrics; this assumes MAE and task performance are tightly correlated.
  • domain assumption The prior Hyper-Compression algorithm from [13] is correct and directly applicable to SAM linear layers (Sec II-C).
    Birkhoff inherits the decompression formulas and codebook search from the authors' own earlier paper without re-derivation; this is self-citation, but the evaluation here is external.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compress Any Segment Anything Model (SAM)." pith.science (2026). https://pith.science/paper/G6WBBX2M

@misc{pith2026250708765,
  author       = {Pith},
  title        = {Pith review of: Compress Any Segment Anything Model (SAM)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G6WBBX2M}},
  note         = {Machine review of arXiv:2507.08765}
}
read the original abstract

Due to the excellent performance in yielding high-quality, zero-shot segmentation, Segment Anything Model (SAM) and its variants have been widely applied in diverse scenarios such as healthcare and intelligent manufacturing. Therefore, effectively compressing SAMs has become an increasingly pressing practical need. In this study, we propose Birkhoff, a novel data-free compression algorithm for SAM and its variants. Unlike quantization, pruning, distillation, and other compression methods, Birkhoff embodies versatility across model types, agility in deployment, faithfulness to the original model, and compactness in model size. Specifically, Birkhoff introduces a novel compression algorithm: Hyper-Compression, whose core principle is to find a dense trajectory to turn a high-dimensional parameter vector into a low-dimensional scalar. Furthermore, Birkhoff designs a dedicated linear layer operator, HyperLinear, to fuse decompression and matrix multiplication to significantly accelerate inference of the compressed SAMs. Extensive experiments on 18 SAMs in the COCO, LVIS, and SA-1B datasets show that Birkhoff performs consistently and competitively in compression time, compression ratio, post-compression performance, and inference speed. For example, Birkhoff can achieve a compression ratio of 5.17x on SAM2-B, with less than 1% performance drop without using any fine-tuning data. Moreover, the compression is finished within 60 seconds for all models.

Figures

Figures reproduced from arXiv: 2507.08765 by the authors.

Figure 1
Figure 1. The basic principle of the Hyper-Compression, where [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our Birkhoff compression framework. First, model parameters are subjected to per-tensor compression and then stored with auxiliary global parameters. Then, the compressed result of each tensor is packed and stored with its corresponding auxiliary global parameters. Lastly, parameters from a model’s linear layers are loaded into the HyperLinear operator layer by layer for inference. hyperfunction h(θ; … view at source ↗
Figure 3
Figure 3. The relationship between the parameter l and the proportion of inner points. 2) Hyper-Compression needs to scale points into the box [¯xi − l 2 , x¯i + l 2 ] × [¯yi − l 2 , y¯i + l 2 ]. To deal with outliers, it adopts different scaling factors for different points. Specifi￾cally, all the points w (1) , w (2) , . . . , w (G) are divided into M cetegories based on their distances from the centroid [¯xi , y¯i ]. For e… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Technical evolution of decompression: from theoretical element-wise operation to vector-wise operation, and finally to our proposed block-wise one [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The comparison between the block-wise matrix multiplication and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The compression time and achieved compression ratios for 18 SAMs [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The accuracy and sizes (MB) of 18 SAMs and their corresponding [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: We select SAM-B, SAM-L, SAM-H, and MedSAM-B for a visual [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 19 canonical work pages

  1. [1]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026

    2023

  2. [2]

    Medsam-u: Uncertainty-guided auto multi-prompt adaptation for reliable medsam,

    N. Zhou, K. Zou, K. Ren, M. Luo, L. He, M. Wang, Y . Chen, Y . Zhang, H. Chen, and H. Fu, “Medsam-u: Uncertainty-guided auto multi-prompt adaptation for reliable medsam,”arXiv preprint arXiv:2409.00924, 2024

    arXiv 2024

  3. [3]

    Segment anything in high quality,

    L. Ke, M. Ye, M. Danelljan, Y .-W. Tai, C.-K. Tang, F. Yuet al., “Segment anything in high quality,” Advances in Neural Information Processing Systems, vol. 36, pp. 29 914–29 934, 2023

    work page 2023

  4. [4]

    Mobilesamv2: Faster segment anything to everything,

    C. Zhang, D. Han, S. Zheng, J. Choi, T.-H. Kim, and C. S. Hong, “Mobilesamv2: Faster segment anything to everything,” arXiv preprint arXiv:2312.09579, 2023

    arXiv 2023

  5. [5]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R¨adle, C. Rolland, L. Gustafson et al., “Sam 2: Segment anything in images and videos,” arXiv preprint arXiv:2408.00714 , 2024

    arXiv 2024

  6. [6]

    Edgesam: Prompt-in-the- loop distillation for on-device deployment of sam,

    C. Zhou, X. Li, C. C. Loy, and B. Dai, “Edgesam: Prompt-in-the- loop distillation for on-device deployment of sam,” arXiv preprint arXiv:2312.06660, 2023

    arXiv 2023

  7. [7]

    Efficientsam: Leveraged masked image pretraining for efficient segment anything,

    Y . Xiong, B. Varadarajan, L. Wu, X. Xiang, F. Xiao, C. Zhu, X. Dai, D. Wang, F. Sun, F. Iandola et al. , “Efficientsam: Leveraged masked image pretraining for efficient segment anything,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 16 111–16 121

    work page 2024

  8. [8]

    Tinysam: Pushing the envelope for efficient segment anything model,

    H. Shu, W. Li, Y . Tang, Y . Zhang, Y . Chen, H. Li, Y . Wang, and X. Chen, “Tinysam: Pushing the envelope for efficient segment anything model,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 19, 2025, pp. 20 470–20 478

    work page 2025

Show all 37 references
  1. [9]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,” Nature Communications, vol. 15, no. 1, p. 654, 2024

    2024

  2. [10]

    An empir- ical study of catastrophic forgetting in large language models during continual fine-tuning,

    Y . Luo, Z. Yang, F. Meng, Y . Li, J. Zhou, and Y . Zhang, “An empir- ical study of catastrophic forgetting in large language models during continual fine-tuning,” arXiv preprint arXiv:2023.09602 , 2023

    2023

  3. [11]

    A survey on model compression for large language models,

    X. Zhu, J. Li, Y . Liu, C. Ma, and W. Wang, “A survey on model compression for large language models,” Transactions of the Association for Computational Linguistics , vol. 12, pp. 1556–1577, 2024

    2024

  4. [12]

    Model compression for deep neural networks: A survey,

    Z. Li, H. Li, and L. Meng, “Model compression for deep neural networks: A survey,” Computers, vol. 12, no. 3, p. 60, 2023

    2023

  5. [13]

    Hyper-compression: Model compression via hyperfunction,

    F. Fan, J. Fan, D. Wang, J. Zhang, Z. Dong, S. Zhang, G. Wang, and T. Zeng, “Hyper-compression: Model compression via hyperfunction,” arXiv preprint arXiv:2409.00592 , 2024

    2024

  6. [14]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13 . Springer,...

    2014

  7. [15]

    Lvis: A dataset for large vocabulary instance segmentation,

    A. Gupta, P. Dollar, and R. Girshick, “Lvis: A dataset for large vocabulary instance segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 5356– 5364

    2019

  8. [16]

    Sparsegpt: Massive language models can be accurately pruned in one-shot,

    E. Frantar and D. Alistarh, “Sparsegpt: Massive language models can be accurately pruned in one-shot,” in International Conference on Machine Learning. PMLR, 2023, pp. 10 323–10 337

    2023

  9. [17]

    Pruning neural networks without any data by iteratively conserving synaptic flow,

    H. Tanaka, D. Kunin, D. L. Yamins, and S. Ganguli, “Pruning neural networks without any data by iteratively conserving synaptic flow,” Advances in neural information processing systems , vol. 33, pp. 6377– 6389, 2020

    2020

  10. [18]

    The lottery ticket hypothesis: Finding sparse, trainable neural networks,

    J. Frankle and M. Carbin, “The lottery ticket hypothesis: Finding sparse, trainable neural networks,” in International Conference on Learning Representations, 2018

    2018

  11. [19]

    Autodfp: Automatic data-free pruning via channel similarity reconstruction,

    S. Li, J. Chen, J. Xiang, C. Zhu, and Y . Liu, “Autodfp: Automatic data-free pruning via channel similarity reconstruction,” arXiv preprint arXiv:2403.08204, 2024

    2024 arXiv

  12. [20]

    Billm: Pushing the limit of post-training quantization for llms,

    W. Huang, Y . Liu, H. Qin, Y . Li, S. Zhang, X. Liu, M. Magno, and X. Qi, “Billm: Pushing the limit of post-training quantization for llms,” arXiv preprint arXiv:2402.04291 , 2024

    2024 arXiv

  13. [21]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2...

    2018

  14. [22]

    Integer quanti- zation for deep learning inference: Principles and empirical evaluation,

    H. Wu, P. Judd, X. Zhang, M. Isaev, and P. Micikevicius, “Integer quanti- zation for deep learning inference: Principles and empirical evaluation,” arXiv preprint arXiv:2004.09602 , 2020

    2004 arXiv

  15. [23]

    Low-bit quantiza- tion of neural networks for efficient inference,

    Y . Choukroun, E. Kravchik, F. Yang, and P. Kisilev, “Low-bit quantiza- tion of neural networks for efficient inference,” in 2019 IEEE/CVF In- ternational Conference on Computer Vision Workshop (ICCVW). IEEE, 2019, pp. 3009–3018

    2019

  16. [24]

    Zeroq: A novel zero shot quantization framework,

    Y . Cai, Z. Yao, Z. Dong, A. Gholami, M. W. Mahoney, and K. Keutzer, “Zeroq: A novel zero shot quantization framework,” in Proceedings of JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13 the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp....

    2021

  17. [25]

    Tensor-train decomposition,

    I. V . Oseledets, “Tensor-train decomposition,” SIAM Journal on Scien- tific Computing, vol. 33, no. 5, pp. 2295–2317, 2011

    2011

  18. [26]

    Tensor ring decomposition,

    Q. Zhao, G. Zhou, S. Xie, L. Zhang, and A. Cichocki, “Tensor ring decomposition,” arXiv preprint arXiv:1606.05535 , 2016

    2016 arXiv

  19. [27]

    Data- freeweight compress and denoise for large language models,

    R. Peng, Y . Zhou, Q. Guo, Y . Gao, H. Yan, X. Qiu, and D. Lin, “Data- freeweight compress and denoise for large language models,” arXiv preprint arXiv:2402.16319, 2024

    2024 arXiv

  20. [28]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  21. [29]

    Faster segment anything: Towards lightweight sam for mobile applications,

    C. Zhang, D. Han, Y . Qiao, J. U. Kim, S.-H. Bae, S. Lee, and C. S. Hong, “Faster segment anything: Towards lightweight sam for mobile applications,” arXiv preprint arXiv:2306.14289 , 2023

    2023 arXiv

  22. [30]

    Ptq4sam: Post-training quantization for segment anything,

    C. Lv, H. Chen, J. Guo, Y . Ding, and X. Liu, “Ptq4sam: Post-training quantization for segment anything,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 15 941–15 951

    2024

  23. [31]

    I. P. Cornfeld, S. V . Fomin, and Y . G. Sinai, Ergodic theory. Springer Science & Business Media, 2012, vol. 245

    2012

  24. [32]

    Vector quantization,

    R. Gray, “Vector quantization,” IEEE Assp Magazine , vol. 1, no. 2, pp. 4–29, 1984

    1984

  25. [33]

    Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,

    X. Wei, R. Gong, Y . Li, X. Liu, and F. Yu, “Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,” arXiv preprint arXiv:2203.05740, 2022

    2022 arXiv

  26. [34]

    Exploring plain vision transformer backbones for object detection,

    Y . Li, H. Mao, R. Girshick, and K. He, “Exploring plain vision transformer backbones for object detection,” in European conference on computer vision . Springer, 2022, pp. 280–296

    2022

  27. [35]

    Yolox: Exceeding yolo series in 2021,

    Z. Ge, S. Liu, F. Wang, Z. Li, and J. Sun, “Yolox: Exceeding yolo series in 2021,” arXiv preprint arXiv:2107.08430 , 2021

    2021 arXiv

  28. [36]

    Up or down? adaptive rounding for post-training quantization,

    M. Nagel, R. A. Amjad, M. Van Baalen, C. Louizos, and T. Blankevoort, “Up or down? adaptive rounding for post-training quantization,” in International conference on machine learning. PMLR, 2020, pp. 7197– 7206

    2020

  29. [37]

    Brecq: Pushing the limit of post-training quantization by block reconstruction,

    Y . Li, R. Gong, X. Tan, Y . Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu, “Brecq: Pushing the limit of post-training quantization by block reconstruction,” arXiv preprint arXiv:2102.05426 , 2021

    2021 arXiv

Pith tools

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