Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

iFormer: Integrating ConvNet and Transformer for Mobile Application

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A hybrid CNN-transformer design called iFormer claims a new Pareto-optimal point for mobile vision, reaching 80.4% ImageNet-1k accuracy at 1.10 ms on an iPhone 13.

desk verdict iFormer-M's 80.4% at 1.10 ms is a credible new point in the mobile latency-accuracy trade-off, but the Pareto-optimal claim rests on cross-paper latency numbers that were not re-measured; conditional accept. read the letter →

arxiv 2501.15369 v2 pith:KAA6WJPC submitted 2025-01-26 cs.CV cs.AIcs.LG

Chuanyang Zheng This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords mobilevisionhybridCNN-transformersingle-headattentionmodulationmechanismon-devicelatencyImageNetclassificationmodelacceleration
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

The paper proposes iFormer, a family of hybrid vision backbones that combine fast convolutional stages with a new single-head modulation attention (SHMA) in later low-resolution stages. The central claim is that this design achieves a better accuracy-latency trade-off than existing lightweight networks on real mobile hardware: iFormer-M reaches 80.4% ImageNet-1k top-1 accuracy at 1.10 ms on an iPhone 13, surpassing MobileNetV4-Conv-M (79.9% at 1.08 ms) and matching or beating FastViT-SA12 while running about 1.4x faster. If correct, iFormer would currently be the best reported accuracy per millisecond among compared mobile backbones. The paper also shows gains on COCO object detection, instance segmentation, and ADE20K semantic segmentation at mobile-friendly latencies, and demonstrates that a scaled-up iFormer-H (99M parameters) reaches 84.8% top-1 accuracy without architectural changes.

What carries the argument

The central object is the Single-Head Modulation Attention (SHMA) block. It replaces multi-head self-attention with one attention head applied to the full feature map (with channel-reduced query and key projections), and then modulates the output: the attention-derived context vector gates a parallel feature projection through element-wise sigmoid multiplication. The paper also identifies the reshaping and transposing operations of MHA as the memory-access bottleneck on mobile hardware; SHMA removes them. The supporting machinery is the lightened ConvNeXt base: overlapping convolutional stem, BatchNorm everywhere, stage-ratio reallocation to (2,2,18,2), and 7x7 depthwise convolutions, which together shape the latency budget that SHMA spends on global modeling.

What would settle it

Independently re-measure the on-device latencies of MobileNetV4-Conv-M, FastViT-SA12, and iFormer-M on the same iPhone 13 with the same compiler settings and batch size 1. If the gap between iFormer-M (1.10 ms) and MobileNetV4-Conv-M (1.08 ms) reverses, or if FastViT-SA12 comes out faster than iFormer-M at equal accuracy under identical measurement, the Pareto-optimal claim is false.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the multi-head mechanism in vision transformers is the main latency bottleneck on mobile devices, because the reshaping and transpose operations drive memory-access costs, and that a single attention head equipped with a modulation branch can recover the lost accuracy. Replacing MHA with single-head attention yields a 1.25x speedup with only a 0.1% accuracy drop; adding the modulation branch, which computes a context signal with single-head attention and gates a parallel feature branch via sigmoid multiplication, lifts accuracy to 80.4% at 1.10 ms. The broader architectural recipe is a ConvNeXt-derived convolutional stem and early stages (overlapping convolutional stem, BatchNorm instead of LayerNorm, deepened stage 3, 7x7 and 5x5 kernels) followed by SHMA blocks in the last two stages, plus conditional positional encodings. The paper's headline inference is that this configuration is Pareto-optimal in the latency-accuracy plane on an iPhone 13 among the compared models.

Load-bearing premise

The paper's speed advantage rests on comparing its own measured latencies with latency numbers reported by other papers, assuming those were measured under identical device, compiler, and batch settings; it does not re-measure the competitors itself.

Editorial extensions

If this is right

  • Mobile backbones no longer need to choose between convolution speed and attention quality: a single-head attention with modulation can match or exceed multi-head accuracy at a fraction of the latency.
  • iFormer-M yields a better accuracy-latency point than MobileNetV4-Conv-M despite training for fewer epochs (300 vs 500) and at lower resolution (224 vs 256).
  • The design transfers to dense prediction: iFormer-M beats FastViT-SA12 on COCO detection and instance segmentation by +1.9 and +2.0 AP while running 1.32x faster, and iFormer-L leads ADE20K mIoU among the compared backbones.
  • The architecture scales: iFormer-H (99M parameters) reaches 84.8% ImageNet top-1 accuracy, on par with larger general-purpose transformers, suggesting the hybrid recipe is not limited to tiny mobile models.
  • With knowledge distillation, iFormer-M reaches 81.1% and iFormer-L reaches 82.7% under the same latency, indicating that the architecture benefits from stronger training recipes without architectural changes.

Reading between the lines

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

  • If the latency measurements hold, the single-head attention insight should transfer to other transformer-based mobile networks: any model that currently pays the reshape cost of multi-head attention could adopt the modulation-attention pattern, potentially shifting the whole latency-accuracy frontier rather than just one architecture.
  • A testable extension is to measure how the head-similarity curve in the paper's Figure 3 changes with model scale; if head redundancy grows with depth and width, the optimal attention-head count may shrink as models scale up, which would strengthen the single-head design case.
  • The paper's window-attention plus channel-chunking variant is only evaluated for high-resolution dense prediction; an obvious next test is whether the same chunking trick reduces latency for large-input classification (for example, 384x384 or 512x512 ImageNet), where the paper reports plain SHMA fails.
  • Because the Pareto-optimal claim depends on quoted baseline latencies, an independent rebenchmark of MobileNetV4 and FastViT on identical device, compiler, and batch settings would settle whether iFormer's margin is real or an artifact of heterogeneous measurement; the paper itself does not re-measure baselines.
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 / 6 minor

Summary. The paper proposes iFormer, a family of lightweight hybrid CNN-transformer backbones for mobile devices. The architecture uses convolutional blocks in early high-resolution stages and a novel Single-Head Modulation Attention (SHMA) in later stages, with an optional channel-chunked hybrid window attention for high-resolution dense prediction. The headline result is iFormer-M, which reaches 80.4% ImageNet-1k top-1 accuracy at 1.10 ms measured on an iPhone 13 with Core ML Tools, claimed to be Pareto-optimal in the accuracy-latency plane. The paper also reports object detection, instance segmentation, and semantic segmentation results, plus an extensive set of ablations and scaling experiments. The main claims are empirical and rely on cross-model latency comparisons.

Significance. If the latency measurements are mutually consistent, iFormer-M would be a strong mobile backbone, improving accuracy over MobileNetV4-Conv-M at comparable latency and offering a favorable accuracy-latency trade-off. The architecture is simple, includes a plausible mechanism for reducing MHA memory overhead (single-head modulation), and the paper provides detailed training settings and architecture tables. The code and models are promised. The main empirical evidence is extensive, covering classification, detection, segmentation, ablations, and training-strategy variations. However, the central Pareto-optimal and speed-ratio claims depend on latency values quoted from different papers without a documented unified measurement protocol, which is a substantive correctness risk.

major comments (4)
  1. [Section 4.1, Tables 3, 5, 18] The central claims of Pareto optimality (Fig. 1) and speed advantages (e.g., 1.32x faster than FastViT-SA12 in Table 5) rely on latency numbers for competitor models that are quoted from their respective papers. The paper states only that iFormer latencies were measured on an iPhone 13 with Core ML Tools at batch size 1, "as done in prior works," and does not document the coremltools version, compute-unit policy, precision, warm-up, number of runs, or averaging method. More importantly, there is no statement that baseline latencies were re-measured in the authors' own environment. Because on-device latency is sensitive to these details, a 10-20% systematic difference could move a baseline such as FastViT-SA12 (80.6%, 1.50 ms) to a point that dominates iFormer-M (80.4%, 1.10 ms), or reverse the 1.32x speed advantage in Table 5. To support the Pareto-optimal and speed-ratio claims, the authors must re-measure all compared models under one documented Core ML protocol on the same iPhone 13 device, or explicitly weaken the claims to reflect the reported (non-unified) nature of the comparison.
  2. [Table 3 vs. Table 18] The ImageNet-1k top-1 accuracy for iFormer-L is reported as 81.9% in Table 3 but 81.7% in Table 18, for what should be the same configuration. This discrepancy affects the accuracy coordinate used in Fig. 1 and the Pareto-optimal claim. The authors should clarify whether these are different checkpoints, different training runs, or a typographical error, and ensure that all tables and figures use the same official result.
  3. [Table 5 and Appendix F] The downstream-task results in Table 5 label the backbones as iFormer-M and iFormer-L, but Section F of the appendix shows that these models use a different attention variant, the Channel Chunking Hybrid SHMA, rather than the standard SHMA used in the classification models in Table 3. Table 5 does not state this variant explicitly. Since the architecture differs, the text's comparison "iFormer-M surpasses FastViT-SA12 ... while running 1.32x faster" applies to a modified model, not to the exact iFormer-M from the classification results. The caption and main text should explicitly identify the variant and note that it differs from the classification checkpoint.
  4. [Section 1 (Introduction)] The introduction claims iFormer-M is "1.4x faster than FastViT-SA12" with similar accuracy, but the ratio from Table 3 (1.50/1.10 ms) is 1.36x and from Table 5 (5.27/4.00 ms) is 1.32x. This inconsistency in a quantitative speedup claim, combined with the cross-paper latency issue, makes the speed advantage difficult to verify precisely. The authors should ensure all speed ratios are computed from a single, consistently measured latency table.
minor comments (6)
  1. [Conclusion] There is a typo in the concluding paragraph: "bottlenecks sociated with high-resolution images" should read "bottlenecks associated with high-resolution images."
  2. [Table 15 caption] The caption spells "Singe-Head Modulation Attention" instead of "Single-Head Modulation Attention."
  3. [Table 4 caption] The table reports results with distillation but does not mark iFormer-M and iFormer-L in that table with the same symbol used in the text to indicate training with distillation; consider adding a footnote for clarity.
  4. [Figure 1] The x-axis label appears as "Latency (ms)" but the tick labels are crowded; also consider marking the iFormer points distinctly and noting which table each point comes from, given the iFormer-L discrepancy between Tables 3 and 18.
  5. [Section 2.2 (Related Work)] The sentence discussing SHViT says "the key difference lies in the novel modulation attention," but the paragraph does not cite a specific comparison table in the main text; Table 14 and Appendix D provide the comparison, which could be referenced directly.
  6. [Section 3.3 (Eq. 6)] The complexity formula for SHMA includes a term 2P^2HW C for self-attention; for P=14 this equals 2*196*196*C, which may be worth clarifying as the linear-in-HW expression when P is constant. As written, the text suggests a quadratic dependence on the number of tokens, which is correct when P scales with resolution, but the formula's interpretation would be clearer with an explicit statement of when P is fixed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: iFormer's architecture and benchmark claims are empirical, not derived from their own predictions.

full rationale

The paper's central claims are architectural and empirical: iFormer-M reaches 80.4% ImageNet-1k top-1 accuracy at 1.10 ms on an iPhone 13, and the paper argues this is Pareto-optimal against published mobile backbones. Nothing in the manuscript derives a result from a quantity defined in terms of that same result. The design process in Sec. 3 is an ablation-driven evolution from a ConvNeXt baseline, with each modification (convolutional stem, BatchNorm, depth, stage ratio, kernel size, attention type, width reduction, positional encoding) measured independently in terms of accuracy and latency; no fitted parameter is later renamed as a prediction. The single-head attention motivation in Sec. 3.3 is supported by an empirical cosine-similarity analysis of the paper's own MHA baseline and by a separately trained SHA baseline, not by a self-referential construction. The modulation formulation in Eq. 1-3 is explicitly attributed to external prior work (Ma et al. 2024; Shazeer 2020), and conditional positional encodings are attributed to Chu et al. 2021; neither citation is self-referential. The only self-citation, Zheng et al. 2022 (SAViT), appears in the related-work list of model-compression approaches and plays no load-bearing role in the architecture, training, or evaluation claims. The latency-comparison methodology, in which competitor latencies appear to be quoted rather than re-measured in a single documented Core ML protocol, is a legitimate fairness and correctness risk for the Pareto-optimal claim, but it is not circularity: the iFormer results are not constructed to equal the baseline numbers. The claimed accuracy and latency for iFormer are produced by training on ImageNet-1k and measuring on iPhone 13, independent of the values used for comparison. Accordingly, no circular step satisfies the required standard of exhibiting an equation or fitted parameter that reduces to its own input.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central claim is empirical, not derivational. No free physical constants are fitted. The listed free parameters are architecture hyperparameters selected by hand on ImageNet validation. The axioms are domain assumptions about training recipes, hardware measurement, and the validity of prior building blocks.

free parameters (5)
  • R (query/key channel reduction ratio in SHMA) = 2
    Chosen by hand to balance latency and accuracy (Sec 3.3, Fig 4); affects the width of attention projections.
  • FFN expansion ratio (r) = 3 for most variants (reduced from 4)
    Reduced in the width reduction step to lower latency (Sec 3.3); value selected via ablation-style design.
  • Head dimension (hd) per variant = 64, 88, 96, 128 for T/S/M/L
    Hand-set for each model variant (Table 15); controls the capacity of the single attention head.
  • Stage block distribution = (2,2,18,2) in Sec 3.2 discussion, final Table 15 shows different counts
    Reallocated from (3,3,9,3) to improve latency-accuracy; a design choice validated on ImageNet.
  • Channel chunk count (n) for high-resolution window attention = 16
    Introduced in Sec F to reduce memory traffic; set to 16 for 512x512 inputs without a formal sensitivity study.
assumptions (5)
  • domain assumption The ConvNeXt training recipe (300 epochs, AdamW, cosine schedule) transfers to the iFormer hybrid architecture.
    Used as the default training setup (Sec B.1, Table 9); if this recipe is suboptimal for the new hybrid blocks, reported accuracies could be understated or overstated.
  • domain assumption Apple Core ML Tools latency on iPhone 13 is a reliable proxy for real-world mobile latency.
    All latency claims (Tables 3, 5, 18) depend on this measurement proxy (Sec 3.1, Sec B.2). Different compiler versions or device thermal states could change numbers.
  • domain assumption Single-head attention captures most of multi-head capacity in mobile-scale hierarchical models.
    Justified by cosine similarity analysis in Fig 3 and the SHA baseline in Table 1; the analysis is on the authors' own model, not a proof across architectures.
  • domain assumption The modulation mechanism from EfficientMod (Eq. 1) is a sound building block.
    The paper adopts the abstracted modulation framework (Sec 3.3) and cites Ma et al. 2024; it does not re-derive or test this framework independently.
  • domain assumption Conditional positional encodings (CPE) from Chu et al. improve attention in the iFormer setting.
    CPE is added in the final step of the architecture evolution (Sec 3.3) and is treated as a known good component from prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of iFormer: Integrating ConvNet and Transformer for Mobile Application." pith.science (2026). https://pith.science/paper/KAA6WJPC

@misc{pith2026250115369,
  author       = {Pith},
  title        = {Pith review of: iFormer: Integrating ConvNet and Transformer for Mobile Application},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KAA6WJPC}},
  note         = {Machine review of arXiv:2501.15369}
}
read the original abstract

We present a new family of mobile hybrid vision networks, called iFormer, with a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network, \textit{i.e.}, ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks. Notably, iFormer achieves an impressive Top-1 accuracy of 80.4\% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13, surpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios.

Figures

Figures reproduced from arXiv: 2501.15369 by the authors.

Figure 1
Figure 1. Comparison of latency and accu￾racy between our iFormer and other exist￾ing methods on ImageNet-1k. The latency is measured on an iPhone 13. Our iFormer is Pareto-optimal. Through the decades, convolutional neural net￾works (CNNs) (Krizhevsky et al., 2012; Szegedy et al., 2015; He et al., 2016) have emerged as the primary choice for balancing latency and per￾formance on resource-constrained mobile devices. However, … view at source ↗
Figure 3
Figure 3. The distribution of aver￾age cosine similarity among multiple heads within the MHA mechanism. As the layer depth increases, the simi￾larity goes higher [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Overview of iFormer architecture, detailed convolutional stem, block design, and SHMA. The hatched area in SHMA indicates extra memory-intensive reshaping operations that are eliminated by SHMA. S(·) denotes the softmax function. R is the ratio for reducing channels of query and key. It is set to 2 in iFormer. We omit BN following project or convolution for simplicity. division and square root operations, leading to… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Comparison of SHMA and SHA in SHViT. In SHViT, rC channels are utilized for spatial attention, where r is set to 1 4.67 . SHMA projects the input into a higher dimension of 1 2 C (i.e., R=2) and avoids split and concatenation operations. Furthermore, we integrate iForm…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation

    cs.CV 2026-08 conditional novelty 4.0 of 10

    A unified, reparameterized single-encoder RGB-D segmentation network with linear gated attention and a pyramid merging decoder achieves strong accuracy-efficiency trade-offs on NYUDepth V2 and SUN-RGBD.

  2. GasTwinFormer: A Hybrid Vision Transformer for Livestock Methane Emission Segmentation and Dietary Classification in Optical Gas Imaging

    cs.CV 2025-08 conditional novelty 4.0 of 10

    GasTwinFormer, a hybrid of two existing attention mechanisms, segments cattle methane plumes in thermal video at 74.47% mIoU and contributes a new 11,694-frame OGI beef cattle dataset.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages · cited by 2 Pith papers

  1. [1]

    Output Size(Downs

    r means the expansion ratio in FFN. Output Size(Downs. Rate) iFormer-T iFormer-S iFormer-M iFormer-L Stem 56×56(4×) Conv-BN-GELU 5×5 s2 d16 ×1 Conv-BN-GELU 5×5 s2 d16 ×1 Conv-BN-GELU 5×5 s2 d24 ×1 Conv-BN-GELU 5×5 s2 d24 ×1 Conv-BN-GELU 5×5 s2 d64Conv-BN 1×1 s1 d32 ×1 Conv-BN-GELU 5×5 s2 d64Conv-BN 1×1 s1 d32 ×1 Conv-BN-GELU 5×5 s2 d96Conv-BN 1×1 s1 d48 ×...

    work page 2025

  2. [3]

    As summarized in Table 14, split and concatenate operations introduce additional runtime

    concatenating the attention output with the residual input X2. As summarized in Table 14, split and concatenate operations introduce additional runtime. Fur- thermore, the performance of the SHA in the SHViT exhibits a decline compared to its counterpart in iFormer under similar latency conditions (79.8 v.s. 79.5). This degraded performance may be attribu...

    work page 2025

  3. [7]

    Learning efficient vision transformers via fine-grained manifold distillation.arXiv preprint arXiv:2107.01378,

    Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation.arXiv preprint arXiv:2107.01378,

    arXiv

  4. [9]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861,

    arXiv

  5. [11]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision , pp. 10012–10022, 2021a. Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quanti- z...

    arXiv

  6. [12]

    MoCoViT: Mobile Convolutional Vision Transformer

    12 Published as a conference paper at ICLR 2025 Hailong Ma, Xin Xia, Xing Wang, Xuefeng Xiao, Jiashi Li, and Min Zheng. Mocovit: Mobile convolutional vision transformer. arXiv preprint arXiv:2205.12635,

    work page Pith review arXiv 2025

  7. [14]

    Mobilevit: light-weight, general-purpose, and mobile- friendly vision transformer

    Sachin Mehta and Mohammad Rastegari. Mobilevit: light-weight, general-purpose, and mobile- friendly vision transformer. arXiv preprint arXiv:2110.02178,

    arXiv

  8. [15]

    Separable self-attention for mobile vision transformers

    Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680,

    arXiv

Show all 34 references
  1. [16]

    Lowformer: Hardware efficient design for convolutional transformer backbones

    Moritz Nottebaum, Matteo Dunnhofer, and Christian Micheloni. Lowformer: Hardware efficient design for convolutional transformer backbones. arXiv preprint arXiv:2409.03460,

    arXiv

  2. [17]

    Mobilenetv4-universal models for the mobile ecosystem

    Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518,

    arXiv

  3. [18]

    Glu variants improve transformer

    Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202,

    2002 arXiv

  4. [19]

    Ghostnetv2: Enhance cheap operation with long-range attention

    13 Published as a conference paper at ICLR 2025 Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Chao Xu, and Yunhe Wang. Ghostnetv2: Enhance cheap operation with long-range attention. Advances in Neural Information Processing Systems , 35:9969–9982,

    2025

  5. [20]

    Fastvit: A fast hybrid vision transformer using structural reparameterization

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 5785–5795, 2023a. Pavan Kumar An...

    arXiv

  6. [21]

    Seaformer: Squeeze-enhanced axial transformer for mobile semantic segmentation

    Qiang Wan, Zilong Huang, Jiachen Lu, Gang Yu, and Li Zhang. Seaformer: Squeeze-enhanced axial transformer for mobile semantic segmentation. arXiv preprint arXiv:2301.13156,

    arXiv

  7. [22]

    Focal modulation networks

    Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks. Advances in Neural Information Processing Systems , 35:4203–4217, 2022a. 14 Published as a conference paper at ICLR 2025 Qiming Yang, Kai Zhang, Chaoxiang Lan, Zhi Yang, Zheyang Li, Wenming Tan...

    2025

  8. [23]

    Edgeformer: Improving light-weight convnets by learning from vision transformers

    Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Edgeformer: Improving light-weight convnets by learning from vision transformers. arXiv preprint arXiv:2203.03952, 2,

    arXiv

  9. [24]

    Rethinking mobile block for efficient attention-based models

    Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, and Chengjie Wang. Rethinking mobile block for efficient attention-based models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 138...

    2023

  10. [25]

    Cas-vit: Convolu- tional additive self-attention vision transformers for efficient mobile applications

    Tianfang Zhang, Lei Li, Yang Zhou, Wentao Liu, Chen Qian, and Xiangyang Ji. Cas-vit: Convolu- tional additive self-attention vision transformers for efficient mobile applications. arXiv preprint arXiv:2408.03703,

    arXiv

  11. [26]

    Repnext: A fast multi-scale cnn using structural repa- rameterization

    Mingshu Zhao, Yi Luo, and Yong Ouyang. Repnext: A fast multi-scale cnn using structural repa- rameterization. arXiv preprint arXiv:2406.16004,

    arXiv

  12. [27]

    Lightweight vision transformer with cross feature attention

    Youpeng Zhao, Huadong Tang, Yingying Jiang, Qiang Wu, et al. Lightweight vision transformer with cross feature attention. arXiv preprint arXiv:2207.07268,

    arXiv

  13. [28]

    training config iFormer-T/S/M/L/H resolution 2242 weight init trunc

    15 Published as a conference paper at ICLR 2025 A A PPENDIX B E XPERIMENTAL SETTINGS B.1 I MAGE CLASSIFICATION Table 9: ImageNet-1K training settings. training config iFormer-T/S/M/L/H resolution 2242 weight init trunc. normal (0.2) optimizer AdamW base learning rate 4e-3 (T/S...

    2025

  14. [30]

    We hypothesize that implementing more effective spatial mixing before the FFN diminishes its significance

    attempt to insert a depthwise convolution (DW Conv) within the FFN to perform spatial mixing on the expanded features activations. We hypothesize that implementing more effective spatial mixing before the FFN diminishes its significance. In our iFormer, depthwise convolution w...

    2022

  15. [31]

    Table 13: Object detection & Semantic segmentation results using backbone pretrained for 450 epochs

    It shows that training iFormer-L2 for 450 epochs yields improved performance, obtaining a Top-1 accuracy of 83.9%, even surpassing the ConvNeXt-Base model. Table 13: Object detection & Semantic segmentation results using backbone pretrained for 450 epochs. Backbone Param(M) La...

    2016

  16. [34]

    Channel Chunking

    From the latency comparison in Table 17, we see that simply applying SHMA will encounter a memory bottleneck on mobile devices. Instead, our hybrid SHMA can significantly reduce memory access costs, achieving a mobile latency of 11.46 ms. However, hybrid SHMA still lags much b...

    2018

  17. [2016]

    Token merging: Your vit but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461,

    arXiv

  18. [2017]

    Batch normalization: Accelerating deep network training by reducing internal covari- ate shift

    Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covari- ate shift. arXiv preprint arXiv:1502.03167,

    arXiv

  19. [2018]

    Efficient modulation for vision networks

    Xu Ma, Xiyang Dai, Jianwei Yang, Bin Xiao, Yinpeng Chen, Yun Fu, and Lu Yuan. Efficient modulation for vision networks. arXiv preprint arXiv:2403.19963,

    arXiv

  20. [2019]

    Mixformer: Mixing features across windows and dimensions

    Qiang Chen, Qiman Wu, Jian Wang, Qinghao Hu, Tao Hu, Errui Ding, Jian Cheng, and Jingdong Wang. Mixformer: Mixing features across windows and dimensions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 5249–5259, 2022a. Yinpeng Chen, ...

    arXiv

  21. [2020]

    In https://github.com/apple/coremltools

    CoreML. In https://github.com/apple/coremltools. Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hi- erarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pp. 248–255. Ieee,

    2009

  22. [2021]

    Fastervit: Fast vision transformers with hierarchical attention

    11 Published as a conference paper at ICLR 2025 Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189,

    2025 arXiv

  23. [2022]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  24. [2023]

    Mmdetection: Open mmlab detection toolbox and benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155,

    1906 arXiv

  25. [2024]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450,

    arXiv

  26. [4096]

    It is possible to further improve performance by adjusting the learning rates for different model variants, which we will explore in the future

    We employ the same learning rate across all models. It is possible to further improve performance by adjusting the learning rates for different model variants, which we will explore in the future. For distillation, we use the RegNetY-16GF model as the teacher model and apply a...

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.