Pith. sign in

REVIEW 2 major objections 3 minor 31 references

Sound Field Reconstruction Using Physics-Informed Boundary Integral Networks

T0 review · 2 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that sound field reconstruction reduces to learning a boundary pressure function, and that the resulting boundary integral network outperforms existing physics-informed baselines on sparse microphone data.

desk verdict New application of boundary-integral networks to sound field reconstruction with promising simulated results, but the written method omits how the normal derivative in the Kirchhoff-Helmholtz integral is computed, making it not reproducible as described. read the letter →

arxiv 2506.03917 v1 pith:EUOP3HZM submitted 2025-06-04 eess.AS

Stefano Damiano , Toon van Waterschoot This is my paper
classification eess.AS
keywords soundfieldreconstructionboundaryintegralequationphysics-informedneuralnetworksKirchhoff-HelmholtzHelmholtzacousticpressureroomacoustics
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 tries to establish that the acoustic pressure throughout a source-free region can be reconstructed from sparse microphone recordings by learning only the pressure distribution on the region's boundary. A shallow neural network outputs the boundary pressure, and the Kirchhoff-Helmholtz boundary integral equation converts it into pressure estimates at any interior point, so every reconstruction is automatically a solution of the homogeneous Helmholtz equation. Training requires only a mean-squared-error loss at the measured microphone positions, with no PDE residual term and no hyperparameter to balance it. In a simulated 2D reverberant room, the method is reported to outperform a physics-informed dictionary learning baseline and a standard PINN across frequency, microphone count, and number of integration points, and to need fewer microphones and fewer boundary points to match the baselines' accuracy. If these results hold, sound field reconstruction becomes more data-efficient and more physically interpretable than current physics-informed approaches.

What carries the argument

The load-bearing object is Eq. (8), the Kirchhoff-Helmholtz boundary integral equation, which writes the pressure $p(r)$ inside a region as an integral over the boundary of the Green's function $G(y,r)$, the boundary pressure $h(y)$, and the normal derivative $\partial h/\partial n_y$. The Green's function is known in closed form, and the paper uses a three-layer MLP with 64 neurons per layer to represent $h(y)$ at a discrete set of integration points on the boundary. The integral is discretized, the pressure at microphone positions is compared with measurements in a mean-squared-error loss, and no PDE residual term is added. This machinery makes every reconstruction an exact Helmholtz solution by construction, eliminating the PDE loss weighting hyperparameter and reducing the sampling problem from the interior of the region to its boundary.

What would settle it

Sample the learned boundary pressure $h(y)$ at the integration points, evaluate Eq. (8) in two versions—one with the normal-derivative term $\partial h/\partial n_y$ computed numerically and one with that term set to zero—and compare the resulting interior fields. If both versions give the same reconstruction, the model is not using the stated integral equation. A complementary check is to compute $(\nabla^2 + k^2)\hat{p}(r)$ at interior checkpoints: large residuals would directly contradict the paper's claim that every output is a Helmholtz solution.

Watch

Extended reading notes

Core claim

The central claim is that the Helmholtz equation's boundary integral representation turns sound field reconstruction into a low-dimensional fitting problem: a shallow network outputs the boundary pressure $h(y)$, and the Kirchhoff-Helmholtz integral (8) carries that boundary data, together with its normal derivative and the known Green's function, into the interior. The authors argue that because the predicted field is generated by the integral equation, it always satisfies the physics, which removes the need for interior collocation points and for the PDE-loss weighting term used by PINNs. Their simulations show lower NMSE and higher NCC than both baselines over the tested frequency range, continued advantage down to roughly 15 microphones, and equal or better performance with fewer boundary integration points than the PINN's interior collocation points.

Load-bearing premise

The method's load-bearing premise is that the boundary pressure $h(y)$ produced by the network is sufficient to evaluate the full Kirchhoff-Helmholtz integral, but the integral also contains the normal derivative $\partial h/\partial n_y$, and the paper never explains how that derivative is obtained; if it is omitted or only implicit, the network is not implementing the stated physical equation.

Editorial extensions

If this is right

  • The PDE-residual loss term and its weighting hyperparameter can be dropped entirely, simplifying the training objective to a single data-fidelity term.
  • Fewer measurement microphones are needed: the paper reports that a PIBI-Net with 30 microphones matches a PINN's best NMSE at 50 microphones, directly lowering sensing cost.
  • Because integration points live on the boundary rather than throughout the region, the problem has one fewer dimension to sample, which the paper connects to faster convergence and better performance at low microphone counts.
  • The method requires no prior knowledge of source positions, room geometry, or boundary conditions, while still producing reconstructions that are Helmholtz solutions by construction.
  • The learned boundary pressure $h(y)$ is itself an interpretable physical quantity that can be inspected or reused, something the black-box field predictions of PINNs do not offer.

Reading between the lines

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

  • Editorial inference: the paper does not state how the normal derivative $\partial h/\partial n_y$ in Eq. (8) is computed; if it is neglected, the method is in effect solving a modified integral equation, and the learned $h(y)$ would silently absorb the missing term. A natural extension is to supply the derivative explicitly through automatic differentiation or a second network.
  • Editorial inference: the boundary-only formulation connects to classical techniques such as nearfield acoustic holography and equivalent source methods; a focused comparison against those model-based baselines would clarify whether the reported gains come from the network or from the integral formulation itself.
  • Editorial inference: because all experiments use simulated data from a single 2D shoebox room, the most direct next test is a real-room recording with known microphone positions, where sensor noise and position errors will show whether the physical interpretability survives.
  • Editorial inference: the claim that the reconstruction is always a Helmholtz solution can be checked independently by evaluating the PDE residual at interior points; if the normal derivative is not computed correctly, this residual will not vanish.
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

2 major / 3 minor

Summary. The paper proposes PIBI-Net, a sound field reconstruction method in which a shallow MLP learns a boundary pressure density that is inserted into the Kirchhoff-Helmholtz boundary integral representation to predict pressure inside a source-free region. The network is trained only on microphone measurements, and experiments in a simulated rectangular room compare it with a physics-informed dictionary learning baseline and a PINN across frequency, number of microphones, and number of integration/collocation points. The paper reports lower NMSE and higher NCC for PIBI in most settings and claims greater data-efficiency, lower-dimensional training, and physically interpretable solutions.

Significance. If the implementation gaps are resolved, the approach is attractive: replacing collocation points in the domain with integration points on the boundary reduces the dimensionality of the learned representation, removes the PDE-loss weight hyperparameter, and connects sound field reconstruction to BEM theory. The paper includes a code link, uses two relevant baselines, and reports quantitative comparisons, which are strengths. However, the main technical contribution rests on a boundary integral formulation that is under-specified in the manuscript, so the stated physical-interpretability and Helmholtz-satisfaction guarantees are not currently established.

major comments (2)
  1. [Sec. III, Eq. (8)] The Kirchhoff-Helmholtz representation in Eq. (8) requires both the boundary pressure h(y) and its normal derivative dh/dn_y on the boundary, but the text only says that the MLP outputs h(y) and that its input is a spatial coordinate r inside Omega. The manuscript never states how dh/dn_y is obtained: automatic differentiation with respect to the boundary coordinate y, a second output head, finite differences, or a simplifying assumption. Moreover, an input coordinate r inside Omega cannot directly supply the boundary density h(y) needed to evaluate the integral over dOmega. As written, Eq. (8) is not computable from the described network, and the claims that PIBI-Net always satisfies the Helmholtz equation and yields an interpretable boundary solution are unsupported. Please specify the network evaluation points, the normal derivative computation, and any approximations; if the derivative term is omitted or approximated, the experimental and conceptual claims must be qualified accordingly.
  2. [Eq. (12), Sec. IV] The NCC formula is dimensionally and normalization-wise incorrect. Dividing |p_hat^H p| by ||p_hat||_2^2 ||p||_2^2 does not give a normalized cross-correlation in the stated range, and the value depends on the overall pressure scale. With the absolute value, the proper range is [0,1], not [-1,1] as written in the text. The standard expression is |p_hat^H p| / (||p_hat||_2 ||p||_2). Since NCC is one of the two primary metrics used throughout the experiments, this must be corrected (or the typo confirmed against the provided code) and the reported NCC values should be recomputed.
minor comments (3)
  1. [Sec. IV, Fig. 4 caption] There is a typo in the caption: '625 Hzon' should be '625 Hz on'.
  2. [Sec. III, network description] Please clarify whether 'the real and imaginary parts are treated separately and then combined' means two independent networks, one network with two outputs, or two separate training runs; this affects the parameter count and the reproducibility of the experiments.
  3. [Sec. IV, Fig. 2c/3c] The claim that PIBI 'demands fewer integration points than PINN requires collocation points' should be quantified by an equal-performance comparison (e.g., the smallest Nint needed to reach the best NMSE of PINN at Ncoll=600), since the presented figures compare equal values of Ncoll and Nint rather than equal performance levels.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the Kirchhoff-Helmholtz constraint is architectural, the network is trained only on microphone data, and the comparisons are against independent baselines.

full rationale

The derivation chain is not circular. The model's forward map is the Kirchhoff-Helmholtz integral equation (Eq. 8), with the MLP supplying the boundary density h(y); the network is trained solely by minimizing the data loss LPIBI = Σ|s(rm) − p̂(rm)|² on M microphone pressures. There is no fitted parameter that is later renamed as a prediction, and no quantity in the evaluation (NMSE/NCC on the 30×30 grid) is used to define the model. The claim that p̂ satisfies the Helmholtz equation follows from the integral representation, not from the learned weights. Prior PIBI-Net work [24]–[27] is external, and the only overlapping-author citation, [23] (PIDL), is used as a comparison baseline rather than as a premise for PIBI's correctness. The paper explicitly lists remaining limitations (sources inside the region, explicit boundary conditions, real data), confirming that the method is not defined by its target benchmark. One non-circular caveat: Eq. (8) requires ∂h/∂n_y, but the paper never states how this normal derivative is obtained (automatic differentiation, a second output head, or a simplifying assumption), and the network input/output description (interior r versus boundary y) is ambiguous. This is a reproducibility and completeness gap, not a circularity, because even an approximate or implicit treatment of ∂h/∂n_y does not make PIBI's results equal to its training data by construction.

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

The method introduces no new physical entity or constant. The free parameters are the learned neural network weights (8,577 parameters), which are estimated from microphone data in a standard supervised manner, and experimental hyperparameters chosen by the authors (learning rate, Nint, M, architecture size), none of which are fitted to make the derivation work. The only potentially ad hoc element is the unstated handling of the normal derivative in the boundary integral, which is a gap in the written derivation.

assumptions (4)
  • standard math The Kirchhoff-Helmholtz boundary integral equation (Eq. 8) provides an explicit representation of any solution to the homogeneous Helmholtz equation inside a source-free region.
    Invoked as background in Section II-B and Section III; it is the core physical and mathematical foundation of the method.
  • domain assumption The reconstruction region Ω is source-free, so the homogeneous Helmholtz equation (1) holds inside it.
    Stated in Section II-A; the method does not handle sources inside Ω, as acknowledged in the conclusions.
  • domain assumption A shallow MLP with 3 hidden layers of 64 neurons can sufficiently approximate the boundary pressure distribution h(y).
    The architecture is taken from [27] and used without justification or ablation; the paper assumes this capacity is enough for the tested frequencies and geometries.
  • standard math The boundary integral can be accurately discretized using Nint equally spaced integration points on the boundary.
    The discretization follows [27] and is described in Section III; the paper experimentally varies Nint and observes convergence, but no quadrature error analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sound Field Reconstruction Using Physics-Informed Boundary Integral Networks." pith.science (2026). https://pith.science/paper/EUOP3HZM

@misc{pith2026250603917,
  author       = {Pith},
  title        = {Pith review of: Sound Field Reconstruction Using Physics-Informed Boundary Integral Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUOP3HZM}},
  note         = {Machine review of arXiv:2506.03917}
}
read the original abstract

Sound field reconstruction refers to the problem of estimating the acoustic pressure field over an arbitrary region of space, using only a limited set of measurements. Physics-informed neural networks have been adopted to solve the problem by incorporating in the training loss function the governing partial differential equation, either the Helmholtz or the wave equation. In this work, we introduce a boundary integral network for sound field reconstruction. Relying on the Kirchhoff-Helmholtz boundary integral equation to model the sound field in a given region of space, we employ a shallow neural network to retrieve the pressure distribution on the boundary of the considered domain, enabling to accurately retrieve the acoustic pressure inside of it. Assuming the positions of measurement microphones are known, we train the model by minimizing the mean squared error between the estimated and measured pressure at those locations. Experimental results indicate that the proposed model outperforms existing physics-informed data-driven techniques.

Figures

Figures reproduced from arXiv: 2506.03917 by the authors.

Figure 1
Figure 1. Geometry of the simulated room considered in the experiments. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mean (solid line) and standard deviation (shaded region) of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Mean (solid line) and standard deviation (shaded region) of the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Real part of the sound field at f = 625 Hz on the 30 × 30 point grid, obtained with the PIDL and PINN baselines and the proposed method PIBI, and compared to the ground truth. The problem is solved using M = 50 measurements with microphones randomly arranged as in the …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

  1. [1]

    Zotter, Ambisonics: A practical 3D audio theory for recording, studio production, sound reinforcement, and virtual reality

    F. Zotter, Ambisonics: A practical 3D audio theory for recording, studio production, sound reinforcement, and virtual reality . Cham: Springer Nature, 2019

    work page 2019

  2. [2]

    MYRiAD: a multi-array room acoustic database,

    T. Dietzen, R. Ali, M. Taseska, and T. van Waterschoot, “MYRiAD: a multi-array room acoustic database,” EURASIP J. Audio Speech Music Process., vol. 2023, Apr. 2023

    work page 2023

  3. [3]

    A Para- metric Approach to Virtual Miking for Sources of Arbitrary Directivity,

    M. Pezzoli, F. Borra, F. Antonacci, S. Tubaro, and A. Sarti, “A Para- metric Approach to Virtual Miking for Sources of Arbitrary Directivity,” IEEE/ACM Trans. Audio Speech Language Process. , vol. 28, 2020

    work page 2020

  4. [4]

    Kernel ridge regression with constraint of Helmholtz equation for sound field interpolation,

    N. Ueno, S. Koyama, and H. Saruwatari, “Kernel ridge regression with constraint of Helmholtz equation for sound field interpolation,” in Proc. 2018 Int. Workshop Acoust. Signal Enhancement (IWAENC) , 2018

    work page 2018

  5. [5]

    Theory and design of multizone soundfield reproduction using sparse methods,

    W. Jin and W. B. Kleijn, “Theory and design of multizone soundfield reproduction using sparse methods,” IEEE/ACM Trans. Audio, Speech Language Process., vol. 23, no. 12, 2015

    work page 2015

  6. [6]

    Soundfield Reconstruction in Reverberant Environments Using Higher-order Microphones and Im- pulse Response Measurements,

    F. Borra, I. D. Gebru, and D. Markovic, “Soundfield Reconstruction in Reverberant Environments Using Higher-order Microphones and Im- pulse Response Measurements,” in Proc. 2019 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP) , pp. 281–285, May 2019

    work page 2019

  7. [7]

    Room Impulse Response Interpolation Using a Sparse Spatio-Temporal Representation of the Sound Field,

    N. Antonello, E. De Sena, M. Moonen, P. A. Naylor, and T. van Waterschoot, “Room Impulse Response Interpolation Using a Sparse Spatio-Temporal Representation of the Sound Field,” IEEE/ACM Trans. Audio Speech Language Process. , vol. 25, Oct. 2017

    work page 2017

  8. [8]

    Sparse Representation of a Spatial Sound Field in a Reverberant Environment,

    S. Koyama and L. Daudet, “Sparse Representation of a Spatial Sound Field in a Reverberant Environment,” IEEE J. Selected Topics Signal Process., vol. 13, pp. 172–184, Mar. 2019

    work page 2019

Show all 31 references
  1. [9]

    A Compressive Sensing Approach for the Reconstruction of the Soundfield Produced by Directive Sources in Reverberant Rooms,

    S. Damiano, F. Borra, A. Bernardini, F. Antonacci, and A. Sarti, “A Compressive Sensing Approach for the Reconstruction of the Soundfield Produced by Directive Sources in Reverberant Rooms,” IEEE/ACM Trans. Audio Speech Language Process. , vol. 32, 2024

    2024

  2. [10]

    INRAS: Implicit Neural Repre- sentation for Audio Scenes,

    K. Su, M. Chen, and E. Shlizerman, “INRAS: Implicit Neural Repre- sentation for Audio Scenes,” in Adv. Neur. Inf. Process. Syst. , (New Orleans, USA), 2022

    2022

  3. [11]

    Learning Neural Acoustic Fields,

    A. Luo and Y . Du, “Learning Neural Acoustic Fields,” in Proc. 36th Int. Conf. Neur. Inf. Process. Syst. (NeurIPS) , (New Orleans, USA), 2022

    2022

  4. [12]

    Generative models for sound field reconstruction,

    E. Fernandez-Grande, X. Karakonstantis, D. Caviedes-Nozal, and P. Ger- stoft, “Generative models for sound field reconstruction,” J. Acoust. Soc. Amer., vol. 153, pp. 1179–1190, Feb. 2023

    2023

  5. [13]

    Reconstruction of sound field through diffusion models,

    F. Miotello, L. Comanducci, M. Pezzoli, A. Bernardini, F. Antonacci, and A. Sarti, “Reconstruction of sound field through diffusion models,” in Proc. 2024 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), pp. 1476–1480, 2024

    2024

  6. [14]

    Spatial re- construction of sound fields using local and data-driven functions,

    M. Hahmann, S. A. Verburg, and E. Fernandez-Grande, “Spatial re- construction of sound fields using local and data-driven functions,” J. Acoust. Soc. Amer., vol. 150, pp. 4417–4428, Dec. 2021

    2021

  7. [15]

    Physics-informed machine learning for sound field estimation: Fun- damentals, state of the art, and challenges,

    S. Koyama, J. G. Ribeiro, T. Nakamura, N. Ueno, and M. Pezzoli, “Physics-informed machine learning for sound field estimation: Fun- damentals, state of the art, and challenges,” IEEE Signal Process. Mag., vol. 41, no. 6, pp. 60–71, 2025

    2025

  8. [16]

    Implicit neural representation with physics-informed neural networks for the reconstruction of the early part of room impulse responses,

    M. Pezzoli, F. Antonacci, and A. Sarti, “Implicit neural representation with physics-informed neural networks for the reconstruction of the early part of room impulse responses,” in Proc. 10th Conv. Europ. Acoust. Association Forum Acusticum 2023 , Jan. 2024

    2023

  9. [17]

    Sound field estimation using deep kernel learning regularized by the wave equation,

    D. Sundstr ¨om, S. Koyama, and A. Jakobsson, “Sound field estimation using deep kernel learning regularized by the wave equation,” in Proc. 2024 Int. Workshop Acoust. Signal Enhancement (IWAENC) , 2024

    2024

  10. [18]

    Physics-informed neural network for volumetric sound field reconstruction of speech signals,

    M. Olivieri, X. Karakonstantis, M. Pezzoli, F. Antonacci, A. Sarti, and E. Fernandez-Grande, “Physics-informed neural network for volumetric sound field reconstruction of speech signals,” EURASIP J. Audio Speech Music Process., vol. 2024, no. 1, p. 42, 2024

    2024

  11. [19]

    Room impulse response reconstruction with physics-informed deep learning,

    X. Karakonstantis, D. Caviedes-Nozal, A. Richard, and E. Fernandez- Grande, “Room impulse response reconstruction with physics-informed deep learning,” J. Acoust. Soc. Amer. , vol. 155, Feb. 2024

    2024

  12. [20]

    Physics- informed convolutional neural network with bicubic spline interpolation for sound field estimation,

    K. Shigemi, S. Koyama, T. Nakamura, and H. Saruwatari, “Physics- informed convolutional neural network with bicubic spline interpolation for sound field estimation,” in Proc. 2022 Int. Workshop Acoust. Signal Enhancement (IWAENC), 2022

    2022

  13. [21]

    Sound field estimation based on physics-constrained kernel interpolation adapted to environment,

    J. G. Ribeiro, S. Koyama, R. Horiuchi, and H. Saruwatari, “Sound field estimation based on physics-constrained kernel interpolation adapted to environment,” IEEE/ACM Trans. Audio Speech Language Process. , vol. 32, 2024

    2024

  14. [22]

    Sound field reconstruction using a compact acoustics-informed neural network,

    F. Ma, S. Zhao, and I. S. Burnett, “Sound field reconstruction using a compact acoustics-informed neural network,” J. Acoust. Soc. Amer. , vol. 156, Sept. 2024

    2024

  15. [23]

    A zero-shot physics-informed dic- tionary learning approach for sound field reconstruction,

    S. Damiano, F. Miotello, M. Pezzoli, A. Bernardini, F. Antonacci, A. Sarti, and T. van Waterschoot, “A zero-shot physics-informed dic- tionary learning approach for sound field reconstruction,” in Proc. 2025 IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP) , (Hyderabad,...

    2025

  16. [24]

    BINet: Learning to Solve Partial Differential Equations with Boundary Integral Networks,

    G. Lin, P. Hu, F. Chen, X. Chen, J. Chen, J. Wang, and Z. Shi, “BINet: Learning to Solve Partial Differential Equations with Boundary Integral Networks,” arXiv:2110.00352, Oct. 2021

    2021 arXiv

  17. [25]

    BI- GreenNet: Learning Green’s Functions by Boundary Integral Network,

    G. Lin, F. Chen, P. Hu, X. Chen, J. Chen, J. Wang, and Z. Shi, “BI- GreenNet: Learning Green’s Functions by Boundary Integral Network,” Commun. Math. Stat. , vol. 11, Mar. 2023

    2023

  18. [26]

    BINN: A deep learning approach for computational mechanics problems based on boundary integral equations,

    J. Sun, Y . Liu, Y . Wang, Z. Yao, and X. Zheng, “BINN: A deep learning approach for computational mechanics problems based on boundary integral equations,” Comput. Methods Appl. Mech. Eng. , vol. 410, 2023

    2023

  19. [27]

    Physics-informed boundary integral networks (PIBI-Nets): A data-driven approach for solving partial dif- ferential equations,

    M. Nagy-Huber and V . Roth, “Physics-informed boundary integral networks (PIBI-Nets): A data-driven approach for solving partial dif- ferential equations,” J. Comput. Sci. , vol. 81, p. 102355, Sept. 2024

    2024

  20. [28]

    The Boundary Element Method in Acoustics: A Survey,

    S. Kirkup, “The Boundary Element Method in Acoustics: A Survey,” Applied Sciences, vol. 9, Apr. 2019

    2019

  21. [29]

    E. G. Williams, Fourier Acoustics: Sound Radiation and Nearfield Acoustic Holography. London: Academic Press, 1999

    1999

  22. [30]

    Pyroomacoustics: A Python Package for Audio Room Simulation and Array Processing Algorithms,

    R. Scheibler, E. Bezzam, and I. Dokmanic, “Pyroomacoustics: A Python Package for Audio Room Simulation and Array Processing Algorithms,” in Proc. 2018 IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), (Calgary, AB), Apr. 2018

    2018

  23. [31]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, and others, “Pytorch: An imperative style, high-performance deep learning library,” inAdv. Neural Inf. Process. Syst. , vol. 32, 2019

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.