Unexpected warning (about resizing) from torch_matrix_exp() only when input tensor values are < 1, despite output being correct

[maspotts]

I've encountered an unexpected warning during torch_matrix_exp() which is triggered only when the argument (x) has values in a range (< 1) which causes print(x) to display in "scientific notation" mode (ie. by printing an exponential scaling factor (eg. 0.001 *) before printing the columns). This is consistent across 1000s of calls to torch_matrix_exp(): scientifically-notated inputs x always trigger the warning; non-scientifically-notated inputs x never trigger the warning. The warning is:

[W Resize.cpp:24] Warning: An output with one or more elements was resized since it had shape [1, 16, 16], which does not match the required output shape [1, 1, 16, 16]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (function resize_output_check)

(my inputs are 16x16 tensors). Note: despite the warning, the calculated matrix exponential is still correct (when compared with the output of expm::expm() applied to the corresponding matrix (non-tensor). I'm using torch 0.9.1 under R 4.2.2.

Code to replicate the problem:

a <- matrix(c(1, 2, 5, 3), ncol = 2)
aa <- torch_tensor(a)
b <- a * 1e-3
bb <- torch_tensor(b)
print("a: Not scientific, no warning:")
aa
torch_matrix_exp(aa)
print("b: Scientific, triggers warning:")
bb
torch_matrix_exp(bb)

Output:

[1] "a: Not scientific, no warning:"
> aa
torch_tensor
 1  5
 2  3
[ CPUFloatType{2,2} ]
> torch_matrix_exp(aa)
torch_tensor
  71.3138  153.3390
  61.3356  132.6494
[ CPUFloatType{2,2} ]
> print("b: Scientific, no warning:")
[1] "b: Scientific, triggers warning:"
> bb
torch_tensor
0.001 *
 1.0000  5.0000
 2.0000  3.0000
[ CPUFloatType{2,2} ]
> torch_matrix_exp(bb)
[W Resize.cpp:24] Warning: An output with one or more elements was resized since it had shape [1, 2, 2], which does not match the required output shape [1, 1, 2, 2]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (function resize_output_check)
torch_tensor
 1.0010  0.0050
 0.0020  1.0030
[ CPUFloatType{2,2} ]

I'm concerned that this warning may turn into an error in a future release, and obviously would prefer not to have to worry about an (apparently spurious) warning based on the scale of the inputs. Any ideas?

Session info:

R version 4.2.2 (2022-10-31)
Platform: aarch64-apple-darwin20 (64-bit)
Running under: macOS Ventura 13.1

Matrix products: default
BLAS:   /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRblas.0.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/4.2-arm64/Resources/lib/libRlapack.dylib

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats4    graphics  datasets  stats     utils     grDevices methods
[8] base

other attached packages:
 [1] torch_0.9.1      expm_0.999-6     Matrix_1.5-1     withr_2.5.0
 [5] jsonlite_1.8.3   bbmle_1.0.25     optimx_2022-4.30 assertthat_0.2.1
 [9] plotly_4.10.1    patchwork_1.1.2  scales_1.2.1     gridExtra_2.3
[13] lattice_0.20-45  deSolve_1.34     glue_1.6.2       rlist_0.4.6.2
[17] fuzzyjoin_0.1.6  broom_1.0.1      ggvis_0.4.7      ggplot2_3.4.0
[21] forcats_0.5.2    hms_1.1.2        lubridate_1.9.0  timechange_0.1.1
[25] stringr_1.4.1    readr_2.1.3      purrr_0.3.5      tidyr_1.2.1
[29] dplyr_1.0.10     magrittr_2.0.3   tibble_3.1.8

loaded via a namespace (and not attached):
 [1] Rcpp_1.0.9          bdsmatrix_1.3-6     mvtnorm_1.1-3
 [4] ps_1.7.2            digest_0.6.30       utf8_1.2.2
 [7] mime_0.12           R6_2.5.1            backports_1.4.1
[10] httr_1.4.4          pillar_1.8.1        rlang_1.0.6
[13] lazyeval_0.2.2      data.table_1.14.6   callr_3.7.3
[16] htmlwidgets_1.5.4   bit_4.0.5           munsell_0.5.0
[19] shiny_1.7.3         compiler_4.2.2      numDeriv_2016.8-1.1
[22] httpuv_1.6.6        pkgconfig_2.0.3     htmltools_0.5.3
[25] tidyselect_1.2.0    fansi_1.0.3         viridisLite_0.4.1
[28] tzdb_0.3.0          later_1.3.0         MASS_7.3-58.1
[31] grid_4.2.2          xtable_1.8-4        gtable_0.3.1
[34] lifecycle_1.0.3     DBI_1.1.3           coro_1.0.3
[37] cli_3.4.1           stringi_1.7.8       promises_1.2.0.1
[40] ellipsis_0.3.2      generics_0.1.3      vctrs_0.5.1
[43] tools_4.2.2         bit64_4.0.5         processx_3.8.0
[46] fastmap_1.1.0       colorspace_2.0-3

[skeydan]

Hi,
thanks for asking! This is an internal message originating from libtorch (so, not on the R side).
In fact, this message goes away in the next libtorch version (1.13.1), which we're adding support for right now (#977).

I've checked this in Python:

PyTorch 1.12.1

Python 3.10.9 (main, Dec 20 2022, 00:00:00) [GCC 12.2.1 20221121 (Red Hat 12.2.1-4)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> a = torch.tensor((2., 3., 4., 5.)).reshape(2,2)
>>> torch.matrix_exp(a)
tensor([[ 435.5261,  573.3393],
        [ 764.4525, 1008.8654]])
>>> b = a * 1e-3
>>> torch.matrix_exp(b)
<stdin>:1: UserWarning: An output with one or more elements was resized since it had shape [1, 2, 2], which does not match the required output shape [1, 1, 2, 2]. This behavior is deprecated, and in a future PyTorch release outputs will not be resized unless they have zero elements. You can explicitly reuse an out tensor t by resizing it, inplace, to zero elements with t.resize_(0). (Triggered internally at  ../aten/src/ATen/native/Resize.cpp:17.)
tensor([[1.0020, 0.0030],
        [0.0040, 1.0050]])

Pytorch 1.13.1

Python 3.11.1 (main, Jan  6 2023, 00:00:00) [GCC 12.2.1 20221121 (Red Hat 12.2.1-4)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> a = torch.tensor((2., 3., 4., 5.)).reshape(2,2)
>>> torch.matrix_exp(a)
tensor([[ 435.5261,  573.3393],
        [ 764.4525, 1008.8654]])
>>> b = a * 1e-3
>>> torch.matrix_exp(b)
tensor([[1.0020, 0.0030],
        [0.0040, 1.0050]])
>>> 

[maspotts]

Great to know: thanks very much!