ENH+FIX to_conv1d() and energy bugfix in MuReNNDirect

[xir4n]

closes #45, resolves #47 , resolves #52

Update:

• fix bug: normalization behavior inside dtcwt.transform1d
• new feature: MuReNNDirect.to_conv1d
• fix bug: normalization setting in MuReNNDirect

[lostanlen]

Thank you but this implementation returns an array. I would prefer if to_conv1d() could return a fully functional Conv1D module whose weight.data contains the array you have. The padding should be the same.

Then in tests/ you can check that conv1d(x) is almost equal to murenn(x)

OK?

[xir4n]

ok, I think I've got what you mean

[lostanlen]

Following our conversation today, this is both ENH and FIX now. Let me know when it's ready so I can review.

[xir4n]

@lostanlen What do you think I should do? Can I specify the numpy version in the continuous integration script to make this PR pass the checks?

[lostanlen]

tests are passing on this environment

(murenn) ➜  murenn git:(fafe691) conda list
# packages in environment at /Users/user/miniconda3/envs/murenn:
#
# Name                    Version                   Build  Channel
bzip2                     1.0.8                h6c40b1e_6
ca-certificates           2024.3.11            hecd8cb5_0
dtcwt                     0.14.0                   pypi_0    pypi
exceptiongroup            1.2.1                    pypi_0    pypi
filelock                  3.15.3                   pypi_0    pypi
fsspec                    2024.6.0                 pypi_0    pypi
iniconfig                 2.0.0                    pypi_0    pypi
jinja2                    3.1.4                    pypi_0    pypi
libffi                    3.4.4                hecd8cb5_1
markupsafe                2.1.5                    pypi_0    pypi
mpmath                    1.3.0                    pypi_0    pypi
murenn                    0.3.dev0                  dev_0    <develop>
ncurses                   6.4                  hcec6c5f_0
networkx                  3.3                      pypi_0    pypi
numpy                     1.26.4                   pypi_0    pypi
openssl                   3.0.14               h46256e1_0
packaging                 24.1                     pypi_0    pypi
pip                       24.0                     pypi_0    pypi
pluggy                    1.5.0                    pypi_0    pypi
pytest                    8.2.2                    pypi_0    pypi
python                    3.10.14              h5ee71fb_1
readline                  8.2                  hca72f7f_0
scipy                     1.13.1                   pypi_0    pypi
setuptools                69.5.1                   pypi_0    pypi
six                       1.16.0                   pypi_0    pypi
sqlite                    3.45.3               h6c40b1e_0
sympy                     1.12.1                   pypi_0    pypi
tk                        8.6.14               h4d00af3_0
tomli                     2.0.1                    pypi_0    pypi
torch                     2.2.2                    pypi_0    pypi
typing-extensions         4.12.2                   pypi_0    pypi
tzdata                    2024a                h04d1e81_0
wheel                     0.43.0                   pypi_0    pypi
xz                        5.4.6                h6c40b1e_1
zlib                      1.2.13               h4b97444_1

[lostanlen]

Collecting numpy<2.0.0 (from dtcwt>=0.13.0->murenn==0.3.dev0)
  Downloading numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (61 kB)

😻

[lostanlen]

thanks. i just have a question on the zero-ing of complex-valued coefficients

[lostanlen]

why convert to complex? and why multiply by 1j?

[xir4n]

I converted them to complex because you wanted the to_conv1d method to return a complex-valued PyTorch tensor in #45.

Besides, in dtcwt, we have: (I've uploaded a test to show this. )

$$ \begin{align} idtcwt(x) &= idtcwt.tree_a(x) + idtcwt.tree_b(x)\\ &= idtcwt(x.real) + idtcwt(x.imag) \end{align} $$

I think in the returned Conv1d instance, conv1d.weight.real+conv1d.weight.imag is equivalent to the case if we don't convert Wpsi_j to complex

there is a bug: maybe i should use $x=(1-1j)*\delta$ instead of $x=(1+0j) * \delta$ to convolve with conv1d to visualize the filter

I can change this into a real-valued version if this is too complicated

[lostanlen]

Hello, i don't think $x = (1 - 1j) \delta$ is necessary. $x = \delta$ should suffice.

My question was not a mathematical question but more like a CS question. Usually when you do
z = torch.complex(re, im)
re and im are both real-valued. Here this is not the case, since you have multiplied im by 1j.

[xir4n]

Thanks for your advice, I've made the change.
I was not clear in my previous answer. In the algorithm, only the real part of a complex tensor would be passed through Tree A, and only the imaginary part would be passed through Tree B. That's why I was converging z into a complex-valued tensor.

[lostanlen]

closed because of too high complexity
converted into new PR's: --> #55 and #56