License about fft-conv-pytorch HOT 8 CLOSED

you got it!

from fft-conv-pytorch.

@hmaarrfk All of my projects are MIT license
https://github.com/fkodom/fft-conv-pytorch/blob/master/LICENSE

from fft-conv-pytorch.

Why do you need to package it for conda-forge? This library is already pip installable:

• https://github.com/fkodom/fft-conv-pytorch#install
• https://pypi.org/project/fft-conv-pytorch/

from fft-conv-pytorch.

Sorry, i was blind. I guess the project i'm using is a fork of yours, I was presumed that they had also included your license
https://github.com/yoyololicon/fft-conv-pytorch/

maybe it wasn't forked when you added your license.

from fft-conv-pytorch.

Why do you need to package it for conda-forge? This library is already pip installable:

While you can use pip+conda together having everything managed by conda helps with updates and maintainability.

conda(-forge) really helps manage packages that have C dependencies between them. For example, you can have opencv depend on ffmpeg and HDF5.

Ultimately, if you are to do this with pip, you need to have every single pip package recompile HDF5 itself to built it in with HDF5 support.

in my opinion, conda(-forge), lowers the barrier to integrating C libraries with python, making it an overall stronger scientific development environment.

Small example:

mamba create --name opencv opencv python=3.10

In [1]: import cv2
In [2]: cv2.hdf
<module 'cv2.hdf'>

pip install opencv-python

In [1]: import cv2

In [2]: cv2.hdf
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[2], line 1
----> 1 cv2.hdf

AttributeError: module 'cv2' has no attribute 'hdf'

Again, Thank you for pointing me to the obvious place where your license would be.... sorry for the bother on that front.

Best,

Mark

from fft-conv-pytorch.

Interesting -- is that a specific issue with the Conda opencv package? I have installed and used opencv-python, ffmpeg, and h5py together with pip many times. Maybe the cv2.hdf module does something additional that I'm missing?

I suppose this is a more generic request than cv2.hdf, though. I'm curious how often this is an issue in practice?

from fft-conv-pytorch.

Note that the conda opencv package supports cv2.hdf , while the pip does not.

Using pip, you get to to communicate between all these different modules through python.

So your data must pass effectively through something like a numpy array (this isn't so bad, its just a wrapper around a strided piece of data)

So you can load your data through h5py, coerce it into a numpy array, then pass it to opencv-python.

But if you have more control over how opencv is built for your application, you could for example simply ask opencv to open the HDF5 directly. The HDF5 usecase below is somewhat minor. but you can think of it as "as feature of opencv was traded off to simplify packaging, or to make the installable smaller".

So the recourse is a few fold:

1. Ask nicely for the person that packaged opencv for pip to include HDF5 support. This will take a while.
2. Use the conda-forge package if it supports what you need
3. Create your own conda package compiled with the options you need and upload it to your conda channel.

This last option, is only available with conda, where the concept if channels exists from the start. I know that pip can have sources, but the website and infrastructure created by Anaconda (or binstar) makes it really easy for others to create their own channels.

I'm curious how often this is an issue in practice?

This is difficult to measure. Indeed, going with route 1 tends to in the long term, work out, but I'm guilty of recompiling programs to be compatible with lesser used features for performance enhancements while upstream takes action.

The ability to have your own (hosted) channel gives you "patience"

One concrete example, is the availability of codecs, I used this little snippet of code to test things:
https://stackoverflow.com/a/76173072/2321145

import cv2
from pprint import pprint

def is_fourcc_available(codec):
    try:
        fourcc = cv2.VideoWriter_fourcc(*codec)
        temp_video = cv2.VideoWriter('temp.mkv', fourcc, 30, (640, 480), isColor=True)
        return temp_video.isOpened()
    except:
        return False

def enumerate_fourcc_codecs():
    codecs_to_test = ["DIVX", "XVID", "MJPG", "X264", "WMV1", "WMV2", "FMP4",
                      "mp4v", "avc1", "I420", "IYUV", "mpg1", "H264"]
    available_codecs = []
    for codec in codecs_to_test:
        available_codecs.append((codec, is_fourcc_available(codec)))
    return available_codecs

if __name__ == "__main__":
    codecs = enumerate_fourcc_codecs()
    print("Available FourCC codecs:")
    pprint(codecs)

Pip:

Available FourCC codecs:
[('DIVX', True),
 ('XVID', True),
 ('MJPG', True),
 ('X264', False),
 ('WMV1', True),
 ('WMV2', True),
 ('FMP4', True),
 ('mp4v', True),
 ('avc1', False),
 ('I420', True),
 ('IYUV', True),
 ('mpg1', True),
 ('H264', False)]

conda-forge:

Available FourCC codecs:
[('DIVX', True),
 ('XVID', True),
 ('MJPG', True),
 ('X264', True),
 ('WMV1', True),
 ('WMV2', True),
 ('FMP4', True),
 ('mp4v', True),
 ('avc1', True),
 ('I420', True),
 ('IYUV', True),
 ('mpg1', True),
 ('H264', True)]

How valuable is using x264, avc1, h264 through the opencv API for you? That only you can decide.

But the challenge of using opencv + ffmpeg, or opencv + hdf5, or any pairwise combination (dare I say 3 library combination!) is very difficult in my experience, to do using just pip.

from fft-conv-pytorch.

Ok, I think I understand. Sounds like conda "channels" effectively take the place of a private PyPI server (or pip source)?

For my understanding -- could the issue also be solved by building from source? If so, that feels more familiar to me. (My instinct is to build from source, and containerize anything that needs to be reused or run on a remote machine.)

from fft-conv-pytorch.