Get rid of scipy

[eli-osherovich]

Please make sure that the boxes below are checked before you submit your issue.
If your issue is an implementation question, please ask your question on StackOverflow or on the Keras Slack channel instead of opening a GitHub issue.

Thank you!

• Check that you are up-to-date with the master branch of keras-preprocessing. You can update with:
  pip install git+git://github.com/keras-team/keras-preprocessing.git --upgrade --no-deps

• Provide a link to a GitHub Gist of a Python script that can reproduce your issue (or just copy the script here if it is short).

From my experience, relying on SciPy for image transformation makes it:
a) slow
b) non-suitable for TPU

Is there any particular reason to keep SciPy in the loop?

[Stanfording]

For the m1 apple chip, SciPy is not available yet, but the ImageDataGenerator.flow_from_directory requires it. So getting rid of scipy would be a temporary relief for the m1 users. Maybe not necessary though ( since SciPy is working on it too I guess).

[eli-osherovich]

I can do this if there is a real need from the keras side.

[Dref360]

We use scipy here for SVD (I think np.linalg.svd is called?)
here

And we use scipy.ndimage for matrix multiplication here:
ndimage.interpolation.affine_transform.

Do you think it would require a lot of work to replicate affine_transform behaviour?

[eli-osherovich]

No. It should be pretty simple.

…

On Thu, Jan 14, 2021, 4:30 PM Frédéric Branchaud-Charron < ***@***.***> wrote: We use scipy here for SVD (I think np.linalg.svd is called?) here <https://github.com/keras-team/keras-preprocessing/blob/54b5d6ed65f1608ca7207d585075c2ad3a52929e/keras_preprocessing/image/image_data_generator.py#L986> And we use scipy.ndimage for matrix multiplication here: ndimage.interpolation.affine_transform <https://github.com/keras-team/keras-preprocessing/blob/54b5d6ed65f1608ca7207d585075c2ad3a52929e/keras_preprocessing/image/affine_transformations.py#L387> . Do you think it would require a lot of work to replicate affine_transform behaviour? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#322 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AASS73X5YPPFSVXBJFJPWATSZ35YPANCNFSM4TGMG7JQ> .

[Dref360]

PRs welcome then :)

[eli-osherovich]

Will send some shortly

…

On Fri, Jan 15, 2021, 6:16 PM Frédéric Branchaud-Charron < ***@***.***> wrote: PRs welcome then :) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#322 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AASS73UOJAEUXWZTKVEXEHLS2BS5NANCNFSM4TGMG7JQ> .

[Stanfording]

I can do this if there is a real need from the keras side.

Yes, please! Because scipy is important to Keras and the machine learning community.

[eli-osherovich]

Yes, please! Because scipy is important to Keras and the machine learning community.

Just to make sure: the idea is to remove scipy dependency from keras. Namely, it will not make scipy working on M1, we'll just replace it with something else so as to let keras be used on M1.

[Stanfording]

Yes, please! Because scipy is important to Keras and the machine learning community.

Just to make sure: the idea is to remove scipy dependency from keras. Namely, it will not make scipy working on M1, we'll just replace it with something else so as to let keras be used on M1.

That's exactly what I meant!

[eli-osherovich]

@Stanfording first patch is waiting for approval, as you can see. I chose to replace scipy with numpy, but I cannot test it on M1.... Actually, after thinking about it a bit, I do not know if numpy is working on M1...

[Stanfording]

@Stanfording first patch is waiting for approval, as you can see. I chose to replace scipy with numpy, but I cannot test it on M1.... Actually, after thinking about it a bit, I do not know if numpy is working on M1...

There are some issues with numpy if just downloading by itself. However, numpy works with tensorflow if downloading tensorflow with this link https://github.com/apple/tensorflow_macos . So many people who cannot download numpy on m1 just download the TensorFlow from the link above, which will install numpy automatically. The version of this numpy is currently 1.18.5. Python version is 3.8.2 from xcode command line tool. The way to use keras on m1 is mostly use "from tensorflow import keras" instead of downloading keras directly and just import keras. So if there is any update from keras, we can only use it from the one imported from tensorflow. I hope the information helps.

[Stanfording]

@Stanfording first patch is waiting for approval, as you can see. I chose to replace scipy with numpy, but I cannot test it on M1.... Actually, after thinking about it a bit, I do not know if numpy is working on M1...

However, it is possible for m1 users to download keras directly. But it's easy to run into many issues. I spent a lot of time to solve issues and finally get keras downloaded successfully in the same virtual environment with tensorflow. However, when using it, I could just type "python3" and "import keras" on the terminal in the virtual environment, which imported successfully, but if I import keras in a text editor (spyder) and use "python3 xxx.py" in virtual environment, it will say "No module named 'keras'". If this problem is impossible to solve by users, then the only way we can get keras is this link: https://github.com/apple/tensorflow_macos. I hope this is also helpful.