Support of Volumetric Data

Hello,

I saw such an opinion when creating a model.

Which volumetric data do you support? (binvox or off)

volumetric

Thank you.

Hi Ahmet,
Any dataset which can be saved and read in numpy can be used as volumetric dataset.

If you have 3D data (video, DIACOM or any other 3D data), you can first encode each record as Numpy.

Create train.csv (as you do for any image dataset but instead of image file, it will be a path to numpy file) and upload it.

When you used AutoML with volumetric input type, it will generate network which uses Convolutional 3D layers.

I seem to be doing something wrong here.

I have saved my 3D volumes of numpy.array()-format images using numpy.save(filename, array), created the CSV with filenames and classes, but get an error, regardless if I select numpy or image or array as a datatype.

Selecting image as my type. This will then result in an “OSerror: cannot identify image file”. This is an error raised from PIL/Image.py.

Using numpy as type, it will be “AttributeError: ‘numpy.ndarray’ object has no attribute ‘files’” which directly comes from project.network.get_port_shape.

Any help appreciated…

It may be an issue with how you are saving your NumPy file or the extension of the NumPy file. Chose “.npz” as the file extension.

See the below post for reference on how to save the npz file

Thanks!
Well, this sounds a little more complicated than necessary, at least at first. Of course it’s sensible to use compressed files to save disk space. But it leaves one wondering why the simplest possible way – np.save() – would lead to an error…
Anyways, gonna try!
Any chance to document such things somehow, somewhere? There isn’t accidentially a Wiki where users may help you in providing such documentation?

Hi
You can refer to these 2 articles based on 3-d data.
https://towardsdatascience.com/recognition-of-action-in-video-with-deep-learning-studio-9ffad2283080

https://towardsdatascience.com/3-d-mnist-with-deep-learning-studio-d73af9504369

Regards
Rajat