Installing SpaCy 3.1.X and latest torchtext on Jetson Nano

Hello,

I have tried a few links to no avail on how to install SpaCy 3.1.X and torchtext for GPU on Jetson Nano.

I have tried using the following script:

Which hasn’t worked and tried installing a few wheels from here

To install SpaCy and torchtext for GPU

I was wondering if there were some pre built wheels to install SpaCy and torchtext or if I’m missing something more fundamental in my approach on trying to install?

You don’t give specific resson why the gist-script script failed, but the scripts uses Pytorch 1.4 - which iirc is no longer supported by latest Jetpack. You may try a newer Pytorch version instead.

Sorry, I should have provided more detail earlier, I have also tried building torchtext and SpaCy from the repositories and the installation for SpaCy failed which required thinc package which is provided by the wheel in the description.

However, after attempting to install thinc with the wheel, the version was too low, so decided to compile myself using the repository which failed to install. The python version (3.6.9) I used was the one installed with the image Jetpack 4.6 (I believe the latest version of time of writing)

And it add clarity, Pytorch and torchvision installed perfectly fine which is great but it’s torchtext and SpaCy that I have issues with.

Hi,

Sorry that we don’t have a prebuilt package for SpaCy and Torchtext.
Would you mind checking with the SpaCy team to see if they support ARM and python 3.6 environments?

Thanks.

I have made contact with the spaCy team and got the following response.

Hi,

Could you share the issue ID or exact page with us?
The link points to the spaCy forum rather than a specific discussion.

Thanks.

Hello,

Sorry, I thought the link pasted ok - Installing SpaCy with cuda on Jetson Nano. · Discussion #9367 · explosion/spaCy · GitHub

To round up, use conda forge and aarch installer from the link in the discussion and you can install spaCy.

However, I do get an error when testing if cuda is available.

But this is to do with cuda and thinc package - dependencies of spaCy.

Later on I’ll try different environments and see if that helps but this is a lot closer than I’ve managed to get before.