Create a smaller production pip install?

[crearo]

I want to run an inference on heroku where I've already hosted my website.
Everything worked great on my laptop, but as soon as I push to Heroku, it complains about its maximum size limit being 500MB.

Here's all I've done:

1. pip install fastai
2. pip freeze > requirements.txt # fastai internally installs so many more pypis
3. git push heroku master

Aaaand, heroku fails. Specifically, it redirects me to the slug-size page.

Can we have a separate install for inferencing only? I know PyTorch comes with a CPU only build that's 45MB and is meant for production.

[stas00]

Here are the instructions that let you choose and pick what fastai dependencies you want to install:
https://docs.fast.ai/install.html#custom-dependencies

But dependencies of dependencies is beyond of fastai control.

[crearo]

Thanks! It's still going over the roof even with the CPU build, and I'm kinda clueless at this point.
If I come across something I'll put it up here.

The suggestions on the fastai forums didn't help either. https://forums.fast.ai/t/hosting-a-simple-web-application/13696/19

[crearo]

Alright! The trick was to just type:

fastai
https://download.pytorch.org/whl/cpu/torch-1.0.1.post2-cp36-cp36m-linux_x86_64.whl

in requirements.txt.

I was doing a pip freeze > requirements.txt which does not work. I'm still
This still is problematic for me because every time I pip install something I have to be extra careful to put it in the requirements.txt file manually.