↓Skip to main content

cs2370 Notes: 32 AI

··2 mins 
conda install transformers timm pillow

Now to install Pytorch for the appropriate hardware

https://pytorch.org/get-started/locally/

  • For the lab computers, we pick CUDA.
  • Which version? nvcc --version
  • We need to use pip here rather than the conda command. Mixing them works.
from transformers import pipeline
import time

t0 = time.time()

qa = pipeline("question-answering")

context = """

Leonardo, the leader, is the most disciplined and skilled 
turtle; an expert
swordsman, he wields two katana and wears a blue bandana. Raphael, the
strongest and most hot-headed turtle,[26] wears a red bandana and uses a pair of
sai. Donatello uses his intellect to invent gadgets and vehicles; he
wears a purple bandana and uses a bō staff. Michelangelo is the least
disciplined and most fun-loving, and is usually portrayed as the fastest and
most agile. He wears an orange bandana and uses nunchucks.

"""

question = "Which turtle wears a red bandana?"

answer = qa(question=question, context=context)

t1 = time.time()
print("runtime:", round(t1-t0, 2))

print(answer)

Next, operating on images:

https://www.wallpaperflare.com/static/906/824/973/digital-art-wolf-moon-three-wallpaper.jpg

from transformers import pipeline
from PIL import Image

img = Image.open("/home/nat/Pictures/digital-art-wolf-moon-three-wallpaper.jpg")

od = pipeline('object-detection')
result = od(img)

print(result)

Then:

from transformers import pipeline
from PIL import Image

vqa = pipeline(model="dandelin/vilt-b32-finetuned-vqa")

img = Image.open("/home/nat/Pictures/digital-art-wolf-moon-three-wallpaper.jpg")

answer = vqa(question="How many wolves are there?", image=img)

Jupyter Notebook #

This semester we’ve written a bunch of python code. We’ve primarily written and run our programs in IDLE, but we’ve also tried running our programs from the command line. This makes a lot of sense when the point of the program is primarily to create a completed program that can then be run repeatedly.an image generation example:

Today I want to demonstrate another way to look at writing Python programs that makes sense for situations where:

  • A core purpose of the program is to have other people look at the code in order to communicate (e.g. a scientific paper).
  • The code will produce one output and the only reason to re-run it is to confirm that output.
  • You’re exploring in a way that interactive programming is useful.

New tool: Jupyter Notebook

Install and run with:

$ conda activate
$ conda install jupyter
$ jupyter-notebook

Then File -> New Notebook.

  • Cell -> Markdown
  • ‘# Hello Notebook’
  • Cell -> Code
  • x = 10 + 20
  • Cell -> Code
  • x
  • Cell -> Code
from PIL import Image
image = Image.open("duck1.jpg")
image

Another cell

image.rotate(180)

Notes

  • File -> Export as PDF
    • Deps:
    • Linux: apt-get install texlive texlive-xetex
    • Windows: choco install texlive
    • Mac: brew install texlive
  • File -> Export as HTML