I couldn’t stand everyone talking about GPT this, GPT that, GPT here, there, everywhere and not put GPT on my GoPiGo3.
(Sorta - it is running on the Pi4 2GB processor of my “ROSbot GoPiGo3” over Ubuntu 22.04 jammy jellyfish, and Python3.10.6, but is not integrated…)
Steps:
1) Create a directory for miscellaneous examples: mkdir misc; cd misc
2) Bring down "picoGPT" code: git clone https://github.com/jaymody/picoGPT.git
3) Install required elements: cd picoGPT; pip install -r requirements.txt
4) Run it the first time primarily to download models:
$ python3 gpt2.py "GPT is going to"
5) Now run it with a timer and htop:
$ time python3 gpt2.py "I love my robot, but my robot thinks" --n_tokens_to_generate 20
2023-03-16 22:34:14.672873: W tensorflow/tsl/framework/cpu_allocator_impl.cc:82] Allocation of 154389504 exceeds 10% of free system memory.
generating: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:13<00:00, 1.51it/s]
I'm a robot. I'm not. I'm not a robot. I'm not a robot
real 0m22.468s
user 0m57.425s
sys 0m3.347s
RESULT: “I love my robot, but my robot thinks I’m a robot. I’m not. I’m not a robot. I’m not a robot”
It used 1.4GB of virtual memory, 800MB of physical memory, all four cores at 100% for about 24 seconds. The “1.51it/s” means it generated 1.5 words per second (about as fast as I think before noticing it is past my bedtime.)
More on the options in the README:
If you are interested in understanding the technology, check out:
Indeed theories about language are embodied in various ways in our computing. Language has even been proposed as one of the necessary elements to define life by some.
Much of my career involved building language generation and language understanding systems based on language models either of my creation as a user or from large data collection efforts by giving out movie vouchers in exchange for curated “native speech”.
Today’s language generation and language understanding systems are based on curated data from the Internet with the ability to cover and uncover more regional diversity. Additionally, these systems are self programming that adapt to the input in microscopic detail, where my systems required specifically detailed examples for me to create exception cases.
My systems were all linear and my approach was one problem at a time. Today’s systems are parallel that approach “everything everywhere all at once”, (which might explain why I couldn’t get into EEAAO as a movie).