How to port the Huggingface🤗 into Julia using PyCall?

#machinelearning #ml #huggingface #julia

Ready for a fast and concise delicious Julia thing? Read to the end, you'll be surprised.
First of all, I should mention that you can port any python package into Julia with zero overhead! This means you won't face performance reduction at all!

First, import the PyCall and connect it into an executable Python (I go for python3.10 since the huggingface🤗 isn't available for 3.11 so far):

julia> using PyCall

julia> ENV["PYTHON"] = raw"C:\Users\<USER>\AppData\Local\Programs\Python\Python310\python.exe";

julia> using Pkg

julia> Pkg.build("PyCall")

After a successful build, exit from the Julia and re-run it. Then check your python environment is changed successfully:

julia> PyCall.python
"C:\\Users\\<USER>\\AppData\\Local\\Programs\\Python\\Python310\\python.exe"

Good! I'm on Python3.10 now. Then, open cmd and install the huggingface on the same python path via pip install transformers[torch]. Then start a fresh Julia session and continue:

julia> using PyCall

julia> ts = pyimport("transformers");

julia> gen = ts.pipeline(task="text-generation")

This leads to downloading the requirements (~550MB). Then let's get a text generator:

julia> gen("Hello from Julia, Julia is a lovely programming language that is fast as C in many cases, and readable as Python in everycases.")
Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.
1-element Vector{Dict{Any, Any}}:
 Dict("generated_text" => "Hello from Julia, Julia is a lovely programming language that is fast as C in many cases, and readable as Python in everycases. Julia is a great way of creating beautiful applications and helping you write beautiful code, and I love using it everyday!")