This is a collection of random data-science exercises either gathered from different articles or videos that I have used for learning.
I setup an arguably complicated Makefile structure since I have so many projects in here.
Here is a demonstration of how it works for the related building_ml_powered_apps project:
Here are some of the included projects:
- building_ml_powered_apps/
This project is built from the Building Machine Leaning Powered Applications Going From Idea to Product Python book. Here is the related github repo
- complete_data_science_bootcamp/
The following directory has examples for learning data science / machine learning based on this Udemy course
- derivatives-with-python/
The following examples come from the Derivatives Analytics With Python book. Here is the related github repo
- financial_analyst_course_2023/
The following directory has examples for learning financial analysis based on this Udemy course
The course is meant to work in excel but I have other plans...
- hugging_face_course/
This project follows the hugging face course which provides building and tweaking various types of NLP pipelines. It also exposes a little bit of gradio so you can have a way to showcase your models!
- pycon2020-examples/
This collection of scripts is based off Keith Galli's awesome PyCon2020 NLP youtube video. It runs through various NLP techniques and serves as a good introduction with some guidance for starting off your ML journey
- python_data_science_december_2022/
These are a collection of activities that do some basic data analysis on a wide variety of sources
- scripts/
one off examples usually found from blog posts demonstrating a single topic
gum is a library to make good-looking & interactive shell scripts. I use it on some of the scripts in here.
I have included the installation for my system below:
sudo mkdir -p /etc/apt/keyrings
curl -fsSL https://repo.charm.sh/apt/gpg.key | sudo gpg --dearmor -o /etc/apt/keyrings/charm.gpg
echo "deb [signed-by=/etc/apt/keyrings/charm.gpg] https://repo.charm.sh/apt/ * *" | sudo tee /etc/apt/sources.list.d/charm.list
sudo apt update && sudo apt install gumDuring the creation of some of the projects / examples I have encountered some issues when working with my recent version of python 11. For instance tensorflow will not be happy with you and as a result it's probably a good reason to start to learn pytorch.
Anyways opinions aside (this is programming right??) here are some notes I have on some libraries having trouble with my setup.
In order to try and get umap-learn I needed to get a version of llvm working.
In my attempt it looks like the library is still too old to work with python 3.11 =(
$ poetry add umap-learn
FileNotFoundError: [Errno 2] No such file or directory: 'llvm-config'
RuntimeError: llvm-config failed executing, please point LLVM_CONFIG to the path for llvm-config
$ apt-cache search "llvm-.*-dev" | grep -v ocaml | sort
llvm-11-dev - Modular compiler and toolchain technologies, libraries and headers
$ sudo apt install llvm-11-dev
$ ls /usr/bin/llvm-config*
/usr/bin/llvm-config-11
$ LLVM_CONFIG=/usr/bin/llvm-config-11 poetry add umap-learn
RuntimeError: Building llvmlite requires LLVM 10.0.x or 9.0.x, got '11.1.0'. Be sure to set LLVM_CONFIG to the right executable path.TODO: Add this to Makefile Line count of files in a given directory
find ./quantitative_finance_algorithmic_trading -name '*.py' | xargs wc -l | tail -n1
find ./machine_learning_and_deep_learning_bootcamp -name '*.py' | xargs wc -l | tail -n1Here is a pandas cookbook that would be good for adding some examples from into this repo...
An interesting thing to check out could be this python profiler py-spy which allows you to profile and debug any running python program, even if it is serving production traffic
Face recognition API in python
Deep learning lessons from Fast AI
Interesting transaction manager called pyWave
Tool called Oxen for versioned datasets
Interesting hacker news article talks about challenges they have with AI in development scenarios. It sounds like it is mostly frustrations due to deployment / infrastructure and rapid prototyping
We serve our models with FastAPI, containerize them, and then deploy them to our GKE clusters. Depending on the model, we choose different machines - some require GPUs, most are decent on CPU. We take models up or down based on usage, so we have cold starts unless otherwise specified by customers. We expose access to the model via a POST call through our cloud app. We track inputs and outputs, as we expect that people will become interested in fine tuning models based on their past usage.
For the original "davinci" models (now 3 generations behind if you count Instruct, ChatGPT, and upcoming DV"), OpenAI recommends "Aim for at least ~500 examples" as a starting point for fine-tuning
How do you monitor and debug models?
When your engineering team builds great tooling, monitoring and debugging get much easier. Stitch Fix has built an internal tool that takes in a modeling pipeline and creates a Docker container, validates arguments and return types, exposes the inference pipeline as an API, deploys it on our infrastructure, and builds a dashboard on top of it. This tooling allows data scientists to directly fix any errors that happen during or after deployment.
How do you deploy new model versions?
In addition, data scientists run experiments by using a custom-built A/B testing service that allows them to define granular parameters. They then analyze test results, and if they are deemed conclusive by the team, they deploy the new version themselves.
When it comes to deployment, we use a system similar to canary development where we start by deploying the new version to one instance and progressively update instances while monitoring performance. Data scientists have access to a dashboard that shows the number of instances under each version and continuous performance metrics as the deployment progresses.