AI Cheat Sheet

Tools and tricks to make AI development faster and simpler

Author: bluelul.com

Python profiler with graph

Setup library

sudo apt install graphviz

pip install gprof2dot

Analyse Python program main.py and save profiling result in profile.pstats

python -m cProfile -o profile.pstats main.py

Plot profiling result to graph as svg file

gprof2dot -f pstats profile.pstats | dot -Tsvg -o main_profiled.svg

OpenCV CPP environment installation

Pre-install

Verify your Ubuntu version

Ubuntu 23.04 -> OpenCV 4.6.0
Ubuntu 22.04 -> OpenCV 4.5.4
Ubuntu 21.04 -> OpenCV 4.5.1
Ubuntu 20.04 -> OpenCV 4.2.0
Ubuntu 18.04 -> OpenCV 3.2.0
Ubuntu 16.04 -> OpenCV 2.4.9.1

Install cpp build tools (just in case on fresh Ubuntu)

sudo apt update
sudo apt install build-essential

Main install

Install pkg-config for easier to make cpp without declaring tons of flag

sudo apt install pkg-config

Install prebuilt opencv cpp lib (opencv-contrib included)

sudo apt install libopencv-dev

Test

Create new simple opencv-called cpp file mainprog.cpp

#include <opencv2/opencv.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/tracking.hpp> // only in contrib

using namespace cv;
using namespace std;

int main() {
    cout << "The current OpenCV version is " << CV_VERSION << "\n";
    return 0;
}

Build

• For OpenCV version 4

  g++ mainprog.cpp -o mainprog `pkg-config --cflags --libs opencv4`

• For OpenCV version 2, 3

  g++ mainprog.cpp -o mainprog `pkg-config --cflags --libs opencv`

Run

./mainprog

Crawl image from internet

More on https://icrawler.readthedocs.io/en/latest/builtin.html

from icrawler.builtin import BaiduImageCrawler, BingImageCrawler, GoogleImageCrawler

google_crawler = GoogleImageCrawler(
    feeder_threads=1,
    parser_threads=1,
    downloader_threads=4,
    storage={'root_dir': 'your_image_dir'})
filters = dict(
    size='large',
    color='orange',
    license='commercial,modify',
    date=((2017, 1, 1), (2017, 11, 30)))
google_crawler.crawl(keyword='cat', filters=filters, offset=0, max_num=1000,
                     min_size=(200,200), max_size=None, file_idx_offset=0)

bing_crawler = BingImageCrawler(downloader_threads=4,
                                storage={'root_dir': 'your_image_dir'})
bing_crawler.crawl(keyword='cat', filters=None, offset=0, max_num=1000)

baidu_crawler = BaiduImageCrawler(storage={'root_dir': 'your_image_dir'})
baidu_crawler.crawl(keyword='cat', offset=0, max_num=1000,
                    min_size=(200,200), max_size=None)

YOLO infer

• Install YOLO

pip install ultralytics

• Detect objects and crop results to another folder inside execute_path/yolo/ folder

yolo detect predict model=yolov8x.pt save_crop project='yolo' source='input_folder_or_file'

Distinct integer ID numbers by BGR color

def get_color(number):
    """ Converts an integer number to a color """
    blue = int(number*50 % 256)
    green = int(number*30 % 256)
    red = int(number*103 % 256)

    return blue, green, red

LLM Benchmark

MT-Bench

• Multi-turn open-ended questions
• Radar plot
• Bench field: Writing, Roleplay, Reasoning, Math, Coding, Extraction, STEM, Humanities

Vocal Remover and Isolation

https://vocalremover.org/

ONNX FLOPs counting

pip install onnx-tool

import onnx_tool
modelpath = 'resnet50.onnx'
onnx_tool.model_profile(modelpath, None, None)

The final MACs total count is actually the FLOPs count