daQuincy / kcg-ml-image-pipeline

kcg-ml-image-pipeline

• kcg-ml-image-pipeline
  • Prerequisites
    • Use Python Environment:
    • Install Dependencies
    • Install MongoDB and Running MongoDB
  • Orchestration API
  • Worker
    • Prompt Generator
  • Model Training
    • AB Ranking Linear Model
    • AB Ranking Efficient Net Model

Prerequisites

Use Python Environment:

Create an env by running:

python3 -m venv venv

Then activate it by running:

source ./venv/bin/activate

Install Dependencies

Install dependencies by running:

pip install -r requirements.txt

Install MongoDB and Running MongoDB

For MacOS:

    brew install mongodb-community
    mkdir /data/db
    brew services start mongodb-community

For Ubuntu:

    sudo apt install mongodb
    mkdir /data
    mkdir /data/db
    sudo systemctl start mongodb
    sudo systemctl enable mongodb

For Devboxes (Containers):

    sudo apt install mongodb
    mkdir /data
    mkdir /data/db
    sudo systemctl enable mongodb
    mongod &> mongodb.log &

Orchestration API

To deploy orchestration api, run:

uvicorn orchestration.api.main:app --reload

or

python3 main.py

API Docs should be accessible at:

http://127.0.0.1:8000/docs

Worker

To run the worker, run:

options:
  -h, --help            show this help message and exit
  
  --device DEVICE
                Choose a device for the worker example ('cuda:0', 'cuda:1' ..)

python3 ./worker/worker.py --device cuda:0

Prompt Generator

Generates prompts and saves to a zip file

usage: prompt_generator.py [-h] [--positive-prefix POSITIVE_PREFIX] [--num-prompts NUM_PROMPTS] [--csv-phrase-limit CSV_PHRASE_LIMIT] [--csv-path CSV_PATH] [--output OUTPUT] [--positive-ratio-threshold POSITIVE_RATIO_THRESHOLD] [--negative-ratio-threshold NEGATIVE_RATIO_THRESHOLD]
                           [--use-threshold USE_THRESHOLD] [--proportional-selection PROPORTIONAL_SELECTION]

Prompt Generator CLI tool generates prompts from phrases inside a csv

options:
  -h, --help            show this help message and exit
  --positive-prefix POSITIVE_PREFIX
                        Prefix phrase to add to positive prompts
  --num-prompts NUM_PROMPTS
                        Number of prompts to generate
  --csv-phrase-limit CSV_PHRASE_LIMIT
                        Number of phrases to use from the csv data
  --csv-path CSV_PATH   Full path to the csv path
  --output OUTPUT       Output path for dataset zip containing prompt list npz
  --positive-ratio-threshold POSITIVE_RATIO_THRESHOLD
                        Threshold ratio of positive/negative to use a phrase for positive prompt
  --negative-ratio-threshold NEGATIVE_RATIO_THRESHOLD
                        Threshold ratio of negative/positive to use a phrase for negative prompt
  --use-threshold USE_THRESHOLD
                        True if positive and negative ratio will be used
  --proportional-selection PROPORTIONAL_SELECTION
                        True if proportional selection will be used to get the phrases

To get civitai csv, download via:
https://mega.nz/file/wNIXHKDA#CIB92fLkCquatWvzUAjjcKbrKCUmO2ffRVrzE3XYQVM

Example Usage:

python ./prompt_generation/scripts/prompt_generator.py --num-prompts 10000 --positive-prefix "icon, game icon, crystal, high resolution, contour, game icon, jewels, minerals, stones, gems, flat, vector art, game art, stylized, cell shaded, 8bit, 16bit, retro, russian futurism" --csv-phrase-limit 512 --csv-path ./input/civitai_phrases_database_v6.csv --output ./output/prompt_list_civitai_10000 --positive-ratio-threshold 3 --negative-ratio-threshold 3

Model Training

AB Ranking Linear Model

To train ab ranking linear model:

1. Install dependencies by running pip install -r requirements.txt
2. Run using python ./scripts/ab_ranking_linear.py --minio-access-key access --minio-secret-key secret --dataset-name propaganda-poster --input-type embedding

AB Ranking ELM Model

To train ab ranking linear model:

1. Install dependencies by running pip install -r requirements.txt
2. Run using python ./scripts/ab_ranking_elm_v1.py --minio-access-key access --minio-secret-key secret --dataset-name propaganda-poster --input-type embedding

AB Ranking Efficient Net Model

To train ab ranking linear model:

1. Install dependencies by running pip install -r requirements.txt
2. Go to ab_ranking_efficient_net test run entrypoint
3. Input the minio server's address, your minio secret, and your minio access key.
4. Run using python python ./training_worker/ab_ranking/script/ab_ranking_efficient_net.py

Tools

Check Dataset Clip

This checks clip data for selection datapoints in the dataset and creates a clip calculation task if clip data doesn't exist.

python ./scripts/check_dataset_clip.py --minio-access-key nkjYl5jO4QnpxQU0k0M1 --minio-secret-key MYtmJ9jhdlyYx3T1McYy4Z0HB3FkxjmITXLEPKA1 --dataset-name all