Design a BOT capable of beating reCAPTCHA challenges
Setup Plan/organize Read reCapthca v3 docs Github repo (+ test) / live shared coding env? Python / TensorFlow / AWS account + access
Model Find captcha specific dataset (images first, no text/numbers) Done (drive folder) Use data augmentation online (while training, less storage needed, see google dev code) to have more data (TF ImageDataGenerator module, see Kaggle) Pre-trained models (inception v4?) + tuned top (hyperopt?)
Possible milestones
- image recognition (multi-class classification)
- voting on 9 images, define a confidence threshold (multi-class classification + voting)
- Connect to API or scrape a website to test live safely + button click (production) and obtain human-like (~1.0) scores
Other GPU vs CPU Separate tasks, or do same but diff paths Future: try one-shot learning, structured learning