Download dataset here

• Training set: 100000 labeled captchas.
• Validation set: 20000 labeled captchas.
• Testing set: 20000 unlabeled captchas.
• All captchas are resized to 50 x 100.

Convolutional Neural Network encodes captcha images into simpler representation.

Output from attention layer is fed into GRU cells followed by FC layers to convert encoded captchas into strings.