Improve architecture

[xbankov]

• Choose TensorFlow or PyTorch, not both, it is inconvenient.
• Use some pre-trained models for siamese page detector.
• Rewrite fastai1 to fastai2 or even better use basic PyTorch for semantic segmentation screen detector.
• Successor to VGG16, e.g. xResNet features could be used as a page detector.
• Train GAN for screen to slide restoration translation.