zhihu

Why we included this project

Each folder in this repo is a self-contained TensorFlow or Keras experiment, and every one of them pairs with a detailed article in the author's long-running Zhihu column, 机器不学习, so the reasoning sits right next to the code. Between them the projects cover most of the classic deep learning canon: character-level LSTM text generation from Anna Karenina, Word2Vec skip-gram, Seq2Seq machine translation with and without attention, denoising autoencoders, CNNs for CIFAR classification, GANs and DCGANs, batch normalization, style transfer, and sentiment analysis. Since every experiment comes with a write-up that walks through the theory and the training setup, it's a practical place to reproduce a model from scratch and understand why it behaves the way it does. The CTR prediction models at the end (DeepFM, Deep&Cross, xDeepFM, AutoInt) push the collection past pure NLP and vision into recommender-style ranking. It is not a production service, but as a structured tutorial it maps closely to the ideas you would implement in real projects。