TensorFlow 1.x nostalgia: a messy but honest learning scrapbook

What it does This repo collects half a dozen TensorFlow 1.x experiments: MLP on MNIST, TFRecord conversion for the 102 Flowers dataset, fine-tuning with tf.slim, AC-GAN and Info-GAN implementations, a CGAN-based anime colorizer, Chinese handwritten character recognition, and classic-to-deep text classifiers (Word Count, TF-IDF, CNN, C-LSTM). Most pieces include a Flask wrapper for inference.

The interesting bit The author actually shipped a live demo at demo.duanshishi.com and candidly notes the fine-tuning code came from a “porn detection” side project — a rare flash of autobiography in an otherwise terse README.

Key highlights

• GAN coverage is unusually broad for a personal repo: both AC-GAN and Info-GAN with result screenshots
• Includes practical deployment patterns: TensorFlow Serving for MNIST, Flask for fine-tuned models
• Chinese handwritten character recognition is a niche addition rarely found in English-centric tutorials
• Text classifier section compares classical NLP baselines against CNN/C-LSTM, not just deep learning
• Most experiments link to detailed Chinese blog posts with actual methodology

Caveats

• README is extremely terse; most context lives on external blogs (some links may rot)
• Heavily tied to TensorFlow 1.x APIs (tf.slim, old serving patterns) — largely historical now
• Data downloads depend on Baidu Pan, which is inaccessible without a Chinese account
• No tests, no requirements.txt, no clear dependency versions

Verdict Worth a quick browse if you’re reverse-engineering legacy TensorFlow 1.x patterns or need a working GAN reference in old TF. Skip it if you want production-ready code or modern TF 2.x/Keras idioms.