What it does

This repository holds slides and Python code examples from a 2016 talk on debugging TensorFlow. It covers common pitfalls in the pre-eager-execution era—shape mismatches, uninitialized variables, silent failures in the graph—and walks through techniques for inspecting what TensorFlow is actually doing under the hood.

The interesting bit

The talk predates tfdbg, TensorFlow’s official debugger, which arrived later in 2016. The author then retrofitted the materials to mention tfdbg in updates, creating a small archaeological record of how developers debugged before first-class tooling existed. The video is in Korean; the code and slides are not.

Key highlights

• Slides are web-hosted and browsable without cloning anything
• Code examples updated for TensorFlow 1.0 compatibility (a 2017 patch)
• Includes pre-tfdbg manual techniques: tf.Print, tf.Assert, session inspection
• 609 stars suggest the pain of debugging TensorFlow is timeless, or at least long-lived
• Video recording available for Korean speakers or the autotranslate-curious

Caveats

• Materials target TensorFlow 1.x graph mode; eager execution and TF 2.x are not the focus
• Last meaningful update was February 2017; some APIs shown are likely deprecated
• The video is not subtitled, so non-Korean speakers get only the code and slides

Verdict

Worth a skim if you’re maintaining legacy TensorFlow 1.x code or teaching graph-mode concepts. Skip it if you’ve already internalized tfdbg or live entirely in TF 2.x / PyTorch land.