tf_cnnvis — Open-Source CNN Visualization Library
tf_cnnvis is an open-source library for peering inside convolutional neural networks. CNNs had been winning ImageNet for years, yet for most practitioners the layers themselves stayed a black box — you could see that a model worked, but not why it made a given prediction. We built tf_cnnvis to close that gap: a researcher plugs in their own trained model, and with a couple of lines of code the tool hands back visualizations that demystify what each layer and feature map has actually learned. It runs on TensorFlow and renders results directly in TensorBoard. This was one of the first research projects I ideated at Infocusp.
Three visualization techniques
As one of the key contributors on the project, I helped implement and package three complementary visualization techniques into a single, reusable API. Activation visualization surfaces the raw layer outputs for a given input. Deconvolution — following Zeiler and Fergus's approach — traces a feature map's response back to the input to highlight exactly which pixels drove it. DeepDream visualization works in the opposite direction, synthesizing the input that maximally activates a chosen feature map, which is especially revealing for the high-level concepts learned in later layers.
Together these turn abstract network internals into concrete, inspectable images that support feature engineering and help diagnose model failures — like understanding why parallel-bar and balance-beam athletes collapse into one class.
Impact & availability
The library was designed to work with arbitrary user models rather than a fixed set of architectures, released under an MIT license with a citable DOI, and presented at NVIDIA GTC Europe. It has since been adopted broadly by the community, with 700+ GitHub stars and 200+ forks.