Meet TorchExplorer: A New Interactive Neural Network Visualizer

A new AI research introduces TorchExplorer, a novel AI tool designed for researchers working with unconventional neural network architectures, which provides an interactive and insightful exploration of network layers. Developed to aid in understanding complex neural network models, the tool automatically generates a Vega Custom Chart in wandb, a class showcasing a module-level visualization of the network architecture. Users can also deploy TorchExplorer to a local web server for standalone use.

While exploring the user interface, the left-hand panel of TorchExplorer features a module-level graph extracted from the autograd graph, allowing researchers to navigate through the network’s structure. Clicking on a module reveals its internal submodules, and a convenient expanding list helps users return to parent modules effortlessly. The nodes in the explorer graph represent either input/output placeholders or specific invocations of submodules. Each submodule invocation is distinct, emphasizing individuality even if a submodule is called multiple times in a forward pass.

Edges between nodes signify autograd traces, indicating the flow of information from parent to child modules. The number of edges is unrelated to the number of inputs/outputs in the forward function, providing clarity on information flow in the network. TorchExplorer’s column panels on the right enable users to inspect modules in detail by dragging and dropping them. The histograms accompanying each module visualize the distribution of values at the corresponding x-axis time. These histograms, representing input/output tensors, are subsampled for performance reasons and reject outliers to maintain accuracy.

Input/output histograms showcase values passing into and out of a module’s forward method, providing a detailed view of data distribution. Additionally, input/output gradient norm histograms capture tensor gradients from backward passes, offering insights into the ℓ2-norm of gradients averaged over the batch dimension. Parameter histograms log the immediate parameters of submodules, while parameter gradient histograms depict gradients of the loss concerning each parameter.

A noteworthy feature of TorchExplorer is its ability to handle non-standard architectures effectively. Researchers often dealing with unconventional models can benefit from the tool’s capacity to adapt and provide meaningful insights, even in the face of wacky network designs.

In essence, TorchExplorer is a valuable companion for researchers engaged in deep learning experiments, offering an interactive and visually intuitive way to comprehend network behavior at various layers. Its deployment flexibility, both on wandb and locally, enhances its accessibility, making it a versatile tool for the research community.