Neurolume
An open-source visualization tool for deep learning models that supports PyTorch and Keras and can interactively display a model's computation graph, weight changes and training dynamics in the browser.
Timeline (last 90 days)
Neurolume is an open-source interactive visualization tool that can draw the computation graph of a PyTorch or Keras model in the browser and display weight changes, loss curves, and prediction comparisons during training.
Neurolume was released by a developer on Reddit.
Neurolume generates an interactive chart in the browser by calling neurolume.show(model, sample), and this chart can display the connections between layers and reflect weight changes through color intensity.
Neurolume's live training feature is labeled by the developer as "very beta" and may have stability or performance issues, making it unsuitable for production environments.
Neurolume provides loss curves, overfitting checks, comparisons of predicted values against true values, highlighting of weight changes, and a timeline feature for replaying the training process.
Users can install Neurolume via pip install "neurolume[torch]" and run neurolume-demo --live to launch the demo.
Neurolume uses watch() together with live.log(step, loss=loss) to stream training data to the browser to update the loss curve in real time.
Neurolume's overfitting check relies on the user actively passing in validation metrics; otherwise it can only display the training-side curve and cannot judge generalization performance.
Neurolume's visualization tool must first trace an actual forward pass to obtain a graph instance with concrete tensor shapes, so show(model, sample) must also be passed an input sample.
Neurolume and TensorBoard are not in a fully substitutive relationship: TensorBoard is more mature in distributed training, hyperparameter comparison, and large-scale experiment management, while Neurolume is currently better suited for the author to quickly build intuition about model structure during local single-machine debugging.
2 more timeline events
All Facts (12)
Neurolume is an open-source interactive visualization tool that can draw the computation graph of a PyTorch or Keras model in the browser and display weight changes, loss curves, and prediction comparisons during training.
50%UnverifiedNeurolume was released by a developer on Reddit.
50%UnverifiedNeurolume generates an interactive chart in the browser by calling neurolume.show(model, sample), and this chart can display the connections between layers and reflect weight changes through color intensity.
50%UnverifiedNeurolume's live training feature is labeled by the developer as "very beta" and may have stability or performance issues, making it unsuitable for production environments.
50%UnverifiedNeurolume provides loss curves, overfitting checks, comparisons of predicted values against true values, highlighting of weight changes, and a timeline feature for replaying the training process.
50%UnverifiedUsers can install Neurolume via pip install "neurolume[torch]" and run neurolume-demo --live to launch the demo.
50%UnverifiedNeurolume uses watch() together with live.log(step, loss=loss) to stream training data to the browser to update the loss curve in real time.
50%UnverifiedNeurolume's overfitting check relies on the user actively passing in validation metrics; otherwise it can only display the training-side curve and cannot judge generalization performance.
50%UnverifiedNeurolume's visualization tool must first trace an actual forward pass to obtain a graph instance with concrete tensor shapes, so show(model, sample) must also be passed an input sample.
50%UnverifiedNeurolume and TensorBoard are not in a fully substitutive relationship: TensorBoard is more mature in distributed training, hyperparameter comparison, and large-scale experiment management, while Neurolume is currently better suited for the author to quickly build intuition about model structure during local single-machine debugging.
50%UnverifiedNeurolume currently mainly targets PyTorch and Keras users, and support for other frameworks has not yet been mentioned.
50%UnverifiedThe developer says feedback is being actively collected, with plans to improve the training features.
50%