tensorboardX

Why we included this project

PyTorch users who are tired of watching raw loss numbers scroll past in a console can get the familiar TensorBoard dashboard in their training loop without touching TensorFlow. You create a SummaryWriter and call a few methods to log scalars, images, audio, embeddings, histograms, and model graphs into standard event files that the regular tensorboard server reads. The same API also works with NumPy, Chainer, and MXNet, so one logging path covers different training stacks. Because it writes standard event formats, existing dashboards keep working, and the same data can be sent to cloud tracking services when you want richer comparison across experiments.