ezboard

This project aims to be a very simple replacement for tensorboard during prototyping.

ezboard parses the stdout of a training script for loss values. The lines have to contain the word "loss". Some variants are possible, e.g. the line "Epoch 5/100 Loss: 3.5 Acc: 80%" will be parsed as "3.5". More heuristics will be added in the future.

Installation

Only installation from source is possible right now. You will need a recent Rust toolchain and cargo.

git clone git@github.com:hesom/ezboard.git
cd ezboard
cargo install --path .

Usage

Just pipe the output of your training script into ezboard. By default lines are buffered by python before they are sent to ezboard, which causes a delay. To disable line buffering, pass the -u to the python interpreter:

python -u train.py | ezboard

Alternatively you can pass a logfile directly

ezboard train.log

Use ezboard -h for all command line options (e.g. smoothing).

There are some hotkeys that will be expanded in the future:

Shortcut	Description
p,P	Toggle between graph and raw log output
q,Q	Shutdown ezboard. This doesn't stop the training
↑,↓ or ←,→	Select a different metric to plot