This is a random collection of utilities for inspecting TensorFlow summary files.
Under development (so for personal use at this point)!
pip install -I git+https://github.com/wookayin/tensorboard-tools
The CLI arguments of tensorboard sucks. We can launch it more easily:
tb logdir/1 logdir/shell_expansion/*
tb --port 6006 logdir/*
tb `gsutil ls -d "gs://bucket/exp/*"` logdir/*
Automatically scan train_dirs from currently running TensorFlow processes:
tb --auto
tb --auto <pattern>
Image summary in Tensorboard limits the number of steps being displayed. Run a server that would parse an event file:
python -m tbtools.image_viewer --logdir /tmp/path/to/traindir
python -m tbtools.image_viewer --event_file /tmp/path/to/traindir/events.out.tfevents.1234567890
Then we can inspect the images that are stored as image summary:
http://localhost:7006/<step>/<summary_tag>
http://localhost:7006/0/input_image/image/0
- Advanced Image Viewer
- Draw matplotlib plots for scalar values (e.g. loss)