A simple library for parsing Tensorboard logs.
You can install the library as usual with pip:
pip install git+https://github.com/velikodniy/tbparser.git
The main class is tbparser.summary_reader.SummaryReader.
Creating an instance:
from tbparser.summary_reader import SummaryReader
logdir = 'my_logs'
reader = SummaryReader(logdir)It will try to parse every file in the logs directory and its subdirectories.
Also you can add a filter on tags if you interested in only specific events.
Just list the tags in the tag_filter argument.
reader = SummaryReader(logdir, tag_filter=['loss'])In this case the reader will yield only events with tag 'loss'.
By default reader yields only scalar values.
It's because the argument types is set to ['scalar'].
You can list all the types you interested in.
All the unlisted types will be ignored.
The supported types are:
scalar— a scalar value;image— a decoded image (Numpy array);image_raw— a raw image (bytes of PNG, for example).
Other types will be added later (maybe).
Note that if you are subscribed both on image and image_raw, any particular image event will produce two yields.
There is also an argument stop_on_error that is False by default.
After creating you can use the reader as an ordinary iterator:
while item in reader:
do_something(item)
item is an instance of tbparser.summary_reader.SummaryItem.
It is a named tuple actually with the following keys:
tag,step,wall_time,value,type— a string with the type name (e.g.,'scalar').