Add python script to plot the training log

[kloudkl]

No description provided.

[ghost]

Would be good to have. One can better formulate the solver messages or potentially have the solver write the key parameters like current loss to a separate log file so further scripts could analyze it more easily.

[sguada]

I have a bash script that extracts the relevant information from the log,
which later can be easily plotted with python, gnuplot, matlab, ...
I can share or simply add it to the code base see #89.

2014-02-08 kloudkl notifications@github.com:

Reply to this email directly or view it on GitHubhttps://github.com//issues/86
.

[kloudkl]

@sguada, thanks for sharing your script in #89 and #90! I am inspired a lot.

Due to the diversified requirements of the users of Caffe, the best we can do is to provide a sample python script as the basis for various customizations. To this end, data extraction and plotting should be separated as you have done. For benchmarking purpose, results of multiple experiments often need to be plotted in one chart.

Based on the information present in the current version of training log, possible useful plots are summarized as follows. In the future, this script should be synchronized with the changes of the training log format such as adding training or validation accuracy.

1. Test accuracy (test score 0) vs. training iterations / time;
2. Test loss (test score 1) vs. training iterations / time;
3. Training loss vs. training iterations / time;
4. Learning rate vs. training iterations / time;

[shelhamer]

Good job everyone. This is done with #89 #90 #91 .