Open cocreature opened 7 years ago
Here are the boxplots for random samples and various precisions. Looks pretty reasonable:
Here’s a plot of the mean error rate and the 5% and 95% quantiles for precision 14 (those values are used in the plot in the google paper). Apart from the additional precision in the sparse representation these error rates are more or less the same as the ones shown in the google paper so our implementation seems to be ok.
How did you generate the graphs?
The code for generating a csv with the data for the first plot is already commited. The second code is currently just a slight variation of that. I plan to merge them into a single executable but I’m not sure when I’ll get to that. The actual plots are created using matplotlib, I’ll commit the ipython notebook for that in the next few days.
I’ve commited the updated evaluation code & the ipynb.
I’ll abuse this issue to post any graphs or other results that I produce. This way we can also easily check if the graphs improve.