Sometimes you might want to keep ('pin') some data on GPU. We demonstrate how to do it in Tensorflow and PyTorch.

Continue reading

First part of a tutorial serie to write your own CUDA kernel and use it in Tensorflow or PyTorch.

Continue reading

Second part of a tutorial serie to write your own CUDA kernel and use it in Tensorflow or PyTorch.

Continue reading

A walkthrough of many of the current techniques for profiling time and memory usage for tensorflow.

Continue reading

Long due blog post about our singularity container use!

Continue reading

News from the last week's meeting publications and talk/job opportunities!

Continue reading

News from last 2 weeks, new meeting time, publications and talk/job opportunities!

Continue reading

I train a very simple and basic mnist classification network with a lot of overkill: I use minibatching, batch normalization, and save the network weights to disk. This tutorial can be done on a CPU. In Part 2, I restore the model, run on the validation set, and analyze the results.

Continue reading

See Part 1 first! There, I trained an mnist classifier. Here, I restore the trained network, run the network on the validation script, and do some analysis on the output. This is meant as a template for new users to see "how do I actually use a trained network?" Lots of this information exists elsewhere too, I've just tried to consolidate the basics.

Continue reading

Here's a report for a naive data rate profiling of ThreadProcessor, larcv's threaded data reader we often use for network training. Measured on my macbook pro early 2013 model (old!) but reproduced similar numbers on dell xps15 (9560).

Continue reading