This week I continued training SqueezeNet and prepared for the Symposium.
This day I continued work on getting squeezenet to train. I have found where it is determined whether the model's output is accurate or not, however I have no understanding where the 'target' output comes from; it significantly differs between training and validating. I changed validate to output a metric that is not the expected one (top5.avg instead of top1.val) because the latter is always 0 and the former typically has an actual value. This way (theoretically) as the model trains it still only saves 'better' versions instead of saving the first one and then sitting there.
Concurrently, I worked on my Symposium poster. I am presenting on Thursday, so my poster needs to be done and printed out before then, but I have no information on how to get a poster printed or if that is my director's plan for me. Also, I have no instructions or examples on posters so I have no base to go off of for the design and what to include.
This day I finalized my Sympsium poster. I think it looks nice.
Concurrently, I continued training the squeezenet model on ImageNet. It still occasionally saves a new best model so in theory it is improving; hopefully by Thursday I will have a model that has any validate precision (what is supposed to have values but always is 0) so that I may talk about it during the Symposium.
This day I sent my poster to the print shop so that it will be ready Thursday morning for the Symposium.
Concurrently, I continued training the squeezenet model on ImageNet. It has not improved on best model since yesterday - I suspect my work-around for evaluating validation did not work.
This day I presented a poster at Brown University's 2018 Summer Symposium; my poster can be found here.
This day I attended a workshop on networking, where Ethernet cables, UDP/TCP, IP, and other basics were discussed.