I am using one current competition from Kaggle.com to make experiments with deep learning. They employ pretrained / weighted neural network models for image classification distributed by Keras for the Amazon Planet competition. Pretraining is a way to transfer learning, that is to say to use models trained on standard image datasets to make predictions on new image datasets, only changing the last layer or two in the model. My new laptop with a GTX 1070 8 GB graphics card comes just handy for working locally with the VGG-16 and ResNet-50 models, my best compromise between hardware affordability and results accuracy. Better and smarter approaches are described in the Fast.ai forums here, food for thought!
Getting calls from Milan about my availability to move there for engineering roles, both from the r&d mechanical and the machine learning industries. Time to create a public portfolio then, I would really prefer consultant roles to becoming an employee and I need to show I can add real value. It will be on GitHub as NukePep for both fields, filling it this Summer.
My 2500+ nuclear eng papers collection morphed into a 500 papers database for machine learning uses. Now tracing references for open problems. Targets are materials science, structural integrity and engineering design in the steels domain.
Back to kaggle.com for their competitions, there are three that I find interesting these days: Mercedes, Instacart, Zillow. Mercedes want to optmize their test bench and reduce the time their cars spend there. Data are pretty small and there is a lot to play with models and ensembles, that’s why I’m interested. Following this terrific ensembling guide from one of the kaggle masters, it’s about time I learn their tricks! Deadline 10 July, not expecting to follow the public leaderboard, I just think many are overfitting, just hoping to come up with a robust multilayer solution for my portfolio.
Too much blurb from live events or no interest in ping pong office culture? You would like knowing more about startups though, real startups from the Silicon Valley. Here Y Combinator comes with their 10-weeks free MOOC about startups’ best practices, which they have named Startup School. No need to sign up for videos, transcripts & slides, enjoy from home instead of going to San Francisco. Is that material applicable to different environments? From my experience it is not, there is no market large enough to sustain scalability up to the unicorn level making use of Venture Capital funds.
According to the much acclaimed Report, investigative journalism TV programme from the national Italian broadcaster Rai 3, the nuclear power plant in Krsko, Slovenia, operates in a seismic region; the one in Doel, Belgium is undergoing severe crack issues in the pressure vessel; the one in Fessenheim, France and several from Areva suffer from high carbon steel problems, adding to the new EPR delay and technical problems in France, Finland and the UK. It is a mess, frankly, too many old plants in Europe in need of life extension and no serious plan to replace them.
Structural design of nuclear power plants’ internals is now considering methods that are more elaborated than the historical linear elastic formulas. The Non-Linear Analysis Design Rules; Part 1: Code Comparison report, released by World Nuclear Association on Feb 2017 under their CORDEL project, “reviews and compares the current code requirements in non-linear analysis for different failure modes and some degradation mechanisms in the major nuclear and non-nuclear design codes.” There is a consensus at academic and consultancies level but regulators still need to have a proper dive into these and homogenise their standards, that’s why this work is important.
I am considering a move to Milan, so testing waters for a good job there as an r&d mechanical engineer or a data scientist. Market is flooded by job agencies and body rental shops, therefore I need direct contacts with firms through my own 2016 database, which is made of 5000+ listings in Central and Northern Italy. Neither an optmist nor a pessimist here, just going to have a good look in June and see what happens. First step there may be renting a flat with a good wi-fi service in the recently rejuvenated quartiere Isola for one month.
Deep learning benchmark from Fast.ai course Lesson 1 with a GTX 1070 8GB from my Win 10 laptop with factory settings: 292s stable (260-300s with 1070, 380-430s with a 1060 6GB, 120-150s with a 1080 Ti 11GB). Laptop ready, wearing the helmet, off I go: Fast.ai course, Kaggle.com competitions, own studies.
I am finding difficult to perform a clean install of Python packages (Theano, Tensorflow, Keras on top of Anaconda and xgboost) for deep learning on my Win 10 native laptop with a GTX 1070 8GB. Procedures are cumbersome and the most recent drivers do not work, adding to compiling errors. I am now looking at this tutorial by Phil Ferriere updated to May 2017, which seems reasonable and clean. Alternatively, an Ubuntu 16.04 partition on my second data hard disk should allow a smoother install. I’ll go that route, Windows is too fragile to risk flooding my SSD again. EDIT: the tutorial by Phil worked and I now have all the packages I need up and running on Win 10. Woohoo!