I’ve tried to build a basic model for Favorita, using fasta.ai library and based on the Rossmann notebook, after watching DL1 Lesson 4 and ML1 Lesson 3 videos.
Note: I worked for Carrefour, like Walmart for EU, as hypermarket store manager for 2 years then in Purchasing/Marketing HQ for another 2 years, so I’ve done sales/inventory predictions a lot, old-school style.
I kept things as basic as possible to start, like not including national/local holidays nor working on Durations or adjusting/counting the # of week-ends for each month. I clipped the “Unit_Sales” negative values or returns to 0.
Also I chose to work on June-August 2016 data as it’s complete, with the idea of Training on 8 weeks and Validating on 2 weeks (ie same ones as the Test 2017 data).
I’ve played around different fixes/versions in the past 2 days but I can’t get the model to “take-off”: it won’t figure the learning_rate, running basic epochs generate “NaN” results, so I don’t know what to do next.
To comply with Kaggle rules, I posted this notebook’s code in a public kernel here with a link to this now public thread.
My basic notebook, as ran/saved, can be viewed:
Or downloaded on GitHub:
Any idea/tip ?