3 ms·

I've been collecting data from my phone and some different "sensors" (food delivery orders, bank payments, Activity Watch statistics, I'm adding a video camera

I've been collecting data from my phone and some different "sensors" (food delivery orders, bank payments, Activity Watch statistics, I'm adding a video camera soon. For phone about every sensor I have in my phone like ambient light, wifi APs, GPS, rotation, steps, when it was last unlocked, what app is in use, what app / artist is playing media etc.) and I'm using ML to predict activities I'm currently doing. I guess the end goal is to automate my diary writing to GPT as a joke.. Started as a small project on wanting to learn C# better and being interested about how much sensors collect information about us and how we could analyze that, but it really got out of hand. I've been working on it for a ~month (with some breaks), sometimes all day, built a lot of tooling (like webassembly & android app to tag current activities for ML training data) and analysis things for it and am nearing a burnout point :P but been fun and I've definitely learned a lot about C# and ML.

politelemon 3y ago
That's quite interesting. What form is the collection being fed in to the ml, is it sentences that gpt would understand or more tabular
- bayesianbot 3y ago
  It's just hundreds of variables of data and some calculated features - basically from current point in time and some configurable points in history + some aggregated history data, fed to ML.NET FastTree / FastForest. I think LSTM / some Python frameworks could have been better, but learning C# better was one of the main reasons so ML.NET algorithms it is. I get quite reasonable predictions, though I need to do some more to get better at predicting the end of an activity because of some limitations how I'm giving the data to avoid overfitting. GPT is just going to be last step of the process, being fed todays actions and we'll see if I can get it to blurt out anything interesting based on it. I just checked my manual tags in the training database and I've been on this for 197 hours and 37 minutes since since 12th of October. No wonder I feel like I need a break..
bruceb 3y ago
If you post about this project, have a github, etc. Would follow your progress.

politelemon 3y ago

That's quite interesting. What form is the collection being fed in to the ml, is it sentences that gpt would understand or more tabular

bayesianbot 3y ago
It's just hundreds of variables of data and some calculated features - basically from current point in time and some configurable points in history + some aggregated history data, fed to ML.NET FastTree / FastForest. I think LSTM / some Python frameworks could have been better, but learning C# better was one of the main reasons so ML.NET algorithms it is. I get quite reasonable predictions, though I need to do some more to get better at predicting the end of an activity because of some limitations how I'm giving the data to avoid overfitting. GPT is just going to be last step of the process, being fed todays actions and we'll see if I can get it to blurt out anything interesting based on it. I just checked my manual tags in the training database and I've been on this for 197 hours and 37 minutes since since 12th of October. No wonder I feel like I need a break..

bruceb 3y ago

If you post about this project, have a github, etc. Would follow your progress.