I was curious to see what would happen out of sample, I figured it would just be a straight line (and it is, despite the parameters being quite complex, print out `model.weights` to check them out!)
outS = [[365],[400],[450],[500],[2*365],[3*365],[4*365]]
outY = model.predict(scaler_X.transform(outS))
outY = scaler_y.inverse_transform(outY)
Negative if you keep at it long enough.Not very important, but the weight loss curve usually decelerates with time and this pattern is not captured by a straight line either - as i say not a big deal, it just happens that by using a simple DNN model for a simple curve I don't have to spend time choosing the type of function I want the data to fit.
Also the model will improve as the number of measurements that are available increases.
Your comment does make sense, I just think for this use case and short term prediction it's not a major problem.
I'd be happy to learn how you'd apply some renormalization techniques to reduce the number of parameters and prevent the overfitting - I'm just a noob very curious to learn as most as I can about practical deep learning techniques.
I have done keto for weight loss across several stints, one lasting 4 years (I had a lot of weight to lose). I encountered plateaus at various points were despite pretty consistent caloric intake (at the time I religiously tracked everything), activity, and normal body eliminations, I would have periods of several weeks where my weight would not move at all, only to dramatically fall in a later week and continue the steady loss. I found it demotivating and had to adjust weigh ins from weekly to monthly to avoid the frustration.
I wonder what the model would say about those periods. I wish I would have kept the data from that time, it would have been interesting
I didn’t happen to find a plateau yet but surely will share the data if that happens.
The nice things of using a simple feedforward DNN is that it will automatically infer both the nonlinear function structure and parameters and for simple basic use cases like this that don't require a lot of training it works well and you can even validate the fit visually by looking at the graph output.