What I did

• Monday: Fixed MSD Values of particle averaged structure. I was double counting some bonds before.
• Tuesday: Neural Network now predicts 5 features. Wrote some nice code to visualize the predictions on a nice subplot.
• Wednesday: Finally got the \(g(r)\) code to work with the skew-norm fitting. It’s kind of tempermental though. And I need to fix it so it doesn’t double count some bonds.
• Thursday: Identical spectra don’t have the same msd when comparing particle average to skew-norm average. This is problematic.
• Friday: Conv1 and LSTM working in my model.

What I learned

• Tuesday: I figured out the error I was getting while building an RNN was due to an incompatibility with numpy and my version of Tensorflow. I downgraded numpy and it fixed the problem.
• Thursday: MSD is variance
• Friday: Given a set of data (in my case a histogram), the descriptive statistics of the set don’t match the fit. At least for skew-norm. I’ve justified it by saying that the minimization loss function works differently from how the dataset stats work. Skew-norm is also tricky because when you fit a histogram with stats.skewnorm.fit, it gives you the skew. But the input parameter C of stats.skewnorm.pdf isn’t actually the skew unless it’s zero.</div>

What I will do next

Hyperparameter optimization of my best 1-d convolution model