Similar to the fitting toolbox shipped in MATLAB (at least 10 years ago, last time I used it), this single page app works by inviting the user to upload a csv datafile, pick two of its columns, one for x one for y, inputting a functional form y=f(x) with some free parameters.
There is freedom for setting the initial value of each of these parameters.
Finally, the user can run the fit, the optimizer will run a gradient descent to minimise the mean squared error.
Under the hood, the functional expression is transformed to a mathematical graph, where everything is decomposed into atomic mathematical operators and functions. This graph is then translated into a TensorFlow graph.
- Any function existing within the tfjs API can be used in the function template (
exp,log,pow, etc...). - Visualisation of the data points and the fitted line.
- Goodness of fit metric, R-squared value is printed.
- The fitting process can be interrupted by the user, allowing for regularisation by early stopping.
- Giving the user freedom in the optimiser choice, the learning rate and the error to be minimised.
- Offering cross validation by splitting the data into training and validation.
- Providing L2 regularisation.
- Generalising to higher dimensions, not only
$$\mathbb{R} \to \mathbb{R}$$ mappings. - More freedom in data imports: excel format, google sheets, copy-pasting, ...
This project was generated with Angular 7, using ECharts for visualization, Math.js & TensorFlow.js for modelling and training.