PHYS379 Group Project v1.1.0

PHYS379 Group Project.exe

Version 1.1.0

Result of creating a GUI-supported neural network program. This second release includes light curve data from NASA's LightKurve library, financial stocks data from Yahoo Finance's YFinance library, and Temperature data from the Met Office (namely from the HadCRUT5, NOAAGlobalTemp, and Berkeley Earth datasets that can be downloaded at https://climate.metoffice.cloud/temperature.html#datasets).

New Features:

• Fixed logo to have accurate text.

Notes for use:

• Download size is currently: 141.084MB. This large size is mainly due to how the LightKurve library requires Astropy to be imported to function, which takes up a substantial amount of storage space. Further reductions in size are difficult.
• Command prompt opens when the app is launched. This is intentional, and is helpful for debugging.
• Currently, the final layer in the network must have ONLY ONE neuron, otherwise errors occur.

The pre-release is a significantly smaller file size - feel free to download that, if that is preferable (although it will lack the option to work with the LightKurve and Met Office datasets)

PHYS379 Group Project v1.0.0

PHYS379 Group Project.exe

Version 1.0.0

Result of creating a GUI-supported neural network program. This first release includes light curve data from NASA's LightKurve library, financial stocks data from Yahoo Finance's YFinance library, and Temperature data from the Met Office (namely from the HadCRUT5, NOAAGlobalTemp, and Berkeley Earth datasets that can be downloaded at https://climate.metoffice.cloud/temperature.html#datasets).

New Features:

• Options to plot data from LightKurve and the Met Office.
• New graph displaying flattened data, to make outliers more explicit.
• Improved console display.
• Updated logo.

Notes for use:

• Download size is currently: 141.083MB. This large size is mainly due to how the LightKurve library requires Astropy to be imported to function, which takes up a substantial amount of storage space. Further reductions in size are difficult.
• Command prompt opens when the app is launched. This is intentional, and is helpful for debugging.
• Currently, the final layer in the network must have ONLY ONE neuron, otherwise multiple fits occur in the plot.

The pre-release is a significantly smaller file size - feel free to download that, if that is preferable (although it will lack the option to work with the LightKurve and Met Office datasets)>

PHYS379 Group Project v0.1.0-alpha

PHYS379 Group Project.exe

Version 0.1.0-alpha

Result of creating a GUI-supported neural network program. It outputs similar graphs to Samir's work and works pretty well (in my definitely unbiased opinion).

Features:

• Graphs for the fit of the neural network against data and the RMSE loss over iterations.
• Options to vary the learning rate, batch size, stock ticker, and time range.
• Additional options to vary the optimiser function (set to SGD by default, even though the alternative is better), and to vary the amount of data that is used to train the network from 5% to 95%.
• Finally, there's the option to edit the network itself, and add layers with Sigmoid, ReLU, Tanh, and Linear activation functions.

Notes for use:

• Download size is currently: 54.586MB (I struggled to make it much smaller than this)
• Command prompt opens when the app is launched. This is intentional, and is helpful for debugging. I will remove this in the final release (and hopefully enable print statements to be visible within the GUI, we'll see).
• Currently, the final layer in the network must have ONLY ONE neuron, otherwise multiple fits occur in the plot.

Future plans:

• Possibly extend to work with convolutional neural networks.
• Add polynomial / logarithmic regression types.
• Allow any models to be saved to a .json file, so trained networks can be saved and rerun, if desired.