For some reason each lines intersecting a diagram leads to generate in the patch list for one line, 18 diagrams, causing big issue with the length for visualization, and in general makes no sense at all.

Rotation of line doesn't seem to work properly. Plus they aren't displayed on plot.

Summary

As explained in the main issue note #7, the FF cannot find the angle among the noise, even with the derivative of the diagrams, or when it's trained with synthetic data. Instead of a FF, we could use a CNN, much more accurate for this kind of problem. It seems to be a rational alternative considering all the options found while trying to solve the FF prediction problem.

The model doesn't predict properly the value of the lines when the patch is rotated. It keeps the same value as when the patch is not rotated

Changed the way lines are selected, only takes the segment of a line crossing effectively the patch but nothing is displayed.

Examples of segment:

([52, 17], [2, 7])
([23, -21], [12, 18])
([23, -19], [2, 7])
([-83, 0], [29, 16])

There is either a problem in how the lines are fetch or how they are drawn over the imshow. They

As shown in issue #7 the angles are not well distributed with a strong imbalance for angles between 0.41 and 0.48. This issue explores resampling for better training.

Currently network doesn't give a correct output when running test_network.py. Output is always 0.4202.

• Diagrams are not all the same
• Lines are correctly calculated
• Angles are correctly calculates

In training (simple_network.py) the y_pred is always the same (line 89).

Resampling dataset and reducing network size did not work. Last way to solve the issue could be to train on synthetic data and test on experimental one.

A look at the derivative of stability diagrams to see if it can improves the angle detection.

The patches contain lines either of light color or dark one, the network might struggle differentiating them. Training a network to recognize the angle of lines of various shades could solve the issue.

While rotating random images to test the model with a wider range of angles, the program gets an error while unpacking coordinates. Here is an example

line coords: [([-31.0, 135.0, 229.0, 323.0, 468.0], [9.999999046325684, 41.00000762939453, 61.0000114440918, 79.00001525878906, 110.0000228881836])]

This makes no sense at all.

After exploring downscalling data, it's time to check if network size influences the training.

Model and Plot should be saved together instead of being separated. Could be easier then to see the loss in relation to the model.