New CNN Model

[hvoss-techfak]

As part of a seminar project at my university, I created a CNN model using a Cartesian Genetic Programming algorithm. As you can see from the graph below, it achieves a much higher accuracy than the ConvolutionalNNDropout model. The CGP project adapted for this is: "https://github.com/scheckmedia/cgp-cnn-design".
The model is currently only designed for 48x48 grayscale images, as I trained it on the FER+ dataset.

Here is a visualization of the created model architecture:

[xuv]

Hello @hvoss-techfak thank you for your submission.

I've asked @dianagamedi to have a look at your proposition. She might be quite busy at the moment. So it could take some time. Thank you for your patience.

[DanielUSPS]

Did you perform any data augmentation of the FER dataset? If so, could you upload your code?

What changes did you make to use the FER+ dataset instead of the FER dataset?

Thanks

[hvoss-techfak]

Did you perform any data augmentation of the FER dataset? If so, could you upload your code?

Data augmentation is done in code at line 459 in the neuralnets.py file

What changes did you make to use the FER+ dataset instead of the FER dataset?

I simply used the instructions found here: https://github.com/microsoft/FERPlus#training-data
At the end you should have the same structure as the fer dataset.

Thanks

[DanielUSPS]

The FER dataset uses 1 hots for the 7 emotions.
The FER+ has a 0-10 value for each emotion that could be converted to a probability for each emotion. Did you take advantage of this? If so, did it improve your results?

[hvoss-techfak]

The FER dataset uses 1 hots for the 7 emotions.
The FER+ has a 0-10 value for each emotion that could be converted to a probability for each emotion. Did you take advantage of this? If so, did it improve your results?

No I just encoded it as a onehot vector, as I only needed the prominent emotion