Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

New CNN Model #31

Open
wants to merge 1 commit into
base: master
Choose a base branch
from

Conversation

hvoss-techfak
Copy link

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.

model compare

Here is a visualization of the created model architecture:
model_plot

@xuv xuv requested a review from dianagamedi July 11, 2019 09:35
@xuv
Copy link
Member

xuv commented Jul 11, 2019

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
Copy link

DanielUSPS commented Aug 21, 2019

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
Copy link
Author

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
Copy link

DanielUSPS commented Aug 21, 2019

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
Copy link
Author

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

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants