[BUG] emotions-detector subject requirements reasonable?

[jarmo-seljamaa]

emotions-detector

The Step 1 instructions in the Face Emotions Classication says:

The CNN has to perform more than 70% on the test set.

However looking at very many different CNN implementations at Kaggle forums on the FER 2013 dataset, I have not yet found a single one that achieves more than 66% accuracy on the test set. I have tried more than 50 different configurations of CNN layers and it's starting to feel like a waste of time.

The provided link to a VGG architecture example is not particularly helpful, as takes input image of size 224 224 3 (RGB image) and FER 2013 dataset is 48 48 1.

Can you confirm that it is possible to build a simple CNN that achieves more than 70% accuracy on the test set AND within reasonable computing time?

[jo-eman]

We are now reaching this project as part of our specialization and coming to the same conclusion.

Also during the audit this "optional" part of the project:

"Optional: (very cool) Hack the CNN. Take a picture for which the prediction of your CNN is Happy. Now, hack the CNN: using the same image SLIGHTLY modified make the CNN predict Sad. "

appears as mandatory and platform ask for reason to fail if you check the "no" box

Edit: Possible to pass 70% when transfer learning (method from suggested coursera course) was applied but does the project then still pass "is the model trained only on the training set" audit requirement? Maybe this can be clarified in project description or audit