Objectosphere is computed on the softmax prediction, but not on the embedding layer

[siebenkopf]

When I understand the loss function:

Reducing-Network-Agnostophobia/MNIST/Mnist_Training.py

Line 69 in 581b3f0

pred=K.sqrt(K.sum(K.square(y_pred),axis=1))

correctly, the input to your loss function are the one-hot-encoded target values (y_true) and the estimated softmax probabilities (y_pred). From this y_pred, you compute the magnitude pred for each sample. However, this magnitude should be computed from the output of the layer before the logits, not from the softmax probabilities.

This might be the reason, why there is very little difference between the Entropic Open-Set loss and the Objectosphere loss.

[siebenkopf]

Sorry, I should always try to read the whole thing before writing a bug report. In

Reducing-Network-Agnostophobia/MNIST/Mnist_Training.py

Line 199 in 581b3f0

loss={'softmax': 'categorical_crossentropy','fc':ring_loss},

your clearly apply this function on the fc layer, as expected.

You should consider renaming the y_pred parameter of the function call to fc or something alike to avoid confusion.