Applying softmax stability fix to cost gradient.
This commit is contained in:
parent
b4b00e0c65
commit
a021ce585b
@ -170,7 +170,9 @@ class CrossEntropySoftmaxError(object):
|
||||
Returns:
|
||||
Gradient of error function with respect to outputs.
|
||||
"""
|
||||
probs = np.exp(outputs)
|
||||
# subtract max inside exponential to improve numerical stability -
|
||||
# when we divide through by sum this term cancels
|
||||
probs = np.exp(outputs - outputs.max(-1)[:, None])
|
||||
probs /= probs.sum(-1)[:, None]
|
||||
return (probs - targets) / outputs.shape[0]
|
||||
|
||||
|
Loading…
Reference in New Issue
Block a user