Created
October 12, 2020 03:54
-
-
Save iantimmis/f3cfa3953db4817c54c874b812e5e5f3 to your computer and use it in GitHub Desktop.
Numerically stable softmax with cross entropy in numpy
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import numpy as np | |
def naive_softmax(logits): | |
''' | |
Failure modes: | |
* If any entry is very large, exp overflows | |
* if all entries are very negative, all exps underflow | |
''' | |
exp_logits = np.exp(logits) | |
return exp_logits / np.sum(exp_logits) | |
def stable_softmax(logits): | |
''' | |
Mathematically equivalent to softmax. | |
''' | |
max_val = np.max(logits) | |
safe_exp_logits = np.exp(logits - max_val) | |
return safe_exp_logits / (max_val * np.sum(safe_exp_logits)) | |
def naive_softmax_with_cross_entropy(logits, t): | |
''' | |
Softmax plugged into categorical cross entropy | |
''' | |
probs = naive_softmax(logits) | |
return -np.sum(t * probs) | |
def stable_softmax_with_cross_entropy(logits, t): | |
''' | |
Mathematically equivalent to softmax with cross entropy | |
''' | |
max_val = np.max(logits) | |
safe_logits = logits - max_val | |
safe_logsumexp = max_val + np.log(np.sum(np.exp(safe_logits))) | |
return safe_logsumexp - np.sum(t * safe_logits |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment