66

Which is better for accuracy or are they the same? Of course, if you use categorical_crossentropy you use one hot encoding, and if you use sparse_categorical_crossentropy you encode as normal integers. Additionally, when is one better than the other?

Ethan
  • 1,625
  • 8
  • 23
  • 39
Master M
  • 763
  • 1
  • 6
  • 5

2 Answers2

65

Use sparse categorical crossentropy when your classes are mutually exclusive (e.g. when each sample belongs exactly to one class) and categorical crossentropy when one sample can have multiple classes or labels are soft probabilities (like [0.5, 0.3, 0.2]).

Formula for categorical crossentropy (S - samples, C - classess, $s \in c $ - sample belongs to class c) is:

$$ -\frac{1}{N} \sum_{s\in S} \sum_{c \in C} 1_{s\in c} log {p(s \in c)} $$

For case when classes are exclusive, you don't need to sum over them - for each sample only non-zero value is just $-log p(s \in c)$ for true class c.

This allows to conserve time and memory. Consider case of 10000 classes when they are mutually exclusive - just 1 log instead of summing up 10000 for each sample, just one integer instead of 10000 floats.

Formula is the same in both cases, so no impact on accuracy should be there.

featuredpeow
  • 766
  • 6
  • 6
  • 1
    Do they impact the accuracy differently, for example on mnist digits dataset? – Master M Dec 01 '18 at 08:47
  • 1
    Mathematically there is no difference. If there is significant difference in values computed by implementations (say tensorflow or pytorch), then this sounds like a bug. Simple comparison on random data (1000 classes, 10 000 samples) show no difference. – featuredpeow Dec 01 '18 at 14:20
  • Dear frenzykryger, I guess you forgot a minus for the one sample case only: "for each sample only non-zero value is just -log(p(s $\in$ c))". For the rest, nice answer. – Nicg Sep 13 '19 at 12:48
  • You're right. Thanks! – featuredpeow Sep 14 '19 at 13:54
  • @frenzykryger I am working on multi-output problem. I have 3 seperate output ```o1,o2,o3``` and each one have ```167,11,7``` classes respectively. I've read your answer that it'll make no difference but is there any difference if I use ```sparse__``` or not. Can I go for ```categorical``` for the last 2 and ```sparse``` for the first one as there are 167 classes in the first class? – Deshwal Jan 08 '20 at 04:58
38

The answer, in a nutshell

If your targets are one-hot encoded, use categorical_crossentropy. Examples of one-hot encodings:

[1,0,0]
[0,1,0] 
[0,0,1]

But if your targets are integers, use sparse_categorical_crossentropy. Examples of integer encodings (for the sake of completion):

1
2
3
Ethan
  • 1,625
  • 8
  • 23
  • 39
user78035
  • 381
  • 3
  • 2
  • 2
    Do I need a single output node for `sparse_categorical_crossentropy`? And what does the `from_logits` argument mean? – Leevo Dec 15 '19 at 17:01
  • 3
    @Leevo `from_logits=True` tells the loss function that an activation function (e.g. `softmax`) was **not** applied on the last layer, in which case your output needs to be as the number of classes. This is equivalent to using a `softmax` and `from_logits=False`. However, if you end up using `sparse_categorical_crossentropy`, make sure your target values are 1D. E.g. `[1, 1, 0, 1, ...]` (and **not** `[[1], [1], [0], [1], ...]`). On the other hand, if you use `categorical_crossentropy` and your target values are 1D, you need to apply `keras.utils(targets)` on them first to convert them to 2D. – Alaa M. Jul 03 '21 at 13:04
  • `keras.utils.to_categorical(targets)` * – Alaa M. Jul 06 '21 at 07:08
  • I experienced vastly different results between the two losses and finally saw here that the last dimension needs to be removed. Thanks! – N4ppeL Dec 06 '22 at 18:05