Skip to content

Loss Functions and Metrics

Frank Seide edited this page Aug 25, 2016 · 23 revisions

CNTK contains a number of common predefined loss functions. In addition, custom loss functions can be defined as BrainScript expressions.

CrossEntropy(), CrossEntropyWithSoftmax()

Computes the categorical cross-entropy loss (or just the cross entropy between two probability distributions).

CrossEntropy (y, p)
CrossEntropyWithSoftmax (y, z)

Parameters

  • y: labels (one-hot), or more generally, reference distribution. Must sum up to 1.
  • p (for CrossEntropy()): posterior probability distribution to score against the reference. Must sum up to 1.
  • z (for CrossEntropyWithSoftmax()): input to a Softmax operation to compute the posterior probability distribution to score against the reference

Return value

This operation computes the cross-entropy between two probability distributions y and p, defined as:

ce = E_y{-log p} = -sum_i y_i log p_i

with i iterating over all elements of y and p. For CrossEntropyWithSoftmax(), p is computed from the input parameter z as

p = Softmax (z)

Description

These functions compute the cross-entropy of two probability distributions. Categorical cross-entropy is the most common training criterion (loss function) for single-class classification, where y encodes a categorical label as a one-hot vector. Another use is as a loss function for probability distribution regression, where y is a target distribution that p shall match.

When used as a loss function, CNTK's SGD process will sum up the cross-entropy values of all individual samples in a minibatch, and compute the per-epoch loss by aggregating these over an entire epoch. This is then reported in this form: L * S where L is the average loss per sample, and S the number of samples. For example, -1.213 * 3600000, means an average cross-entropy loss of -1.213, averaged over an epoch of 3,600,000 samples.

CrossEntropyWithSoftmax() is an optimization for the most common use case of categorical cross-entropy, which takes advantage of the specific form of Softmax. Instead of a normalized probability, it accepts as its input the argument to the Softmax operation, which is the same as a non-normalized version of log Softmax, also known as "logit". This is the recommended way in CNTK to compute the categorical cross-entropy criterion.

Note that categorical cross-entropy is not a suitable loss function for multi-class labels, where y contains more than one position containing a 1. For this case, consider using Sigmoid() instead of Softmax, with a Logistic() loss. See also this article.

Alternative definition

CrossEntropyWithSoftmax() is currently a CNTK primitive which has limitations. A more flexible, recommended alternative is to define it manually as:

CrossEntropyWithSoftmax (y, z) = ReduceLogSum (z) - TransposeTimes (y, z)

Sparse labels

To compute the cross entropy with sparse labels (e.g. read using Input(./Inputs#input){..., sparse=true}), the alternative form above must be used.

Softmax over tensors with rank>1

To compute CrossEntropyWithSoftmax() over When applied to tensors of rank>1, e.g. where the task is to determine a location on a 2D grid, yet another alternative form must be used:

CrossEntropyWithSoftmax (y, z, axis=None) = ReduceLogSum (z, axis=axis) - ReduceSum (y .* z, axis=axis)

This form also allows to apply the Softmax operation along a specific axis only. For example, if the inputs and labels have the shape [10 x 20], and the Softmax should be computed over each of the 20 columns independently, specify axis=1.

Example

labels = Input {9000}
...
z = W * h + b
ce = CrossEntropyWithSoftmax (labels, z)
criterionNodes = (ce)

The same with sparse labels:

labels = Input {9000, sparse=true}
...
z = W * h + b
ce = ReduceLogSum (z) - TransposeTimes (labels, z)
criterionNodes = (ce)

Logistic{}, WeightedLogistic{}

Computes the logistic loss function.

Logistic (y, p)
WeightedLogistic (y, p, instanceWeight)

Parameters

  • y: ground-truth label, 0 or 1
  • p: posterior probability of being of class 1

Return value

Computes the (weighted) logistic loss, defined as:

ll = -sum_i { y_i * log(p_i) + (1-y_i)*log(1-p_i)) * weight }

(where for Logistic(), the weight is 1).

Description

This function is the logistic loss function.

Logistic (y, Softmax (z)) is identical to CrossEntropyWithSoftmax (y, z) for two-class problems, where only one of the two (complementary) ground-truth labels is given to Logistic().

See also this article on training multi-class classifiers.

Example

multiLabels = Input {1000}                     # 0 or 1 for each value
...
p = DenseLayer {1000, activation=Sigmoid} (h)  # element-wise predictor for 1000 values
ll = Logistic (multiLabels, p)
trainingCriterion = (ll)

ErrorPrediction{}

Computes the error rate for prediction of categorical labels.

ErrorPrediction (y, z)

Parameters

  • y: categorical labels in one-hot form
  • z: vector of prediction scores, e.g. log probabilities

Return value

1 if the maximum value of z is at a position where y has a 1; 0 otherwise; summed up over an entire minibatch.

Description

This function accepts a vector of posterior probabilities, logits, or other matching scores, where each element represents the matching score of a class or category. The function determines whether the highest-scoring class is equal to the class denoted by the label input y, by testing whether the highest-scoring position contains a 1.

When used as an evaluation criterion, the SGD process will aggregate all values over an epoch and report the average, i.e. the error rate.

ErrorPrediction() cannot be used with sparse labels.

Example

labels = Input {9000}
...
z = W * h + b
errs = ErrorPrediction (labels, z)
evaluationNodes = (errs)
Clone this wiki locally