可以传递给 tf.keras.model.compile 的指标列表

Question

Based on the tensorflow documentation , when compiling a model, I can specify one or more metrics to use, such as 'accuracy' and 'mse'. However, the documentation doesn't say what metrics are available. I tried to replace 'accuracy' with a few other classical metrics such as 'recall' or 'auc', but that didn't work.

model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['recall'])

What metrics are available? Where can I find a list of all the metrics keywords I can use?

Answer 1

You have the following (usually with relation to a classification task)

1. Accuracy via:

    keras.metrics.accuracy(y_true, y_pred)

2. Binary Accuracy given a certain thershold:

    keras.metrics.binary_accuracy(y_true, y_pred, threshold=0.5)

3. Categorical Accuracy via:

    keras.metrics.categorical_accuracy(y_true, y_pred)

4. Sparse Categorical Accuracy via:

    keras.metrics.sparse_categorical_accuracy(y_true, y_pred)

5. Sparse Categorical Accuracy given a certain k:

    keras.metrics.top_k_categorical_accuracy(y_true, y_pred, k=5)

6. Sparse Categorical Accuracy given a certain k:

    keras.metrics.top_k_categorical_accuracy(y_true, y_pred, k=5)

7. Cosine Proximity given a certain axis:

    keras.metrics.cosine_proximity(y_true, y_pred, axis=-1)

In addition to the metrics above, you may use any of the loss functions described in the loss function page as metrics. You may also implement your own custom metric, for example:

import keras.backend as K

def mean_pred(y_true, y_pred):
    return K.mean(y_pred)

model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['accuracy', mean_pred])

Reference: Keras Metrics , Keras Loss Functions

Answer 2

尝试使用此处的指标之一： https : //keras.io/metrics/