custom_callbacks.py

from keras.callbacks import Callback
from sklearn.metrics import precision_recall_fscore_support


class Metrics(Callback):
    def on_train_begin(self, logs={}):
        self.valFMeasureHistory = []
        self.valRecallHistory = []
        self.valPrecisionHistory = []
        self.bestFMeasure = 0

    def on_epoch_end(self, epoch, logs={}):
        valPredictedLabels = self.model.predict(
            [self.validation_data[0], self.validation_data[1]])

        valLabels = self.validation_data[2]

        valPredictedBinaryLabels = valPredictedLabels.argmax(axis=1)
        valBinaryLabels = valLabels.argmax(axis=1)

        valPrecision, valRecall, valFMeasure, _ = precision_recall_fscore_support(
            valBinaryLabels, valPredictedBinaryLabels, average='binary')

        if valFMeasure is None:
            valFMeasure = 0.0

        if valFMeasure > self.bestFMeasure:
            print('Updating best model')
            print('Current best model comes from epoch {}'.format(str(epoch)))
            self.bestFMeasure = valFMeasure
            self.model.save('best-model.h5')

        self.valFMeasureHistory.append(valFMeasure)
        self.valRecallHistory.append(valRecall)
        self.valPrecisionHistory.append(valPrecision)

        valMessageFmt = "val_f1:{}\tval_precision: {}\tval_recall: {}"
        print(
            valMessageFmt.format(
                round(
                    valFMeasure, 2), round(
                    valPrecision, 2), round(
                    valRecall, 2)))
        print()

        return