当查准率和召回率都较高时，如何保存角星模型？

Question

使用python 3.7和tensorflow 2.1.0，我想知道如何在精确性和召回性都更好的时候保存最好的模型。

当只有精度更高时，下面的代码给出了保存最佳模型的元素。

# some dependencies
from tensorflow.keras.metrics import Precision, Recall

# some coding instructions to define ImageDataGenerator flows.

metrics = {"precision":Precision(name="precision"), "recall":Recall(name="recall")}
checkpoints = {
    "precision" : ModelCheckpoint(
        "./models/best.h5",
        mode="max",
        monitor="val_precision",
        save_freq="epoch",
        save_weights_only=False,
        save_best_only=True,
        verbose=1
    )
}
callbacks = list(checkpoints.values())

model = Sequential()
# neural network architecture ...
model.compile(
    loss="binary_crossentropy",
    optimizer='adam',
    metrics=[metric for metric in metrics.values()]
)

model.fit_generator(train_generator,
                    validation_data = val_generator,
                    steps_per_epoch = train_generator.n//train_generator.batch_size,
                    validation_steps = val_generator.n//val_generator.batch_size,
                    class_weight={0:0.75, 1:1.5},
                    callbacks=callbacks,
                    epochs=300)

我不认为为召回添加回调将使我能够为两个指标保存最好的模型.在我看来，如果召回比上个时代计算的召回要好的话，即使精度越来越差，模型也会被保存下来。精度也一样。我认为这种代码只会给我们一种“或”逻辑，而不是“和”逻辑。

也许我错了..。有人能帮我吗？任何帮助都会很感激的。非常欢迎解释！

Answer 1

我想出了一个解决办法，希望能帮到一些人。

class CustomCallback(tf.keras.callbacks.Callback):
    def __init__(self, model, path, metrics):
        self.model = model
        self.path = path
        self.metrics = metrics
        self.history = []

    def on_epoch_end(self, epoch, logs):
        if self.history:
            last_logs = self.history[-1]
            last_metrics_values = {metric : last_logs[metric] for metric in self.metrics}
            current_metrics_values = {metric : logs[metric] for metric in self.metrics}
            checking = [last_metrics_values[metric] < current_metrics_values[metric] for metric in self.metrics]
            decision = reduce(operator.and_, checking)
            if decision:
                print("\nModel performs better at current epoch for all monitored metrics : ", self.metrics)
                print("\nSaving current model in ", self.path)
                self.model.save(self.path)
        else:
            print("\nSaving model at first epoch for initialization purpose")
            self.model.save(self.path)
        self.history.append(logs)

您可以更改代码，使其具有更复杂的逻辑，以决定何时保存模型，以及对您来说最好的模型是什么。决策采用checking和decision进行计算。您必须具有以下依赖关系：

from functools import reduce
import operator

import tensorflow as tf