Keras提前停止回调错误，val_loss指标不可用

Question

我正在训练Keras（在MacBook上为Tensorflow后端，Python），并且在fit_generator函数的早期停止回调中遇到错误。错误如下：

RuntimeWarning: Early stopping conditioned on metric `val_loss` which is not available. Available metrics are:
  (self.monitor, ','.join(list(logs.keys()))),
RuntimeWarning: Can save best model only with val_acc available, skipping.

'skipping.' % (self.monitor), RuntimeWarning
[local-dir]/lib/python3.6/site-packages/keras/callbacks.py:497: RuntimeWarning: Early stopping conditioned on metric `val_loss` which is not available. Available metrics are:
  (self.monitor, ','.join(list(logs.keys()))), RuntimeWarning
[local-dir]/lib/python3.6/site-packages/keras/callbacks.py:406: RuntimeWarning: Can save best model only with val_acc available, skipping.
  'skipping.' % (self.monitor), RuntimeWarning)
Traceback (most recent call last):
  :
  [my-code]
  :
  File "[local-dir]/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
return func(*args, **kwargs)
  File "[local-dir]/lib/python3.6/site-packages/keras/engine/training.py", line 2213, in fit_generator
callbacks.on_epoch_end(epoch, epoch_logs)
  File "[local-dir]/lib/python3.6/site-packages/keras/callbacks.py", line 76, in on_epoch_end
callback.on_epoch_end(epoch, logs)
  File "[local-dir]/lib/python3.6/site-packages/keras/callbacks.py", line 310, in on_epoch_end
self.progbar.update(self.seen, self.log_values, force=True)
AttributeError: 'ProgbarLogger' object has no attribute 'log_values'

我的代码如下（看起来不错）：

:
ES = EarlyStopping(monitor="val_loss", min_delta=0.001, patience=3, mode="min", verbose=1)
:
self.model.fit_generator(
        generator        = train_batch,
        validation_data  = valid_batch,
        validation_steps = validation_steps,
        steps_per_epoch  = steps_per_epoch,
        epochs           = epochs,
        callbacks        = [ES],
        verbose          = 1,
        workers          = 3,
        max_queue_size   = 8)

该错误消息似乎与提早停止的回调有关，但该回调看起来不错。该错误还指出val_loss不适合，但是我不确定为什么...关于这一点的另一件事是，该错误仅在使用较小的数据集时发生。

任何帮助表示赞赏。

Answer 1

如果错误仅在您使用较小的数据集时发生，则您很可能使用的数据集小到在验证集中没有单个样本。

因此它无法计算验证损失。

Answer 2

错误发生在我们身上，因为我们忘记在 fit() 方法中设置validation_data，而使用 'callbacks': [keras.callbacks.EarlyStopping(monitor='val_loss', patience=1)],

导致错误的代码是：

self.model.fit(
        x=x_train,
        y=y_train,
        callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss', patience=1)],
        verbose=True)

添加validation_data=(self.x_validate, self.y_validate),fit() 固定：

self.model.fit(
        x=x_train,
        y=y_train,
        callbacks=[keras.callbacks.EarlyStopping(monitor='val_loss', patience=1)],
        validation_data=(x_validate, y_validate),
        verbose=True)

Answer 3

我投票赞成前面的答案，因为它可以让我深入了解数据和fit_generator功能输入，并找出问题的根本原因。总之，在我的数据集小的情况下，我计算validation_steps和steps_per_epoch这竟然是零（0），这引起了错误。

我认为对于Keras团队来说，更好的长期答案是在fit_generator这些值均为零时导致错误/异常，这可能会导致对如何解决此问题有更好的理解。