错误：TypeError: Fetch参数None的类型无效

我认为这个错误是在使用回调函数ModelCheckpoint保存模型时发生的。查找这个错误时，找到了这个答案，但因为我在使用Keras，所以不能使用这个答案，因为我并没有在TensorFlow中显式调用sess.run()。此外，训练过程中的每个epoch都是完美无缺的，只有在保存模型时才会出现这个错误。

代码：

完整的模型位于Kaggle笔记本中，链接在这里：https://www.kaggle.com/aevinq/cnn-batchnormalization-0-1646/

错误出现的相关代码如下：

early_stopping = EarlyStopping(monitor='val_loss', patience=5, mode='min')mcp_save = ModelCheckpoint('md.hdf5', save_best_only=True, monitor='val_loss', mode='min')reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=5, verbose=1, epsilon=1e-4, mode='min')history = model.fit(train_X, train_y, batch_size=32, epochs=20, verbose=1, validation_split=0.25, callbacks=[early_stopping, reduce_lr_loss, mcp_save])

错误：

Train on 4413 samples, validate on 1471 samplesEpoch 1/204384/4413 [============================>.] - ETA: 1s - loss: 0.5157 - acc: 0.7696---------------------------------------------------------------------------TypeError                                 Traceback (most recent call last)<ipython-input-11-97f0757a1e9c> in <module>()      2 mcp_save = ModelCheckpoint('md.hdf5', save_best_only=True, monitor='val_loss', mode='min')      3 reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=5, verbose=1, epsilon=1e-4, mode='min')----> 4 history = model.fit(train_X, train_y, batch_size=32, epochs=20, verbose=1, validation_split=0.25, callbacks=[early_stopping, reduce_lr_loss, mcp_save])/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/models.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)    970                               initial_epoch=initial_epoch,    971                               steps_per_epoch=steps_per_epoch,--> 972                               validation_steps=validation_steps)    973     974     def evaluate(self, x=None, y=None,/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)   1655                               initial_epoch=initial_epoch,   1656                               steps_per_epoch=steps_per_epoch,-> 1657                               validation_steps=validation_steps)   1658    1659     def evaluate(self, x=None, y=None,/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/engine/training.py in _fit_loop(self, f, ins, out_labels, batch_size, epochs, verbose, callbacks, val_f, val_ins, shuffle, callback_metrics, initial_epoch, steps_per_epoch, validation_steps)   1231                             for l, o in zip(out_labels, val_outs):   1232                                 epoch_logs['val_' + l] = o-> 1233             callbacks.on_epoch_end(epoch, epoch_logs)   1234             if callback_model.stop_training:   1235                 break/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/callbacks.py in on_epoch_end(self, epoch, logs)     71         logs = logs or {}     72         for callback in self.callbacks:---> 73             callback.on_epoch_end(epoch, logs)     74      75     def on_batch_begin(self, batch, logs=None):/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/callbacks.py in on_epoch_end(self, epoch, logs)    413                             self.model.save_weights(filepath, overwrite=True)    414                         else:--> 415                             self.model.save(filepath, overwrite=True)    416                     else:    417                         if self.verbose > 0:/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/engine/topology.py in save(self, filepath, overwrite, include_optimizer)   2563         """   2564         from ..models import save_model-> 2565         save_model(self, filepath, overwrite, include_optimizer)   2566    2567     def save_weights(self, filepath, overwrite=True):/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/models.py in save_model(model, filepath, overwrite, include_optimizer)    145                 if symbolic_weights:    146                     optimizer_weights_group = f.create_group('optimizer_weights')--> 147                     weight_values = K.batch_get_value(symbolic_weights)    148                     weight_names = []    149                     for i, (w, val) in enumerate(zip(symbolic_weights,/opt/conda/lib/python3.6/site-packages/Keras-2.1.2-py3.6.egg/keras/backend/tensorflow_backend.py in batch_get_value(ops)   2208     """   2209     if ops:-> 2210         return get_session().run(ops)   2211     else:   2212         return []/opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)    887     try:    888       result = self._run(None, fetches, feed_dict, options_ptr,--> 889                          run_metadata_ptr)    890       if run_metadata:    891         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)/opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)   1103     # Create a fetch handler to take care of the structure of fetches.   1104     fetch_handler = _FetchHandler(-> 1105         self._graph, fetches, feed_dict_tensor, feed_handles=feed_handles)   1106    1107     # Run request and get response./opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in __init__(self, graph, fetches, feeds, feed_handles)    412     """    413     with graph.as_default():--> 414       self._fetch_mapper = _FetchMapper.for_fetch(fetches)    415     self._fetches = []    416     self._targets = []/opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in for_fetch(fetch)    232     elif isinstance(fetch, (list, tuple)):    233       # NOTE(touts): This is also the code path for namedtuples.--> 234       return _ListFetchMapper(fetch)    235     elif isinstance(fetch, dict):    236       return _DictFetchMapper(fetch)/opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in __init__(self, fetches)    339     """    340     self._fetch_type = type(fetches)--> 341     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]    342     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)    343 /opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in <listcomp>(.0)    339     """    340     self._fetch_type = type(fetches)--> 341     self._mappers = [_FetchMapper.for_fetch(fetch) for fetch in fetches]    342     self._unique_fetches, self._value_indices = _uniquify_fetches(self._mappers)    343 /opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in for_fetch(fetch)    229     if fetch is None:    230       raise TypeError('Fetch argument %r has invalid type %r' %--> 231                       (fetch, type(fetch)))    232     elif isinstance(fetch, (list, tuple)):    233       # NOTE(touts): This is also the code path for namedtuples.TypeError: Fetch参数None的类型无效 <class 'NoneType'>

回答：

这是Keras中的一个bug。最近的更新后，model.optimizer.weights中包含None值，这导致在保存模型时调用K.batch_get_value时出现错误。

我已经提交了一个PR来修复它，并且它已经被合并。你可以安装GitHub上的最新版本的Keras来解决这个问题。