Tensorboard AttributeError: 'ModelCheckpoint' object has no attribute 'on_train_batch_begin'

问题

I'm currently using Tensorboard using the below callback as outlined by this SO post as shown below.

from keras.callbacks import ModelCheckpoint

CHECKPOINT_FILE_PATH = '/{}_checkpoint.h5'.format(MODEL_NAME)
checkpoint = ModelCheckpoint(CHECKPOINT_FILE_PATH, monitor='val_acc', verbose=1, save_best_only=True, mode='max', period=1)

When I run Keras' dense net model, I get the following error. I haven't had any issues running Tensorboard in this manner with any of my other models, which makes this error very strange. According to this Github post, the official solution is to use the official Tensorboard implementation; however, this requires upgrading to Tensorflow 2.0, which is not ideal for me. Anyone know why I'm getting the following error for this specific densenet and is there a workaround/fix that someone knows?

AttributeError Traceback (most recent call last) in () 26 batch_size=32, 27 class_weight=class_weights_dict, ---> 28 callbacks=callbacks_list 29 ) 30

2 frames /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/callbacks.py in _call_batch_hook(self, mode, hook, batch, logs) 245 t_before_callbacks = time.time() 246 for callback in self.callbacks: --> 247 batch_hook = getattr(callback, hook_name) 248 batch_hook(batch, logs) 249 self._delta_ts[hook_name].append(time.time() - t_before_callbacks)

AttributeError: 'ModelCheckpoint' object has no attribute 'on_train_batch_begin'

The dense net I'm running

from tensorflow.keras import layers, Sequential
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.applications.densenet import preprocess_input, DenseNet121
from keras.optimizers import SGD, Adagrad
from keras.utils.np_utils import to_categorical

IMG_SIZE = 256
NUM_CLASSES = 5
NUM_EPOCHS = 100

x_train = np.asarray(x_train)
x_test = np.asarray(x_test)

y_train = to_categorical(y_train, NUM_CLASSES)
y_test = to_categorical(y_test, NUM_CLASSES)


x_train = x_train.reshape(x_train.shape[0], IMG_SIZE, IMG_SIZE, 3)
x_test = x_test.reshape(x_test.shape[0], IMG_SIZE, IMG_SIZE, 3)

densenet = DenseNet121(
    include_top=False,
    input_shape=(IMG_SIZE, IMG_SIZE, 3)
)

model = Sequential()
model.add(densenet)
model.add(layers.GlobalAveragePooling2D())
model.add(layers.Dense(NUM_CLASSES, activation='softmax'))
model.summary()

model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['accuracy'])

history = model.fit(x_train,
                    y_train,
                    epochs=NUM_EPOCHS,
                    validation_data=(x_test, y_test),
                    batch_size=32,
                    class_weight=class_weights_dict,
                    callbacks=callbacks_list
                   )

回答1:

In your imports you are mixing keras and tf.keras, which are NOT compatible with each other, as you get weird errors like these.

So a simple solution is to choose keras or tf.keras, and make all imports from that package, and never mix it with the other.

回答2:

Make all imports from either keras or tensorflow.keras

I hope this will sort it out!

回答3:

Yes imports are mixed from keras and tensorflow

try sticking on to tensorflow.keras for example :

from tensorflow.keras.callbacks import EarlyStopping

回答4:

I replace this line

from keras.callbacks import EarlyStopping, ModelCheckpoint

To this line

from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint

来源：https://stackoverflow.com/questions/57122907/tensorboard-attributeerror-modelcheckpoint-object-has-no-attribute-on-train

标签

python

tensorflow

keras

tensorboard