如何修复使用InceptionV4时出现的'tensorflow'无'get_default_graph'报错

问题背景 #

使用InceptionV4代码时出现上述错误，TensorFlow和Keras版本均为2.5.0。

错误信息 #

AttributeError
module 'tensorflow' has no attribute 'get_default_graph'
File "C:\Labbb\inception\InceptionV4-master\inceptionv4.py", line 219, in create_inception_v4
init = Input((299,299, 8))
File "C:\Labbb\inception\InceptionV4-master\inceptionv4.py", line 259, in
model = create_inception_v4(load_weights=check)
AttributeError: module 'tensorflow' has no attribute 'get_default_graph'

相关代码片段 #

def create_inception_v4(nb_classes=int(args["num_classes"]), load_weights=check):
model = create_inception_v4(load_weights=check)
model.compile(loss='categorical_crossentropy', optimizer=keras.optimizers.SGD(lr=float(args['learning_rate']), decay=1e-6, momentum=0.9, nesterov=True), metrics=["accuracy"])
hist = model.fit_generator(train_datagen,steps_per_epoch=int(args['50']),epochs=int(args['3']),verbose=True,validation_data=val_datagen,validation_steps=10,callbacks=[mc, tensorboard])

当前导入代码 #

from keras.layers import Input, merge, Dropout, Dense, Flatten, Activation
from keras.layers.convolutional import MaxPooling2D, Convolution2D, AveragePooling2D
from keras.layers.normalization import BatchNormalization
from tensorflow.keras.models import Model
from tensorflow.keras import backend as K
from keras.utils.data_utils import get_file
from keras.utils.training_utils import multi_gpu_model
import tensorflow as tf
import keras
from tensorflow.keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
from tensorflow.keras.layers import concatenate
from tensorflow.keras.callbacks import TensorBoard
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.models import Sequential
from tensorflow.keras.utils import Sequence

解决方法 #

问题出在混用了独立Keras和TensorFlow内置Keras的导入，TensorFlow 2.x里已经删掉了tf.get_default_graph()的直接调用，但独立Keras的部分组件还会试图调用这个已废弃的API。

按下面的步骤修改即可：

1. 统一所有导入为tensorflow.keras，删除所有直接从keras.开头的导入，全部替换成tensorflow.keras.：

   # 替换所有keras.开头的导入
   from tensorflow.keras.layers import Input, Dropout, Dense, Flatten, Activation, MaxPooling2D, Conv2D, AveragePooling2D, BatchNormalization, concatenate
   from tensorflow.keras.models import Model, Sequential
   from tensorflow.keras import backend as K
   from tensorflow.keras.utils import get_file, to_categorical, Sequence
   from tensorflow.keras.optimizers import SGD
   from tensorflow.keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img
   from tensorflow.keras.callbacks import TensorBoard
   import tensorflow as tf
   

   注意：Convolution2D在tensorflow.keras中已更名为Conv2D，merge已被concatenate替代，代码中对应调用也要同步修改。

2. 调整代码中的API调用：

   • 将Convolution2D替换为Conv2D
   • 将merge替换为concatenate（确保已从tensorflow.keras.layers导入该方法）
   • 将keras.optimizers.SGD改为tensorflow.keras.optimizers.SGD

3. 卸载独立Keras包：如果环境中安装了独立Keras，建议卸载避免版本冲突：

   pip uninstall keras -y
   

4. 替换fit_generator为fit：TensorFlow 2.2+版本中fit_generator已被弃用，直接使用model.fit()即可，它会自动处理生成器输入：

   hist = model.fit(train_datagen, steps_per_epoch=int(args['50']), epochs=int(args['3']),
                    verbose=True, validation_data=val_datagen, validation_steps=10,
                    callbacks=[mc, tensorboard])
   

这样修改后，就能彻底解决混用API导致的错误，适配TensorFlow 2.5.0的环境。

内容的提问来源于stack exchange，提问作者Syuuuu