AttributeError: 'Node' object has no attribute 'output_masks'

Question

I use Keras pretrained model VGG16. The problem is that after configuring tensorflow to use the GPU I get an error that I didn't have before when using the CPU.

The error is the following one:

    Traceback (most recent call last):
  File "/home/guillaume/Documents/Allianz/ConstatOrNotConstatv3/train_network.py",      line 109, in <module>
    model = LeNet.build(width=100, height=100, depth=3, classes=5)
  File "/home/guillaume/Documents/Allianz/ConstatOrNotConstatv3/lenet.py", line 39,    in build
    output = model(pretrainedOutput)
  File "/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py", line 443, in __call__
    previous_mask = _collect_previous_mask(inputs)
  File "/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py", line 1311, in _collect_previous_mask
mask = node.output_masks[tensor_index]
  AttributeError: 'Node' object has no attribute 'output_masks'

I get it after executing this code :

    pretrained_model = VGG16(
        include_top=False,
        input_shape=(height, width, depth),
        weights='imagenet'
    )
    for layer in pretrained_model.layers:
        layer.trainable = False

    model = Sequential()
    # first (and only) set of FC => RELU layers
    model.add(Flatten())
    model.add(Dense(200, activation='relu'))
    model.add(Dropout(0.5))
    model.add(BatchNormalization())
    model.add(Dense(400, activation='relu'))
    model.add(Dropout(0.5))
    model.add(BatchNormalization())

    # softmax classifier
    model.add(Dense(classes,activation='softmax'))

    pretrainedInput = pretrained_model.input
    pretrainedOutput = pretrained_model.output
    output = model(pretrainedOutput)
    model = Model(pretrainedInput, output)

EDIT1 : I've got keras (2.2.2) and tensorflow(1.10.0rc1). I've also tried on keras 2.2.0 and same error. The thing is that the python environment I use works on others non-pretrained NN.

EDIT2 : I'm able to connect two homemade models. It's only whith the pretrained ones there is a problem and not only VGG16.

Answer 1

You're likely importing tf.keras.layers or tf.keras.applications or other keras modules from tensorflow.keras, and mixing these objects with objects from the "pure" keras package, which is not compatible, based upon version, etc.

I recommend seeing if you can import and run everything from the "pure" keras modules; don't use tf.keras while debugging, as they're not necessarily compatible. I had the same problem, and this solution is working for me.

Answer 2

I had the same error when I import keras and tenerflow.keras simultaneously: from tensorflow.keras.optimizers import Adam from keras.utils import multi_gpu_model

I solved this problem after changing the code into: from tensorflow.keras.optimizers import Adam from tensorflow.keras.utils import multi_gpu_model

Answer 3

I had a similar issue, but with different architecture. As people suggested, it's important not to mix keras with tensorflow.keras, so try swapping code like:

from keras.preprocessing import image
from keras.models import Model
from keras.layers import Dense, GlobalAveragePooling2D
from keras import backend as K

to:

from tensorflow.keras.preprocessing import image 
from tensorflow.keras.models import Model 
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D 
from tensorflow.keras import backend as K

Also make sure, you don't use keras.something inside your code (not only imports) as well, hope it helps : ) Also, I used Keras 2.2.4 with tensorflow 1.10.0