Keras: new keras model to tf Estimator error

[handongfeng]

Hi there,
start a new keras model and try to convert to tf Estimator
get error valueerror: expected model argument to be a model instance, got xxxxxxxxxxxxxxx

any suggestion? thanks
here is simple example

# set model
model =  Sequential()
model.add(Dense(8, input_dim=4))
model.add(Activation('relu'))
model.add(Dense(3, input_dim=8))
model.add(Activation('softmax'))
model.compile(optimizer='SGD', loss='categorical_crossentropy', metrics=['accuracy'])
model_dir = os.path.join(os.getcwd(), "models//123").replace("//", "\\")
estimator_model = tf.keras.estimator.model_to_estimator(keras_model=model,model_dir=model_dir)

[x831617]

I used the Keras Functional API and I encountered the same issue.

[dheepanr]

The to_estimator does a type check on the Model object. Unfortunately, the tensorflow version of the Model object comes from deep within the bowels of TF in the version I am using (1.4.0). So a default keras Model object will be rejected.

I changed my model import to the following:
from tensorflow.python.keras._impl.keras.models import Model
That fixed my issue.

I believe that all the structures have been changed in 1.5, so I would be careful when upgrading.

[handongfeng]

Hi @x831617 @dheepanr
thanks
yes, not use keras. Using everything from tf such as tf.keras.layers, tf.keras.layers.Conv1D ,..
model = tf.keras.models.Model(inp,out_final). to define your model should be fine.
so many tricks. Cost much time... but solved finally. thanks all.

[purohit]

I have the same issue when using an Inception-V3 net I modified, with code similar to the example (Keras 2.1.4, TF 1.5.0 on CPU):

    # Instantiate a Keras inception v3 model.                                                    
    inception_v3_net = inception.InceptionV3(                                                    
            input_shape=input_shape,                                                             
            classes=y_classes,                                                                   
            )
    # Compile model with the optimizer, loss, and metrics you'd like to train with.              
    inception_v3_net.compile(
            optimizer=tf.train.AdamOptimizer(),                                                  
            loss='categorical_crossentropy',                                                     
            metric='accuracy')
    # Create an Estimator from the compiled Keras model. Note the initial model                  
    # state of the keras model is preserved in the created Estimator.                            
    estimator = tf.keras.estimator.model_to_estimator(                                           
            keras_model=inception_v3_net,                                                        
            model_dir=None)

Traceback (most recent call last):
  File "load_test.py", line 79, in <module>
    tf.app.run(main=run_network)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 124, in run
    _sys.exit(main(argv))
  File "load_test.py", line 73, in run_network
    model_dir=None)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/estimator.py", line 292, in model_to_estimator
    _save_first_checkpoint(keras_model, est, custom_objects, keras_weights)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/estimator.py", line 218, in _save_first_checkpoint
    custom_objects)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/estimator.py", line 102, in _clone_and_build_model
    model = models.clone_model(keras_model, input_tensors=input_tensors)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/models.py", line 1557, in clone_model
    return _clone_functional_model(model, input_tensors=input_tensors)
  File "/home/purohit/.local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/models.py", line 1361, in _clone_functional_model
    'to be a `Model` instance, got ', model)
ValueError: ('Expected `model` argument to be a `Model` instance, got ', <keras.engine.training.Model object at 0x7f76223f6d68>)

[calpeyser]

Try importing the tensorflow implementation of Keras instead of the standard?

from tensorflow.python.keras._impl import keras

Then, keras.models.Model will in fact be the same class that the instanceof check is looking for.

[purohit]

That makes sense. What is the plan for this -- is that recommended practice now if we want to use Keras with TensorFlow?

[schulz0r]

Hello calpeyser,
Using tensorflow's implementation of keras did not completely solve the problem, because I use a model with custom layers. I observed that tensorflow's keras version does not call the compute_output_shape function of my custom layers, even though this is crucial for my unpooling layer.
For me, it worked to simply cast the keras model to the tensorflow one:

myKerasModel = build_keras_model(...) # put your keras model here
tfCompatibleMod = tf.keras.models.Model(myKerasModel) # cast to type tensorflow expect
tfCompatibleMod.compile(...) # compile
estimator = tf.keras.estimator.model_to_estimator(keras_model=tfCompatibleMod)

I hope to help anyone who finds it here.

Edit: I encountered bugs when trying to train the network with a casted keras model, for which no solution could be found. According to other reported issues, there is a problem with some cache introduced in the more recent Keras versions, which stores the network and introduces incompatiblities with tensorflow. I will stop using Keras at this point.

[aspratyush]

@schulz0r Thanks for the casting tip. In TF 1.8, the module is now tf.keras.Model However, as observed at your end, casting the keras model as a tf.keras.Model is then resulting in AttributeError: 'Model' object has no attribute '_output_tensor_cache', with the traceback:

Traceback (most recent call last):
  File "mnist_lenet_keras.py", line 44, in <module>
    main()
  File "mnist_lenet_keras.py", line 40, in main
    keras_estimator.run(model, X, Y_oh, nb_epochs=5)
  File "/home/pratyush/workspace/DL/dl_utils/tf/estimators/keras_estimator.py", line 43, in run
    model_est.train(input_fn=train_input_fn, steps=nb_batches)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 363, in train
    loss = self._train_model(input_fn, hooks, saving_listeners)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 843, in _train_model
    return self._train_model_default(input_fn, hooks, saving_listeners)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 856, in _train_model_default
    features, labels, model_fn_lib.ModeKeys.TRAIN, self.config)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/estimator/estimator.py", line 831, in _call_model_fn
    model_fn_results = self._model_fn(features=features, **kwargs)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/estimator.py", line 330, in model_fn
    labels)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/estimator.py", line 291, in _clone_and_build_model
    model._set_inputs(input_tensors)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/training.py", line 873, in _set_inputs
    self._symbolic_set_inputs(inputs, training=training)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/training.py", line 992, in _symbolic_set_inputs
    outputs = self.call(self.inputs[0], training=training)
  File "/home/pratyush/workspace/virtualenv/tf/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/network.py", line 628, in call
    if cache_key in self._output_tensor_cache:
AttributeError: 'Model' object has no attribute '_output_tensor_cache'

[JackLiar]

so how should we solve this problem @aspratyush ?

there are something interesting things I wanna share about this issue

I'm using a StackAutoEncoderClassifier implementation like this

in a nutshell:
AutoEncoder(Sequential) -> StackedAutoEncoder(Sequential) -> StackedAutoEncoderClassifier(Sequential)

when I try to train the classifier, the '_output_tensor_cache' happened, but no error happened when training each AE respectively or training the SAE inside the SAEC

I use pdb to check each model, SAE has _output_tensor_cache, but none for SAEC, it's pretty weird.......

[KhalilGorsan]

Using Tensorflow backend and keras is the appropriate way to allow converting the keras model to an Estimator. If you find a model implemented in keras , it's easy to make some ajustements and build the blocks using tf.keras.

[zmlim]

@schulz0r Hi I'm trying to prune some layers of a custom model too (YOLOv3). May I know if you solved the issue?

[schulz0r]

@zmlim No, I simply stopped using Keras.

[mayowaosibodu]

@zmlim No, I simply stopped using Keras.

I'm curious, what do you use now?

[schulz0r]

@mayowaosibodu I used TF estimators.