Btukfyl

I am trying to use a Policy Network and CNN within the same Python file. I am only training the PN and using the CNN as a feed-forward feature extractor. I am getting an issue where the Train-PC isn't "building" the imported Keras graph into the session, and thus when I try to run a feed-forward function on the CNN (feature_vector = self.extract_feature_vector([X_img_norm])[0]), I get the error posted below.

The only difference between the files on the Home-PC I am developing on and the Train-PC is the way I import the model which I believe may be causing this (possibly it's not recognised). I posted previously about this error here.

Since the Train-PC is not my own, I cannot change the version of the packages. So I am looking for a solution which does not involve downgrading versions, as the issue with importing wasn't solved through that also.

Home-PC (Mac):

tensorflow==1.11.0

keras 2.2.4

Train-PC (Ubuntu):

tensorflow-gpu==1.10.1

keras 2.1.2

Code of CNN:

class ResNetCNN:



    def __init__(self):

        self.graph = tf.Graph()

        self.sess = tf.Session(graph=self.graph)



        with self.sess.graph.as_default():

            # Load pre-trained model from file

            self.CNN_ResNet_Pascal = models.load_model(CNN_MODEL_DIR)

            # self.CNN_ResNet_Pascal = tf.keras.models.load_model(CNN_MODEL_DIR)



            self.extract_feature_vector = K.function([self.CNN_ResNet_Pascal.layers[0].input],

                                                     [self.CNN_ResNet_Pascal.layers[-4].output])



    def feed_forward(self, img):

        """Extract the feature vector of the passed image using the pre-trained CNN."""



        # Resize observable region into input volume

        X_img = np.array(img, dtype=np.float32).reshape(-1, IMG_SIZE, IMG_SIZE, 3)



        # Normalise feature data

        X_img_norm = X_img / 255.0



        # Extract feature vector from the pre-trained CNN (1, 4096)

        feature_vector = self.extract_feature_vector([X_img_norm])[0]



        # Reshape tensor to (4096, )

        feature_vector = np.array(feature_vector).reshape(4096, )



        # Confirm that the correct output layer was used

        assert feature_vector.shape == (4096, ), "Incorrect CNN output layer: shape = {}".format(feature_vector.shape)



        return feature_vector

Error:

Traceback (most recent call last):

  File "keras_pn.py", line 856, in <module>

    s, a, r, d_r, n = rollout(epsilon, RENDER, PolicyNetwork, ResNetCNN)

  File "keras_pn.py", line 682, in rollout

    feature_vector = ResNetCNN.feed_forward(observable_region)

  File "keras_pn.py", line 87, in feed_forward

    feature_vector = self.extract_feature_vector([X_img_norm])[0]

  File "/home/name/anaconda3/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 2357, in __call__

    **self.session_kwargs)

  File "/home/name/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 877, in run

    run_metadata_ptr)

  File "/home/name/anaconda3/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1025, in _run

    raise RuntimeError('The Session graph is empty.  Add operations to the '

RuntimeError: The Session graph is empty.  Add operations to the graph before calling run().

Edit:

def __init__(self):

    self.graph = tf.Graph()



    with self.graph.as_default():

        # Load pre-trained model from file

        self.CNN_ResNet_Pascal = models.load_model(CNN_MODEL_DIR)

        # self.CNN_ResNet_Pascal = tf.keras.models.load_model(CNN_MODEL_DIR)



        self.extract_feature_vector = K.function([self.CNN_ResNet_Pascal.layers[0].input],

                                                 [self.CNN_ResNet_Pascal.layers[-4].output])



    self.sess = tf.Session(graph=self.graph)