ValueError: Tensor Tensor(…) is not an element of this graph. When using global variable keras model

Question

I'm running a web server using flask and the error comes up when I try to use vgg16, which is the global variable for keras' pre-trained VGG16 model. I have no idea why this error rises or whether it has anything to do with the Tensorflow backend. Here is my code:

vgg16 = VGG16(weights='imagenet', include_top=True)

def getVGG16Prediction(img_path):
    global vgg16

    img = image.load_img(img_path, target_size=(224, 224))
    x = image.img_to_array(img)
    x = np.expand_dims(x, axis=0)
    x = preprocess_input(x)

    pred = vgg16.predict(x)
    return x, sort(decode_predictions(pred, top=3)[0])

@app.route("/uploadMultipleImages", methods=["POST"])
def uploadMultipleImages():
    uploaded_files = request.files.getlist("file[]")
    for file in uploaded_files:
        path = os.path.join(STATIC_PATH, file.filename)
        pInput, result = getVGG16Prediction(path)

Here is the full error:

Any comment or suggestion is greatly appreciated. Thank you.

Answer 1

Take a look at avital's answer on this github issue. Quoting the relevant part here:

Right after loading or constructing your model, save the TensorFlow graph:

graph = tf.get_default_graph()

In the other thread (or perhaps in an asynchronous event handler), do:

global graph
with graph.as_default():
    (... do inference here ...)

I modified this a bit and stored the graph in my app's config object instead of making it a global.

The TensorFlow documentation for get_default_graph explains why this is necessary:

NOTE: The default graph is a property of the current thread. If you create a new thread, and wish to use the default graph in that thread, you must explicitly add a with g.as_default(): in that thread's function.