1

I have a python code running KERAS LSTM model perfectly fine when run directly using the python code (either through SPYDER or via Command Prompt).

Python (anaconda) : 3.6.7 FLASK : 0.12.3 KERAS : 2.2.4 TENSORFLOW : 1.10.0

When trying to run it as a FLASK app, the same code throws the below error : Error : The Session graph is empty. Add operations to the graph before calling run().

I had to set session and set seed so that the model results dont vary from different runs. I have used the code from Kera's docs (https://keras.io/getting-started/faq/#how-can-i-obtain-reproducible-results-using-keras-during-development) in my code and it runs absolutely fine when run as a simple python script.

The problem comes only when trying to run it in a Flask app (which isnt making sense).

The error in code comes at the line where I try to load the model from a saved h5 file.

Can someone help me with this problem ?

Have read mostly all the similar issues from google and StackOverflow but none of the options worked out.

I also checked KERAS's github repo and there seems to be the same problem open as a ticket since 1 year. (https://github.com/keras-team/keras/issues/10585)

FLASK app code (app.py) :

from flask import Flask, jsonify

from PREDICT import worker

app = Flask(__name__)

@app.route('/api/Model', methods=['GET', 'POST'])
def Model_predict():
    worker()
    return ('Model run')

if __name__ == '__main__':
    app.run(debug=True)

Code I used to set session (taken from KERAS's documentation site) that is present in PREDICT.py : **

# The below is necessary for starting Numpy generated random numbers in a well-defined initial state.
np.random.seed(42)
# The below is necessary for starting core Python generated random numbers in a well-defined state.
rn.seed(12345)

# Force TensorFlow to use single thread.
# Multiple threads are a potential source of non-reproducible results.
# For further details, see: https://stackoverflow.com/questions/42022950/
session_conf = tf.ConfigProto(intra_op_parallelism_threads=1,
                              inter_op_parallelism_threads=1)
from keras import backend as K
# The below tf.set_random_seed() will make random number generation in the TensorFlow backend have a well-defined initial state.
# For further details, see: https://www.tensorflow.org/api_docs/python/tf/set_random_seed

tf.set_random_seed(1234)

sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)
K.set_session(sess)
**

Expected results are that the model file should be loaded properly to be used for prediction later.

Details of Error :

**
2019-05-15 00:19:19,121 ERROR  : Error at Line : 363
2019-05-15 00:19:19,122 ERROR  : Error : The Session graph is empty.  Add operations to the graph before calling run().
2019-05-15 00:19:19,123 ERROR Stack Trace : 
Traceback (most recent call last):
  File "PREDICT.py", line 363, in worker
    model = loading_model(output_path, weekno, prod)
  File "PREDICT.py", line 104, in loading_model
    m = load_model(model_name)
  File "C:\Users\12345\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\engine\saving.py", line 419, in load_model
    model = _deserialize_model(f, custom_objects, compile)
  File "C:\Users\12345\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\engine\saving.py", line 287, in _deserialize_model
    K.batch_set_value(weight_value_tuples)
  File "C:\Users\12345\AppData\Local\Continuum\anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py", line 2470, in batch_set_value
    get_session().run(assign_ops, feed_dict=feed_dict)
  File "C:\Users\12345\AppData\Local\Continuum\anaconda3\lib\site-packages\tensorflow\python\client\session.py", line 877, in run
    run_metadata_ptr)
  File "C:\Users\12345\AppData\Local\Continuum\anaconda3\lib\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().
**
Shankar Pandey
  • 451
  • 1
  • 4
  • 22

1 Answers1

0

After thorough searhing, i came across this blog which helped me understand how Flask and Keras operate in tandem.

https://towardsdatascience.com/deploying-keras-deep-learning-models-with-flask-5da4181436a2

The solution (if any one needs it later) is that you need to do 2 things :

  1. define the graph as global
global graph
graph = tf.get_default_graph()
sess = tf.Session(graph=graph, config=session_conf)
  1. before loading model and predicting, wrap it under a loop so that a new graph is generated for each loop 
with graph.as_default():

note: If you get some error like "ValueError: signal only works in main thread", check your FLask-SocketIO package - you may need to uninstall and re-install it with the latest version .

Shankar Pandey
  • 451
  • 1
  • 4
  • 22
  • I'm facing the same problem. What do you mean by wrapping the model under a loop ? Thanks – Arthur Clerc-Gherardi Dec 17 '19 at 14:03
  • @ArthurClerc-Gherardi - in my case, i had to run the model for each Product in a loop, hence the comment. you can do this without the loop just enclose your code within the WITH clause as mentioned above and it should work fine. – Shankar Pandey Dec 19 '19 at 06:21
  • mmh.. ok. Do you have an open github for this project ? I'd like to see the whole implementation. I'm stucking using the same model for every http request. – Arthur Clerc-Gherardi Dec 19 '19 at 10:34