Keras model errors out on prediction

Hi,

I’m trying to replicate a CSRNET model for counting people in images, it works good for the first input , but errors out when I change to second image.

I did add the allow_mutation_inputs tag to the cache which removed the st.cache warnings

def run_model(path): @st.cache(allow_output_mutation=True) def load_model(): # Function to load and return neural network model json_file = open('/models/Model.json', 'r') loaded_model_json = json_file.read() json_file.close() loaded_model = model_from_json(loaded_model_json) loaded_model.load_weights("CSRNet-keras/weights/weights1.hdf5") return loaded_model

Below is the error I get:

Output Console Log:

Not sure if this is because of caching or I’m missing something. It works fine separately on a Jupyter Notebook.

Debug info

• Streamlit version: 0.50.2
• Python version: 3.6.5
• Using Conda
• OS version: Ubuntu 16.04
• Browser version: Google Chrome Version 79

Thanks !

[mohammedayub44]

Also, somtimes it break giving “Tensor X is not a part of the graph…”. Guessing this is because of multi-threading in Keras. Running the loaded_model._make_predict_function() after loading the model helps to avoid the error.

@st.cache(allow_output_mutation=True) def load_model(): json_file = open('Model.json', 'r') loaded_model_json = json_file.read() json_file.close() loaded_model = model_from_json(loaded_model_json) loaded_model.load_weights("weights/epoch200.hdf5") loaded_model._make_predict_function() *** session = K.get_session() return loaded_model, session

[mohammedayub44]

@carolmanderson works like a charm. Thanks a lot. I was trying to reset the session for some reason and that did not work. Just passing the current TF session solved it.