Converting Keras model as tensorflow model gives me error

Question

Hi I am trying to save my 'saved models' (h5 files) as tensorflow file.This is the code I used.

import tensorflow as tf
def tensor_function(i):
    tf.keras.backend.set_learning_phase(0)  # Ignore dropout at inference
    model = tf.keras.models.load_model('/home/ram/Downloads/AutoEncoderModels_ch2/19_hour/autoencoder_models_ram/auto_encoder_model_pos_' + str(i) + '.h5')
    export_path = '/home/ram/Desktop/tensor/' + str(i)
    #sess = tf.Session()

    # Fetch the Keras session and save the model
    # The signature definition is defined by the input and output tensors
    # And stored with the default serving key
    with tf.keras.backend.get_session() as sess:
        tf.saved_model.simple_save(
            sess,
            export_path,
            inputs={'input_image': model.input},
            outputs={t.name: t for t in model.outputs})
        sess.close()

for i in range(4954):
    tensor_function(i)

I tried to open the session manually by using sess = tf.session() (removed with as well) as well but in vain

And the above error I got when I used jupyter notebook and when I ran the same in linux terminal.I get the following error

tensorflow.python.framework.errors_impl.FailedPreconditionError: Error while reading resource variable dense_73/bias from Container: localhost. This could mean that the variable was uninitialized. Not found: Container localhost does not exist. (Could not find resource: localhost/dense_73/bias)
     [[{{node dense_73/bias/Read/ReadVariableOp}} = ReadVariableOp[_class=["loc:@dense_73/bias"], dtype=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"](dense_73/bias)]]

And when I tried to save just the one 'saved model file' it ran successfully.Problems happen only when I try to run it in a loop(probably some session problem).

I tried this answer in SO but didnt help much.

Answer 1

For me the following two options work:

Option 1: Add tf.keras.backend.clear_session() at the beginning of your tensor_function and use a 'with' block:

def tensor_function(i):
    tf.keras.backend.clear_session()
    tf.keras.backend.set_learning_phase(0)  # Ignore dropout at inference

    model = ...

    export_path = 'so-test/' + str(i)

    with tf.keras.backend.get_session() as sess:
        tf.saved_model.simple_save(
            sess,
            export_path,
            inputs={'input_image': model.input},
            outputs={t.name: t for t in model.outputs})
        sess.close()

Option 2: Use tf.Session() instead of the 'with' block but add the line sess.run(tf.global_variables_initializer()) :

def tensor_function(i):
    tf.keras.backend.set_learning_phase(0)  # Ignore dropout at inference

    model = ...

    export_path = 'so-test/' + str(i)
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    tf.saved_model.simple_save(
        sess,
        export_path,
        inputs={'input_image': model.input},
        outputs={t.name: t for t in model.outputs})
    sess.close()