ImageNet Weights being reset - only able to achieve 53% accuracy, whereas 69.5% was achieved in the paper

[s36srini]

Here is how I'm creating the model:

# Initialization and Compilation
mobileNet = keras.applications.mobilenet.MobileNet(weights='imagenet') # ImageNet 2012 trained weights

pruning_params = {'pruning_schedule': sparsity.PolynomialDecay(
                        initial_sparsity=0.0, final_sparsity=0.5,
                        begin_step=0, end_step=40000, frequency=200)}

model_for_pruning = sparsity.prune_low_magnitude(mobileNet, **pruning_params)

mobileNet.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

When using the backend context manager below, do the imagenet weights get reset?

    iterator = dataset.make_initializable_iterator()
    next_batch = iterator.get_next()
    with keras.backend.get_session().as_default() as sess:
        sess.run(iterator.initializer, feed_dict={filenames: files})
        while True:
            inputs, labels = sess.run(next_batch)
            yield inputs, labels`

Here is my call for training:

# Train the MobileNet model with its imagenet weights (faster convergence) and have it prune
mobileNet.fit_generator(
    generator=input_fn('train'),
    validation_data=input_fn('validation'),
    steps_per_epoch=1024,
    validation_steps=128,
    workers=0,
    verbose=1,
    epochs=1,
    callbacks=[checkpoint, tensorboard, update_step]
)