Gist showing a weird caveat in model saving and model loading in keras. Uncommenting lines 25-26 causes an error.

keras_round_trip_model_saving.py

from keras.layers import Input, Dense, merge, Flatten
from keras.layers.convolutional import Convolution2D
from keras.models import Model, model_from_json
from keras import backend as K

def make_model(img_edge_size):

    img_shape = (3, img_edge_size, img_edge_size)
    mask_shape = (1, img_edge_size, img_edge_size)

    generic_img = Input(shape=img_shape)
    layer = Convolution2D(20, 3, 3, activation='relu', border_mode='same')(generic_img)
    reusable_img_featurizer = Model(generic_img, layer)

    start_img = Input(shape=img_shape, name='start_img')
    start_img_features = reusable_img_featurizer(start_img)
    end_img = Input(shape=img_shape, name='end_img')
    end_img_features = reusable_img_featurizer(end_img)


    start_box_mask = Input(shape=mask_shape, name='start_box_mask')

    start_img_features = merge([start_img_features, start_box_mask],
                                mode='concat', concat_axis=1)
    #start_img_features = Convolution2D(40, 3, 3, activation='relu', border_mode='same',
    #                                   dim_ordering='th')(start_img_features)

    start_img_features = Flatten()(start_img_features)
    end_img_features = Flatten()(end_img_features)
    layer = merge([start_img_features, end_img_features],
                   mode='concat',
                   concat_axis=1)

    x0 = Dense(1, activation='linear', name='x0')(layer)

    my_model = Model(input=[start_img, end_img, start_box_mask],
                    output=[x0])
    return my_model


def load_model(model_spec, weights_fname):
    print('Loading saved model')
    my_model = model_from_json(open(model_spec, 'r').read())
    my_model.load_weights(weights_fname)
    return my_model


def save_model(my_model, model_spec, weights_fname):
    print('Saving model')
    json_string = my_model.to_json()
    open(model_spec, 'w').write(json_string)
    my_model.save_weights(weights_fname, overwrite=True)


if __name__ == "__main__":
    img_edge_size = 100
    weights_fname = 'model_weights.h5'
    model_spec = 'model_architecture.json'

    my_model = make_model(img_edge_size)
    save_model(my_model, model_spec, weights_fname)
    load_model(model_spec, weights_fname)