gaphex

json.dump(model.to_json(), open("model.json", "w"))

model = tf.keras.models.model_from_json(json.load(open("model.json")), 
                                        custom_objects={"BertLayer": BertLayer})
model.load_weights("bert_tuned.hdf5")

gaphex

from tensorflow.python.framework.graph_util import convert_variables_to_constants
from tensorflow.python.tools.optimize_for_inference_lib import optimize_for_inference

def freeze_keras_model(model, export_path=None, clear_devices=True):
    """
    Freezes the state of a session into a pruned computation graph.

    @param model The Keras model to be optimized for inference.
    @param clear_devices Remove the device directives from the graph for better portability.
    @return The frozen graph definition.

gaphex

frozen_graph = freeze_keras_model(model, export_path="frozen_graph.pb")

gaphex

def load_graph(frozen_graph_filename):
    with tf.gfile.GFile(frozen_graph_filename, "rb") as f:
        graph_def = tf.GraphDef()
        graph_def.ParseFromString(f.read())

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(graph_def)
    return graph
    
restored_graph = load_graph("frozen_graph.pb")

gaphex

graph_ops = restored_graph.get_operations()
input_op, output_op = graph_ops[0].name, graph_ops[-1].name

x = restored_graph.get_tensor_by_name(input_op + ':0')
y = restored_graph.get_tensor_by_name(output_op + ':0')

gaphex

preprocessor = build_preprocessor("./uncased_L-12_H-768_A-12/vocab.txt", 64)
py_func = tf.numpy_function(preprocessor, [x], [tf.int32, tf.int32, tf.int32], name='preprocessor')

gaphex

sess = tf.Session(graph=restored_graph)

t_data = ["What would happen if earth stopped rotating? ||| What would happen if the Earth rotation suddenly stop?"]

y_out = sess.run(y, feed_dict={
        x: np.array(t_data).reshape((-1,1))
    })

y_out

gaphex

inp = tf.keras.Input(shape=(1,), dtype=tf.string)
encoder = BertLayer(bert_path="./bert-module/", seq_len=48, 
                    tune_embeddings=False, do_preprocessing=True,
                    pooling='cls', n_tune_layers=3, verbose=False)

pred = tf.keras.layers.Dense(1, activation='sigmoid')(encoder(inp))
model = tf.keras.models.Model(inputs=[inp], outputs=[pred])

model.compile(optimizer="adam", loss="binary_crossentropy", metrics=["accuracy"])
model.fit(trX, trY, validation_data=[tsX, tsY], batch_size=128, epochs=5)

gaphex

-0.033	-0.578	0.1926	0.2186	-0.5674	0.2001	0.4424	-0.2903	0.1769	0.355	0.2542	-1.124	0.1263	-0.5547	0.157	-0.1073	-0.02002	-0.1493	0.02702	0.1525	0.1299	0.3364	0.276	-0.004074	-0.3472	0.4946	-0.1622	0.3628	0.03714	0.3384	0.3384	-0.4062	-0.392	-0.3782	0.2084	0.09204	-0.01868	0.1995	-0.03073	0.4304	0.087	-0.04553	-0.6196	0.05472	0.337	0.4192	0.4705	-0.2595	-0.2952	-1.256	-0.10583	0.4216	-0.03677	0.04718	-0.414	0.651	-0.1908	0.2788	-0.125	0.1112	-0.3457	-0.4543	0.256	0.444	0.372	0.007282	0.254	-0.0879	-0.9517	0.1906	0.0641	0.3645	0.09	-0.4607	-0.4106	-0.1879	0.4492	0.12067	-0.1442	0.3267	0.1013	-0.4993	-0.2394	0.814	-0.2183	0.3274	-0.2252	-0.02669	0.2515	-0.457	-0.0087	-0.6934	0.312	0.1274	-0.3042	0.05402	-0.06024	-0.3518	0.5913	0.063	-0.4602	-0.6064	-0.04974	-0.01523	0.4531	-0.07745	-0.01627	0.2477	0.2498	-0.02473	0.10626	0.164	0.3872	-0.4756	0.2184	-0.7964	-0.3525	0.4448	0.02672	-0.432	0.1124	0.3257	0.3655	-0.203	-0.2278	0.01701	0.1403	0.235	0.0183	-0.1366	-0.5996	0.1203	0.3403	0.04422	-0.16	-0.657	0.362	-0.18

gaphex