import tensorflow as tf
from tensorflow.keras.models import load_model
import numpy as np
if __name__ == '__main__':
norm_size=192
graph = tf.Graph()
with graph.as_default():
session = tf.Session()
with tf.Session().as_default():
input = tf.placeholder(tf.float32, [None, 192, 192, 3], name='inputs')
model = load_model('../model/192-v69.m', compile=False)
output = tf.identity(model(input), name='outputs')
ops = [tf.assign(v, w) for v, w
in zip(tf.global_variables(), model.get_weights())]
session.run(ops)
frozen = tf.graph_util.convert_variables_to_constants(
session, graph.as_graph_def(), ['outputs'])
# input: 'conv2d_1_input:0'
with open('../model/192-v69.pb', 'wb') as f:
f.write(frozen.SerializeToString())
graph = tf.Graph()
with graph.as_default():
with open('../model/192-v69.pb', 'rb') as f:
pb = tf.GraphDef.FromString(f.read())
tf.import_graph_def(pb, name='')
session = tf.Session()
func1 = session.make_callable('outputs:0', ['inputs:0'])
kg = tf.Graph()
with kg.as_default():
session2 = tf.Session()
with session2.as_default():
keras = load_model('../model/192-v69.m', compile=False)
# keras.predict(np.zeros([1,192,192,3]))
func2 = keras.predict
testdata = np.random.uniform(0., 1., [1,192,192,3]).astype(np.float32)
testoutput1 = func1(testdata)
with kg.as_default():
with session2.as_default():
testoutput2 = func2(testdata)
print(testoutput1)
print(testoutput2)
assert np.allclose(testoutput1, testoutput2)