TensorFlow 将 checkpoint 冻结为 frozen_graph

import tensorflow as tf
from tensorflow.python.tools.freeze_graph import freeze_graph


def freeze(ckpt_dir, meta_file_path, output_node_names):
	# 导入结构、加载权重
    saver = tf.train.import_meta_graph(meta_file_path)
    with tf.Session() as sess:
        saver.restore(sess, tf.train.latest_checkpoint(ckpt_dir))

        tf.train.write_graph(sess.graph_def, 'temp/', 'temp.pb')  # temp/、temp.pb
        freeze_graph(input_graph='temp/temp.pb',  # temp/temp.pb

                     input_checkpoint=tf.train.latest_checkpoint(ckpt_dir),
                     output_graph='frozen_graph.pb',
                     output_node_names=output_node_names,
					 
					 # 以下为固定写法
                     clear_devices=True,
                     input_binary=False,
                     input_saver='',
                     restore_op_name='save/restore_all',
                     filename_tensor_name='save/Const:0',
                     initializer_nodes='')


if __name__ == '__main__':
    freeze('./ckpt', './ckpt/model.meta', 'input,classifier/Softmax,regression/BiasAdd')