#正確步驟1
import tensorflow as tf
def freeze_graph(input_checkpoint, output_graph):
output_node_names = "strided_slice_13,strided_slice_23" #獲取的節點
saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=True)
graph = tf.get_default_graph() # 獲得默認的圖
input_graph_def = graph.as_graph_def() # 返回一個序列化的圖代表當前的圖
with tf.Session() as sess:
saver.restore(sess, input_checkpoint) # 恢復圖並得到數據
output_graph_def = tf.graph_util.convert_variables_to_constants( # 模型持久化,將變量值固定
sess=sess,
input_graph_def=input_graph_def, # 等於:sess.graph_def
output_node_names=output_node_names.split(",")) # 如果有多個輸出節點,以逗號隔開
with tf.gfile.GFile(output_graph, "w") as f: # 保存模型
f.write(output_graph_def.SerializeToString()) # 序列化輸出
# f.write(output_graph_def.SerializeToOstream()) # 序列化輸出
# print("%d ops in the final graph." % len(output_graph_def.node)) # 得到當前圖有幾個操作節點
if __name__ == '__main__':
modelpath="./model"
freeze_graph(modelpath,"frozen.pbtxt")
print("finish!")
#正確步驟2
import tensorflow as tf
convert=tf.lite.TFLiteConverter.from_frozen_graph("frozen.pbtxt",input_arrays=["waveform"],output_arrays=["strided_slice_23"])
convert.post_training_quantize=False
tflite_model=convert.convert()
open("model.tflite","w").write(tflite_model)
#當需要給定輸入數據形式時,給出輸入格式:
# import tensorflow as tf
# convert=tf.lite.TFLiteConverter.from_frozen_graph("frozen.pbtxt",input_arrays=["waveform"],output_arrays=["strided_slice_23"],
# input_shapes={"waveform":[1,2]})
# convert.post_training_quantize=True
# tflite_model=convert.convert()
# open("model.tflite","w").write(tflite_model)
print("finish!")
如果轉換失敗,看失敗日誌:
Converting unsupported operation: RFFT 說明有不支持的函數、方法