用mmdnn实现模型转换
参考链接:https://www.twblogs.net/a/5ca4cadbbd9eee5b1a0713af
- 安装mmdnn
pip install mmdnn
- 准备好mxnet模型的.json文件和.params文件, 以InsightFace MxNet r50为例 https://github.com/deepinsight/insightface
- 用mmdnn运行命令行
python -m mmdnn.conversion._script.convertToIR -f mxnet -n model-symbol.json -w model-0000.params -d resnet50 --inputShape 3,112,112
会生成resnet50.json(可视化文件) resnet50.npy(权重参数) resnet50.pb(网络结构)三个文件。
- 用mmdnn运行命令行
python -m mmdnn.conversion._script.IRToCode -f tensorflow --IRModelPath resnet50.pb --IRWeightPath resnet50.npy --dstModelPath tf_resnet50.py
生成tf_resnet50.py文件,可以调用tf_resnet50.py中的KitModel函数加载npy权重参数重新生成原网络框架。
-
打开tf_resnet.py文件,修改load_weights()中的代码 (tensorflow=1.14.0报错)
try: weights_dict = np.load(weight_file).item() except: weights_dict = np.load(weight_file, encoding='bytes').item()
改为
try: weights_dict = np.load(weight_file, allow_pickle=True).item() except: weights_dict = np.load(weight_file, allow_pickle=True, encoding='bytes').item()
-
基于resnet50.npy和tf_resnet50.py文件,固化参数,生成PB文件:
import tensorflow as tf import tf_resnet50 as tf_fun def netWork(): model=tf_fun.KitModel("./resnet50.npy") return model def freeze_graph(output_graph): output_node_names = "output" data,fc1=netWork() fc1=tf.identity(fc1,name="output") 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 init = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init) 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, "wb") as f: # 保存模型 f.write(output_graph_def.SerializeToString()) # 序列化輸出 if __name__ == '__main__': freeze_graph("frozen_insightface_r50.pb") print("finish!")
- 采用tensorflow的post-train quantization离线量化方法(有一定的精度损失)转换成tflite模型,从而完成端侧的模型部署:
import tensorflow as tf convert=tf.lite.TFLiteConverter.from_frozen_graph("frozen_insightface_r50.pb",input_arrays=["data"],output_arrays=["output"], input_shapes={"data":[1,112,112,3]}) convert.post_training_quantize=True tflite_model=convert.convert() open("quantized_insightface_r50.tflite","wb").write(tflite_model) print("finish!")