onnx转tflite后，推理速度变慢

[Yakuho]

我先使用onnx转成tensorflow pb，代码如下

from onnx_tf.backend import prepare
import onnx

onnx_model = onnx.load("version-RFB-320.onnx")
tf_rep = prepare(onnx_model)
tf_rep.export_graph("version-RFB-320-tensorflow")

然后我再将pb转为tflite

import tensorflow as tf

saved_model_dir = 'version-RFB-320-tensorflow'      # version-RFB-320-tensorflow 路径下含有pb和variables
save_tf_model = "version-RFB-320.tflite"

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
tflite_model = converter.convert()
open(save_tf_model, "wb").write(tflite_model)

但是我在本机上尝试使用version-RFB-320.tflite的时候，能够正常预测，进行nms选出人脸框，效果很不错。但是我测试了一下推理时间，花费了180ms左右！！！
本机配置 (CPU i5 8th, GPU GTX1050)

请作者们给我一些指导~~感谢

[smilemakc]

look at the graph of the model after prepare(onnx_model), and you will most likely be shocked by what you see

[7aughing]

look at the graph of the model after prepare(onnx_model), and you will most likely be shocked by what you see

Truly, add too many transpose node in tflite

[wei8171023]

look at the graph of the model after prepare(onnx_model), and you will most likely be shocked by what you see

Truly, add too many transpose node in tflite

add many others node in tflite!!