Problem to convert a pre-trained keras h5 model into a tf.savedmodel.save model, just conversion is okay, no further trainning is needed,

General Discussion
, ,
#1

I had my pre-trained model implemented in Keras which is okay for our application. And since I am trying to optimize its computational performance, I have to convert it into a tf savedmodel model. I have converted the same architecture of it into tf.keras and loaded all variables from the pre-trained h5 file. but when I applied tf.savedmodel.save on this loaded model, it didn’t save the model’s weights into the output asset file (just empty random weights). The pb file I got is only ~2 Mb.
It would be very appreciated if someone knows how to write loaded weights into asset output.

#3

Just to be clear, do you have the Keras model loaded in memory?

if so, you can just use something like model.save() (more details here: 保存和加载 Keras 模型  |  TensorFlow Core)
it will save in the saved_model format