如何在没有源代码的情况下修改预训练的张量流模型？

Question

I would like to modify a pre-trained model(to be specific, the ssdlit-mobilenetv2 in the tensorflow detection zoo). What I want to do is to extract the ouput of one layer, do some modification to the output, then feed it back and continue to run the model. I only got the .pb file and the ckpt files, how can I do this without the source code?

So what code should I use?

Answer 1

I think, you have no easy way of doing this.. Just write your own tool for your concrete modification based on existing data, than saving it again as checkpoint and continue train process from that point.

If file has the information of values, so you can change it manually, but i don't think it could be effective.

Also, as you know the name (mobilentv2) you can find the net architecture and the source codes