Migrating TensorFlow from SaaS to your server. Really?

Question

Neural networks

Good afternoon, there was a task to quickly mold an application prototype with an object recognition function on it. I don’t have much experience with neural networks, so I decided to start everything with the help of a third-party SaaS for now, and later migrate to my own server.
The essence of the question is whether it will be enough for migration to have a pack of pictures, on the basis of which the network was taught, and a pack of xml files with annotations to achieve transparent migration from the client's point of view? In other words, is such migration real? Will the output of the network, after migration, be identical to what the third-party SaaS gives out, or will everything break down and it’s better not to do this? This SaaS is based on TensorFlow. This is the same TensorFlow, I think I will raise it on my server later.
Pros, Cons, Pros, Cons, Pitfalls. I would appreciate any experienced opinion. Is this a good idea or not?

PS If you think that this is real, then please advise, if possible, a tool for manually creating annotations to pictures for pascal voc, it is desirable that it can be run on Windows. These annotations will come in handy later.

Thank you very much in advance!