[Feature] Reduce overfitting when training

[adamltyson]

Users are generating small amounts of data, and the network seems to be over fitting. They're having to manually get existing data to constrain their models. This could be improved by:

• Providing old training data via GIN
• Allowing upper layers to be frozen

Interested to know what aspects of training have helped @FedeClaudi @larsrollik

[FedeClaudi]

I'm trying training with my data + a large set of training examples from the initial cellfinder network, will let you know if it helps.

Meanwhile I will also try to create a larger and more balanced training set for my data, in my case, let's see if it helps.