Train neural network sequentially

Research & Models
#1

Hi there,
I want to train the neural network based on a partial part of my data and then re-train iteratively the neural network until reaches the entire data domain, rather than training the neural network given the entire points at once. I am using TF version 1.15. I wonder if someone can help how to do it?

#3

Hi,

Some things here:
1- Why not use TF 2.x?
2- Where will the iteratively part be run? which environment? (on-device or on servers?)

if you’re thinking of just continuing the training, that doable if you have the checkpoints from your regular training

if it’s a different env than it might be a little bit more challenging and require more work. That depends on what you are trying to do

#4

Also try to take a look to this thread:

https://tensorflow-prod.ospodiscourse.com/t/continuous-adaptation-for-machine-learning-system-to-data-changes/6273?u=bhack

1 Like
#6

Hi bhack,

I used TF 1.15 since my code is written there and it’s easier to work on it rather than adapt to the new version. Also it’s on-device.

Second, I have data points for the entire space but instead of throwing them all instantly to the neural network, i want to give it in a sequential manner. check out the graph (link below), Meaning train for the first part, then go to the second time step and so on and so forth (attached photo). Let me know if its not clear.

for now im thinking of looping over each time interval and call (self.optimizer.minimize…) but im not sure if its a correct way.