You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Note that the TF-Hub checkpoints do not contain LAMB optimizer momentum estimates, so it won't be exactly the same as resuming from a training checkpoint (which we do plan to release at some point). In particular, the momentum estimates will be reset to zero.
That said, it's still probably faster than starting from a randomly-initialized ALBERT :)
If you want to try it, I can think of two ways:
Extract the checkpoint files manually from the TF-Hub module. Look inside the variables/ folder. These files can be copied into your training dir as the .data and .index files of a checkpoint. You may need to create a file called checkpoint that points to them (look inside an ALBERT training directory for the format).
Add a flag to run_pretraining.py for initializing from TF-Hub. It should be similar to this recent commit, which did the same thing but for run_classifier.py. If you want to put together a PR for this, that'd really be helpful.
Training from scratch is very expensive. Anybody know how to continue train ALBERT from the exported model..
Thanks
The text was updated successfully, but these errors were encountered: