Training Tokenizers
medtokenizers ships a Trainer with a callback
system for training tokenizers from scratch.
python scripts/train.py --help
The trainer supports continuous (reconstruction + KL) and discrete
(reconstruction + quantization/commitment + optional entropy) objectives, mixed
precision, and checkpointing. Callbacks such as
EarlyStopping,
Checkpoint,
LRScheduler, and
Logger hook into the training loop.
See Training for the full API.