Resume Training

Resume training only works with Train checkpoints. PyTRIO loads the saved model weights and optimizer state, then continues training from that point.

If you have not saved a Train checkpoint yet, use save_state first. See Save Weights.

Get the Checkpoint Path

Find the train-checkpoint path in the Web UI:

Restore Model and Optimizer State

Pass that path to create_training_client_from_state_with_optimizer:

import pytrio as trio
 
service_client = trio.ServiceClient()

training_client = service_client.create_training_client_from_state_with_optimizer(
    path="YOUR_TRAIN_CHECKPOINT_PATH",
)

After this, continue with forward_backward and optim_step; training resumes from the checkpoint state.

Load Weights Only

If you want to load saved weights without optimizer state, use create_training_client_from_state:

import pytrio as trio
 
service_client = trio.ServiceClient()

training_client = service_client.create_training_client_from_state(
    path="YOUR_ADAPTER_PATH",
)

This creates a new training client initialized from existing weights. Because optimizer state is not restored, it is not strict checkpoint resume training.