Exporting Trained Models and Checkpoints

After training a model that achieves satisfactory performance, it is often important to export the trained model for use outside PEDL (e.g., to deploy the model as part of a web service or on an embedded device). In the current version of PEDL, exporting models is supported by copying the checkpoints that are automatically created during the training process.

A checkpoint consists of the model architecture and the values of the model's parameters (i.e., weights) and hyperparameters. When using a stateful optimizer during training, checkpoints will also consist of optimizer state (i.e., learning rate). The exact checkpoint format depends on the application framework being used:

  • TensorFlow trials are checkpointed using the SavedModel format. Please consult the TensorFlow documentation to restore models from the SavedModel format.
  • Keras trials are checkpointed as an HDF5 file with the weights of a model and a JSON file describing the model's architecture. When using a stateful optimization method, the optimizer state is also saved in a separate HDF5 file. To restore the architecture and/or weights of a Keras trial checkpoint, use the keras.models.model_from_json() and model.load_weights() Keras APIs, respectively. Optimizer state is currently saved in a custom experimental format—to restore it outside of PEDL, please contact the Determined AI team for a consultation.
  • TF Keras trials are checkpointed to a file named pedl-keras-model using tf.keras.models.save_model. You can learn more from the TF Keras docs.
  • PyTorch trials are checkpointed as a file. This file is created very similar to what is described in the PyTorch documentation. Instead of the fields in the documentation linked above, there are three fields for the dictionary that is saved by PEDL: "model_state_dict", "optimizer_state_dict", and "hparams", which are the model's state_dict, optimizer's state_dict, and trial's hyperparameters respectively.

PEDL will automatically create checkpoints for all the trials in an experiment. Checkpoints are created in three situations:

  1. When PEDL suspends training of a trial at one agent, before later resuming training that trial at a different agent.
  2. If min_checkpoint_period is set in the experiment configuration, each trial will be checkpointed whenever the specified number of training steps are completed since the last checkpoint.
  3. When a trial has been trained to completion (according to the configuration of the experiment's trial search method).

The PEDL CLI can be used to view all the checkpoints associated with an experiment:

$ pedl experiment list-checkpoints <experiment-id>

Checkpoints are saved to external storage, as specified in the experiment configuration; see the discussion of checkpoint_storage above. Each checkpoint has a UUID, which is used as the name of the checkpoint directory on the external storage system. For example, if the experiment is configured to save checkpoints to a shared file system:

  type: shared_fs
  host_path: /mnt/nfs-volume-1

A checkpoint with UUID b3ed462c-a6c9-41e9-9202-5cb8ff00e109 can be found in the directory /mnt/nfs-volume-1/b3ed462c-a6c9-41e9-9202-5cb8ff00e109.

The PEDL CLI can be used to download a checkpoint saved on S3 or GCS:

$ pedl checkpoint download <trial-id> <step-id>

Checkpoint Garbage Collection

Typically only some checkpoints are appropriate for deployment. Once an experiment has finished, PEDL can optionally garbage collect some or all of the checkpoints taken when the experiment was running. The following parameters in the checkpoint_storage section of the experiment configuration specify which checkpoints to save:

  • save_experiment_best: The number of the best checkpoints with validations over all trials to save (where best is measured by the validation metric specified in the searcher configuration).
  • save_trial_best: The number of the best checkpoints with validations of each trial to save.
  • save_trial_latest: The number of the latest checkpoints of each trial to save.

If multiple save_* parameters are specified, the union of the specified checkpoints are saved.

Default GC Policy

Any GC policy parameter that isn't specified will default to the following respective value:

    save_experiment_best: 0,
    save_trial_best: 1,
    save_trial_latest: 1

This policy will save the most recent and the best checkpoint per trial. In other words, if the most recent checkpoint is also the best checkpoint for a given trial, only one checkpoint will be saved for that trial. Otherwise, two checkpoints will be saved for that trial.


Suppose an experiment has the following trials, checkpoints and validation metrics (where smaller_is_better is true):

Trial ID Checkpoint ID Validation Metric
1 1 null
1 2 null
1 3 0.6
1 4 0.5
1 5 0.4
2 6 null
2 7 0.2
2 8 0.3
2 9 null
2 10 null

The effect of various policies is enumerated in the following table:

save_experiment_best save_trial_best save_trial_latest Saved Checkpoint IDs
0 0 0 none
2 0 0 8,7
>= 5 0 0 8,7,5,4,3
0 1 0 7,5
0 >= 3 0 8,7,5,4,3
0 0 1 10,5
0 0 3 10,9,8,5,4,3
2 1 0 8,7,5
2 0 1 10,8,7,5
0 1 1 10,7,5
2 1 1 10,8,7,5

If aggressive reclamation is desired, set save_experiment_best to a 1 or 2 and leave the other parameters zero. For more conservative reclamation, set save_trial_best to 1 or 2; optionally set save_trial_latest as well.

Checkpoints of an existing experiment can be garbage collected by changing the GC policy using the pedl experiment set gc-policy subcommand of the PEDL CLI.

Checkpoint Storage Configuration

A default checkpoint storage will be used when each experiment is created. It is recommended to create a default checkpoint configuration that will be used by all pertinent tasks started by the cluster. To configure this value, edit the checkpoint_storage configuration in cluster configuration with the desired type and location of checkpoint storage. See Cluster Configuration for details. However, every experiment can still be configured with a specific checkpoint storage. See Experiment Configuration for details.