Save and Load models

Work with a TF Checkpoint

The ModelSaver callback saves the model to the directory defined by logger.get_logger_dir(), in TensorFlow checkpoint format. A TF checkpoint typically includes a .data-xxxxx file and a .index file. Both are necessary.

tf.train.NewCheckpointReader is the offical tool to parse TensorFlow checkpoint. Read TF docs for details. Tensorpack also provides a small tool to load checkpoints, see load_chkpt_vars for details.

scripts/ls-checkpoint.py demos how to print all variables and their shapes in a checkpoint.

Tensorpack includes another tool to save variables to TF checkpoint, see save_chkpt_vars.

Work with .npz Files in the Model Zoo

Most models provided by tensorpack are in npz (dictionary) format, because it’s easy to use without TF dependency. You can read/write them with np.load and np.savez.

scripts/dump-model-params.py can be used to remove unnecessary variables in a checkpoint and save results to a npz. It takes a metagraph file (which is also saved by ModelSaver) and only saves variables that the model needs at inference time. It dumps the model to a var-name: value dict saved in npz format.

Load a Model to a Session

Model loading (in both training and inference) is through the session_init interface. For training, use session_init in TrainConfig(...) or Trainer.train(...). For inference, use session_init in PredictConfig(...).

There are a few ways a session can be initialized:

session_init=SmartInit("path/to/checkpoint")  # load a TF checkpoint
session_init=SmartInit("path/to/model_zoo.npz")  # load tensorpack model zoo
session_init=SmartInit(dict_of_parameters)  # load a dictionary
session_init=SmartInit(["path1", dict2])  # load them sequentially

SmartInit is in fact a small helper which uses some heuristics to return you one of SaverRestore or DictRestore. They are responsible for the actual initialization work.

Whatever you use in session_init, this is what happens during the loading:

  • Variable restoring is completely based on exact name match between variables in the current graph and variables in the session_init initializer.

  • Variables that appear in only one side will be printed as warning.

  • Variables of the same name but incompatible shapes will cause exceptions. If you set ignore_mismatch=True, then such errors will only be printed as warnings.

Transfer Learning

Therefore, transfer learning is trivial.

If you want to load a pre-trained model, just use the same variable names. If you want to re-train some layer, either rename the variables in the graph, or rename/remove the variables in your loader.

Resume Training

“Resume training” is mostly just “loading the last known checkpoint”. To load a model, you should refer to the previous section: Load a Model to a Session.

Note

A checkpoint does not resume everything!

Loading the checkpoint does most of the work in “resume trainig”, but note that TensorFlow checkpoint only saves TensorFlow variables, which means other Python state that are not TensorFlow variables will not be saved and resumed. This means:

  1. Training epoch number will not be resumed. You can set it by providing a starting_epoch to your TrainConfig.

  2. State in your callbacks will not be resumed. Certain callbacks maintain a state (e.g., current best accuracy) in Python, which cannot be saved automatically.

The AutoResumeTrainConfig is an alternative of TrainConfig which applies some heuristics to load the lastest epoch number and lastest checkpoint.