Train Interface

Note: Args that start with ‘--’ (eg. --save-config) can also be set in a config file (specified via --config). The config file uses YAML syntax and must represent a YAML ‘mapping’ (for details, see http://learn.getgrav.org/advanced/yaml). If an arg is specified in more than one place, then command line values override config file values which override defaults.

For generic options see: here

Model specific options can be seen here:

• NuQE training
• Predictor training
• Predictor-Estimator training
• QUETCH training
• Linear training

Usage examples of the python api.

Run:

import kiwi

train_predest_config = 'experiments/examples/train_predest.yaml'
kiwi.train(train_predest_config)

train_nuqe_config = 'experiments/examples/train_nuqe.yaml'
kiwi.train(train_nuqe_config)

Usage example of the cli api.

Run:

kiwi train --config {model_config_file} [OPTS]

usage: kiwi train [-h] [--epochs EPOCHS] [--train-batch-size TRAIN_BATCH_SIZE]
                  [--valid-batch-size VALID_BATCH_SIZE]
                  [--optimizer {sgd,adagrad,adadelta,adam,sparseadam}]
                  [--learning-rate LEARNING_RATE]
                  [--learning-rate-decay LEARNING_RATE_DECAY]
                  [--learning-rate-decay-start LEARNING_RATE_DECAY_START]
                  [--checkpoint-validation-steps CHECKPOINT_VALIDATION_STEPS]
                  [--checkpoint-save [CHECKPOINT_SAVE]]
                  [--checkpoint-keep-only-best CHECKPOINT_KEEP_ONLY_BEST]
                  [--checkpoint-early-stop-patience CHECKPOINT_EARLY_STOP_PATIENCE]
                  [--resume [RESUME]]

training

--epochs	Number of epochs for training. Default: 50
--train-batch-size
	Maximum batch size for training. Default: 64
--valid-batch-size
	Maximum batch size for evaluating. Default: 64

training-optimization

--optimizer	Possible choices: sgd, adagrad, adadelta, adam, sparseadam Optimization method. Default: “adam”
--learning-rate
	Starting learning rate. Recommended settings: sgd = 1, adagrad = 0.1, adadelta = 1, adam = 0.001 Default: 1.0
--learning-rate-decay
	Decay learning rate by this factor. Default: 1.0
--learning-rate-decay-start
	Start decay after this epoch. Default: 0

training-save-load

--checkpoint-validation-steps
	Perform validation every X training batches. Saves model if checkpoint-save is true. Default: 0
--checkpoint-save
	Save a training snapshot when validation is run. If false it will never save the model. Default: True
--checkpoint-keep-only-best
	Keep only n best models according to main metric (F1Mult by default); 0 will keep all. Default: 1
--checkpoint-early-stop-patience
	Stop training if evaluation metrics do not improve after X validations; 0 disables this. Default: 0
--resume	Resume training a previous run. If –output-dir is not none, Kiwi will load from a checkpoint folder in that location. If –output-dir is not specified, then the –run-uuid option must be set. Files are then searched under the “runs” directory. If not found, they are downloaded from the MLflow server (check the –mlflow-tracking-uri option). Default: False