Command Line Options

Command line options may be specified in either a config file or directly passed as flags. If an option exists in both the config file and flags, flags take precedent and override the config file.

The config file is a yaml file, examples may be found in example_config. In the config file, each key-value pair key: value corresponds to passing the flag --key value. The following example sets the training data path in the config file

training_file: path/to/data

For an option with multiple values, it can be specified as a list

monitor_metrics: [P@1, P@3, P@5]

The corresponding command line flag is passed as

python3 main.py --monitor_metrics P@1 P@3 P@5

For options such as liblinear_options, the value may contain -. Use = to pass the value on the command line

python3 main.py --liblinear_options='-s 2 -B 1 -e 0.0001 -q'

Second-level parameters such as those under``network_config`` must be specified in a config file.

List of Options

Name	Description
-c --config	Path to configuration file.
--result_dir	The directory to save checkpoints and logs (default: ./runs).
--data_name	Dataset name (default: unnamed_data).
--training_file	Path to training data (default: None).
--val_file	Path to validation data (default: None).
--test_file	Path to test data (default: None.
--val_size	Training-validation split: a ratio in [0, 1] or an integer for the size of the validation set (default: 0.2)..
--min_vocab_freq	The minimum frequency needed to include a token in the vocabulary (default: 1).
--max_seq_length	The maximum number of tokens of a sample (default: 500).
--shuffle	Whether to shuffle training data before each epoch (default: True).
--merge_train_val	Whether to merge the training and validation data. (default: None).
--include_test_labels	Whether to include labels in the test dataset. (default: None).
--remove_no_label_data	Whether to remove training and validation instances that have no labels. (default: None).
--add_special_tokens	Whether to add the special tokens for inputs of the transformer-based language model. (default: True).
--seed	Random seed (default: None).
--epochs	The number of epochs to train (default: 10000).
--batch_size	Size of training batches (default: 16).
--optimizer	Optimizer (default: adam). One of {adam, adamw, adamax, sgd}.
--learning_rate	Learning rate for optimizer (default: 0.0001).
--weight_decay	Weight decay factor (default: 0).
--momentum	Momentum factor for SGD only (default: 0.9).
--lr_scheduler	Name of the learning rate scheduler (default: None).
--patience	The number of epochs to wait for improvement before early stopping (default: 5).
--early_stopping_metric	The metric to monitor for early stopping. Set to val_metric if specified as None. (default: None).
--normalize_embed	Whether the embeddings of each word is normalized to a unit vector (default: None).
--model_name	Model to be used (default: unnamed_model).
--init_weight	Weight initialization to be used (default: kaiming_uniform).
--loss_function	Loss function (default: binary_cross_entropy_with_logits).
--eval_batch_size	Size of evaluating batches (default: 256).
--metric_threshold	The decision value threshold over which a label is predicted as positive (default: 0.5).
--monitor_metrics	Metrics to monitor for evaluation (default: [‘P@1’, ‘P@3’, ‘P@5’]).
--val_metric	The metric to select the best model for testing (default: P@1).
--vocab_file	Path to a file holding vocabuaries (default: None).
--embed_file	Path to a file holding pre-trained embeddings (default: None).
--label_file	Path to a file holding all labels (default: None).
--save_k_predictions	Save top k predictions on test set. k=100 if not specified. (default: 0).
--predict_out_path	Path to the output file holding label results (default: ./predictions.txt).
--limit_train_batches	Percentage of train dataset to use for auto-testing (default: 1.0).
--limit_val_batches	Percentage of validation dataset to use for auto-testing (default: 1.0).
--limit_test_batches	Percentage of test dataset to use for auto-testing (default: 1.0).
--cpu	Disable CUDA.
--silent	Enable silent mode.
--data_workers	Use multi-cpu core for data pre-processing (default: 4).
--embed_cache_dir	For parameter search only: path to a directory for storing embeddings for multiple runs. (default: None).
--eval	Only run evaluation on the test set (default: None).
--checkpoint_path	The checkpoint to warm-up with (default: None).
--linear	Train linear model.
--data_format	‘svm’ for SVM format or ‘txt’ for LibMultiLabel format (default: txt).
--liblinear_options	Options passed to liblinear (default: None).
--linear_technique	Technique for linear classification (default: 1vsrest). One of {1vsrest, thresholding, cost_sensitive, cost_sensitive_micro, binary_and_multiclass, tree}.
--save_positive_predictions	Save all the predictions with decision value larger then 0. If used, the save_k_predictions must be set to 0.
--tree_degree	Degree of the tree (default: 100).
--tree_max_depth	Maximum depth of the tree (default: 10).
--beam_width	The width of the beam search (default: 10).
--cluster_size	the maximal number of labels inside a cluster (default: 8).
-h --help	If you are trying to specify network config such as dropout or activation or config of the learning rate scheduler, use a yaml file instead. See example configs in example_config.