Command line arguments¶
Usage for model files being run with ExperimentBuilder
.
usage: <model_file>.py [-h] [--model_kwargs MODEL_KWARGS] [--train] [--valid]
[--test] [--start_epoch START_EPOCH]
[--end_epoch END_EPOCH]
[--checkpoint_path CHECKPOINT_PATH]
[--ema_checkpoint_path EMA_CHECKPOINT_PATH]
[--batch_size BATCH_SIZE]
[--learning_rate LEARNING_RATE]
[--lr_schedule_name LR_SCHEDULE_NAME]
[--lr_schedule_kwargs LR_SCHEDULE_KWARGS]
[--weight_decay WEIGHT_DECAY] [--ema_decay EMA_DECAY]
[--num_data_threads NUM_DATA_THREADS]
[--model_checkpoint_interval MODEL_CHECKPOINT_INTERVAL]
[--train_output_interval TRAIN_OUTPUT_INTERVAL]
[--valid_output_interval VALID_OUTPUT_INTERVAL]
[--test_output_interval TEST_OUTPUT_INTERVAL]
[--data_root DATA_ROOT] [--train_dir TRAIN_DIR]
[--valid_dir VALID_DIR] [--test_dir TEST_DIR]
[--train_id_list TRAIN_ID_LIST]
[--valid_id_list VALID_ID_LIST]
[--test_id_list TEST_ID_LIST]
[--normalisation_dir NORMALISATION_DIR]
[--experiments_base EXPERIMENTS_BASE] --experiment_name
EXPERIMENT_NAME [--sample_rate SAMPLE_RATE]
Named Arguments¶
- --model_kwargs
Settings for the model, a Python dictionary written in quotes.
Default: {}
- --train
If True, model will be trained for –num_epochs on –train_id_list.
Default: True
- --valid
If True, model will be evaluated on –valid_id_list every epoch.
Default: True
- --test
If True, generation for –test_id_list will be performed after training.
Default: False
- --start_epoch
The epoch number to start training at (will effect checkpoint saves).
Default: 1
- --end_epoch
Epoch to end training at.
Default: 50
- --checkpoint_path
If specified, the model will first load parameters from an existing checkpoint.
- --ema_checkpoint_path
If specified, the EMA model will first load parameters from an existing checkpoint.
- --batch_size
Batch size used for iteration over train/valid data.
Default: 32
- --learning_rate
Learning rate for Adam optimiser to use during training.
Default: 0.01
- --lr_schedule_name
Learning rate schedule to use during training.
Default: “constant”
- --lr_schedule_kwargs
Settings for learning rate schedule, a Python dictionary written in quotes.
Default: {}
- --weight_decay
L2 regularisation weight, default of 0 indication no L2 loss term.
Default: 0.0
- --ema_decay
If not 0, track exponential moving average of model parameters, used for generation.
Default: 0.0
- --num_data_threads
Number of threads used to load the data with.
Default: 1
- --model_checkpoint_interval
The number of epochs to wait between saving the model.
Default: 1
- --train_output_interval
The number of epochs to wait between generating output for training data.
Default: 10
- --valid_output_interval
The number of epochs to wait between generating output for validation data.
Default: 10
- --test_output_interval
The number of epochs to wait between generating output for test data.
Default: 10
- --data_root
Base directory containing all data.
Default: “data”
- --train_dir
Name of the sub-directory in –data_root containing training data.
Default: “train”
- --valid_dir
Name of the sub-directory in –data_root containing validation data.
Default: “valid”
- --test_dir
Name of the sub-directory in –data_root containing test data.
Default: “test”
- --train_id_list
File name in –train_dir containing basenames of training samples.
Default: “train_file_id_list.scp”
- --valid_id_list
File name in –valid_dir containing basenames of validation samples.
Default: “valid_file_id_list.scp”
- --test_id_list
File name in –test_dir containing basenames of test files.
Default: “test_file_id_list.scp”
- --normalisation_dir
Name of the sub-directory in –data_root containing normalisation data.
Default: “train”
- --experiments_base
Base directory where all experiments direct their output.
Default: “experiments”
- --experiment_name
Name of the sub-directory in –output_dir used for any output.
- --sample_rate
Sample rate of the waveforms generated.
Default: 16000