Required:
--dict-save-path: Dictionary save path after training (str)--dict-train-path: Dictionary training data folder or file (str)
Optional:
-
--vocab-size: Size of vocabulary, (int, default: 30000) -
--min-freq: Minimum token frequency (float, default: 0.0) -
--max-freq: Maximum token frequency (float, default: 1.0) -
--file-batch-size: Dictionary is trained in batches of this size (int, default: 8192) -
--prune-every-n: Dictionary is pruned every n batch (int, default: 200)
Required:
--model-dir: Model checkpoint and log save path (str)--train-data-path: Training data folder or file' (str)
Required when creating the model for the first time:
--dict-path: Path to existing Dictionary (str, default: None)
Optional, locked in when creating the model for the first time:
--cell-type: Cell type, either 'gru' or 'lstm' (str, default: 'lstm--hidden-dim: Number of neurons in hidden layers (int, default: 32)--attn-dim: Number of neurons in to use in attention. None means attn-dim = hidden-dim (int, default: None)--embedding-dim: Embedding dimension (int, default: 16)--depth: Number of hidden layers in encoder and decoder (int, default: 2)--attn-type: Attention type, either 'bahdanau' or 'luong' (str, default: 'bahdanau--attn-input-feeding: Whether attention is fed to decoder inputs (bool, default: True)--use-residual: Whether to use residual connection (bool, default: False)--reverse-source: Whether to reverse encode input sequences (bool, default: True)--keep-every-n-hours: Checkpoint keep interval (int, default: 1)
Optional:
--max-seq-len: Maximum sequence length (int, default: 100)--optimizer: Optimizer to use, should be 'adadelta', 'adam' or 'rmsprop' (str, default: 'adam--learning-rate: Learning rate of the optimizer (float, default 0.0001)--max-gradient-norm: Maximum norm to clip gradient (float, default: 1.0)--dropout-rate: Hidden layer dropout (float, default: 0.2)
Optional:´
--batch-size: Batch size to feed into model (int, default: 32)--file-batch-size: Number of rows to read in-memory from each file (int, default: 8192)--max-file-pool-size: Maximum number of files to cycle at a time (int, default: 50)--shuffle-files: Shuffle files before reading (bool, default: True)--shuffle-file-batches: Shuffle file batches before training (bool, default: True)--save-every-n-batch: Save model checkpoint every n batch (int, default: 1000)--validation-data-path: Validation data folder or file (str, default: None)--validate-every-n-batch: Save model checkpoint every n batch (int, default 100)--validate-n-rows: Number of rows to read from validation file (int, default: None)
Required:
--model-dir: Model checkpoint and log save path (str)--test-data-path: Path to source data to decode (str)--decoded-data-path: Output file for decoded documents (str)
Optional:
--beam-width: Number of beams when using beamsearch. When beam-width=1, greedy decoder will be used instead. (int, default: 1)--max-decode-step: Maximum sequence length when decoding (int, default: 30)--batch-size: Batch size to feed into model (int, default: 32)--file-batch-size: Number of rows to read in-memory from each file (int, default: 8192)--max-file-pool-size: Maximum number of files to cycle at a time (int, default: 50)