The Poet class¶
-
class
tflyrics.Poet(name: str = None, vocabulary: list = None, embedding_dim: int = 256, rnn_units: int = 1024, batch_size: int = 32)¶ An artificial poet.
A Poet object is an object that wraps a recurrent predictive TensorFlow model, and with it can predict the next character in a sequence of unicode characters extracted from a possibly large text corpus.
-
property
batch_size¶ Get the batch size of the poet’s internal model.
Get the model’s batch size, i.e. the number of inputs that can be fed at once to the model for traning or evaluation.
- Returns
the current batch size
-
build_model(batch_size: int = 1) → None¶ Build the Poet’s internal model.
- Parameters
batch_size – the number of inputs to be fed at once to the model
-
generate(start_string: str, n_gen_chars: int = 1000, temperature: float = 1.0) → str¶ Generate text using the poet’s internal model.
Generate a specified number of characters by feeding an initial string to the model and using this string as a starting point to repeatedly sample a character that is likely to appear after the previously generated character.
- Parameters
start_string – initial string fed to the model
n_gen_chars – number of characters to generate
temperature – how surprising vs. predictable characters should be
- Returns
a new, generated text
-
restore() → None¶ Restore the state of the Poet’s model from a checkpoint.
If checkpoints have been saved during training, set the Poet’s model parameters (weights) to the state of the latest checkpoint saved.
- Raises
ResourceWarning
-
train_on(train_data: object, val_data: object = None, n_epochs: int = 1, checkpoints: bool = False) → dict¶ Train the Poet’s internal model on a dataset.
Train the Poet’s internal model on a TensorFlow Dataset containing batches of sequences of text (encoded as integers). You can also add a validation dataset as a separate argument. Optionally specify the number of epochs (by default just 1 epoch), and whether a checkpoint of the model should be saved at the end of each training epoch. Returns the training history, as a dictionary whose values are lists; each list represents a metric, and each element in the list is the value of that metric at a certain epoch.
- Parameters
train_data – LyricsGenerator or TensorFlow Dataset
val_data – LyricsGenerator or TensorFlow Dataset
n_epochs – number of training epochs
checkpoints – whether or not to save checkpoints of the model
- Returns
the training history, as a dictionary of lists
-
property