def sample_uniformly(self,
num_samples,
min_seq_len=None,
max_seq_len=None,
pad=True,
seed=None):
"""Samples valid integer-encoded sequences from the domain.
Args:
num_samples: The number of samples.
min_seq_len: The minimum sequence length of samples (inclusive).
max_seq_len: The maximum sequence length of samples (inclusive).
pad: Whether to pad sequences to the maximum length.
seed: Optional seed of the random number generator.
Returns:
A list with `num_samples` samples.
"""
if min_seq_len is None:
min_seq_len = self.min_length
if max_seq_len is None:
max_seq_len = self.length
random_state = utils.get_random_state(seed)
valid_token_ids = np.delete(self.vocab.token_ids, self.vocab.eos)
lengths = random_state.randint(min_seq_len, max_seq_len + 1,
num_samples)
seqs = [
random_state.choice(valid_token_ids, length) for length in lengths
]
if pad:
seqs = seq_utils.pad_sequences(seqs, self.length, self.vocab.eos)
return seqs
def sample_uniformly(self, num_samples, seed=None):
random_state = utils.get_random_state(seed)
return np.int32(
random_state.randint(size=[num_samples, self.length],
low=0,
high=self.vocab_size))