def test_processor_list(self):
batch_size = 4
cur_len = 10
vocab_size = 15
eos_token_id = 0
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, cur_len), vocab_size)
input_ids_comp = tf.identity(input_ids)
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = tf.identity(scores)
# instantiate all dist processors
min_dist_proc = TFMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
temp_dist_warp = TFTemperatureLogitsWarper(temperature=0.5)
rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
top_k_warp = TFTopKLogitsWarper(3)
top_p_warp = TFTopPLogitsWarper(0.8)
no_repeat_proc = TFNoRepeatNGramLogitsProcessor(2)
no_bad_words_dist_proc = TFNoBadWordsLogitsProcessor(bad_words_ids=[[1]], eos_token_id=eos_token_id)
# no processor list
scores = min_dist_proc(input_ids, scores, cur_len)
scores = temp_dist_warp(input_ids, scores)
scores = rep_penalty_proc(input_ids, scores, cur_len)
scores = top_k_warp(input_ids, scores)
scores = top_p_warp(input_ids, scores)
scores = no_repeat_proc(input_ids, scores, cur_len)
scores = no_bad_words_dist_proc(input_ids, scores, cur_len)
# with processor list
processor = TFLogitsProcessorList(
[
min_dist_proc,
temp_dist_warp,
rep_penalty_proc,
top_k_warp,
top_p_warp,
no_repeat_proc,
no_bad_words_dist_proc,
]
)
scores_comp = processor(input_ids, scores_comp, cur_len=cur_len)
# remove inf
scores = tf.where(tf.math.is_inf(scores), -1e9, scores)
scores_comp = tf.where(tf.math.is_inf(scores_comp), -1e9, scores_comp)
# scores should be equal
tf.debugging.assert_near(scores, scores_comp, atol=1e-3)
# input_ids should never be changed
self.assertListEqual(input_ids.numpy().tolist(), input_ids_comp.numpy().tolist())
def test_processor_list(self):
batch_size = 4
sequence_length = 10
vocab_size = 15
eos_token_id = 0
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
input_ids_comp = tf.identity(input_ids)
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = tf.identity(scores)
# instantiate all dist processors
min_dist_proc = TFMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
rep_penalty_proc = TFRepetitionPenaltyLogitsProcessor(penalty=2.0)
no_repeat_proc = TFNoRepeatNGramLogitsProcessor(2)
no_bad_words_dist_proc = TFNoBadWordsLogitsProcessor(bad_words_ids=[[1]], eos_token_id=eos_token_id)
# no processor list
scores = min_dist_proc(input_ids, scores)
scores = rep_penalty_proc(input_ids, scores)
scores = no_repeat_proc(input_ids, scores)
scores = no_bad_words_dist_proc(input_ids, scores)
# with processor list
processor = TFLogitsProcessorList(
[
min_dist_proc,
rep_penalty_proc,
no_repeat_proc,
no_bad_words_dist_proc,
]
)
scores_comp = processor(input_ids, scores_comp)
# remove inf
scores = set_tensor_by_indices_to_value(scores, tf.math.is_inf(scores), -1e9)
scores_comp = set_tensor_by_indices_to_value(scores_comp, tf.math.is_inf(scores_comp), -1e9)
# scores should be equal
tf.debugging.assert_near(scores, scores_comp, atol=1e-3)
# input_ids should never be changed
self.assertListEqual(input_ids.numpy().tolist(), input_ids_comp.numpy().tolist())
def test_no_bad_words_dist_processor(self):
vocab_size = 5
batch_size = 2
eos_token_id = 4
input_ids = tf.constant([[0, 1, 3, 1], [0, 1, 0, 1]], dtype=tf.int32)
bad_word_tokens = [[1], [4], [1, 0], [0, 1, 2], [1, 3, 1, 3]]
scores = self._get_uniform_logits(batch_size, vocab_size)
no_bad_words_dist_proc = TFNoBadWordsLogitsProcessor(bad_words_ids=bad_word_tokens, eos_token_id=eos_token_id)
filtered_scores = no_bad_words_dist_proc(input_ids, tf.identity(scores))
# batch 1: 1st, 2nd, and 4th (0, 1, 3) token are forbidden
# batch 2: 1st, 2nd, and 3rd (0, 1, 2) token are forbidden
self.assertListEqual(
tf.math.is_inf(filtered_scores).numpy().tolist(),
[[True, True, False, True, True], [True, True, True, False, True]],
)