def test_temperature_dist_warper(self):
input_ids = None
length = 20
scores = self._get_uniform_logits(batch_size=2, length=length)
# tweak scores to not be uniform anymore
scores[1, 5] = (1 / length) + 0.1 # peak, 1st batch
scores[1, 10] = (1 / length) - 0.4 # valley, 1st batch
# compute softmax
probs = nn.functional.softmax(scores, dim=-1)
temp_dist_warper_sharper = TemperatureLogitsWarper(temperature=0.5)
temp_dist_warper_smoother = TemperatureLogitsWarper(temperature=1.3)
warped_prob_sharp = nn.functional.softmax(temp_dist_warper_sharper(input_ids, scores.clone()), dim=-1)
warped_prob_smooth = nn.functional.softmax(temp_dist_warper_smoother(input_ids, scores.clone()), dim=-1)
# uniform distribution stays uniform
self.assertTrue(torch.allclose(probs[0, :], warped_prob_sharp[0, :], atol=1e-3))
self.assertTrue(torch.allclose(probs[0, :], warped_prob_smooth[0, :], atol=1e-3))
# sharp peaks get higher, valleys get lower
self.assertLess(probs[1, :].max(), warped_prob_sharp[1, :].max())
self.assertGreater(probs[1, :].min(), warped_prob_sharp[1, :].min())
# smooth peaks get lower, valleys get higher
self.assertGreater(probs[1, :].max(), warped_prob_smooth[1, :].max())
self.assertLess(probs[1, :].min(), warped_prob_smooth[1, :].min())
def _get_warper_and_kwargs(num_beams):
warp_kwargs = {"top_k": 10, "top_p": 0.7, "temperature": 0.7}
logits_warper = LogitsProcessorList(
[
TopKLogitsWarper(top_k=warp_kwargs["top_k"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
TopPLogitsWarper(top_p=warp_kwargs["top_p"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
TemperatureLogitsWarper(warp_kwargs["temperature"]),
]
)
return warp_kwargs, logits_warper
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 = input_ids.clone()
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = scores.clone()
# instantiate all dist processors
min_dist_proc = MinLengthLogitsProcessor(min_length=10,
eos_token_id=eos_token_id)
temp_dist_warp = TemperatureLogitsWarper(temperature=0.5)
rep_penalty_proc = RepetitionPenaltyLogitsProcessor(penalty=2.0)
top_k_warp = TopKLogitsWarper(3)
top_p_warp = TopPLogitsWarper(0.8)
no_repeat_proc = NoRepeatNGramLogitsProcessor(2)
no_bad_words_dist_proc = NoBadWordsLogitsProcessor(
bad_words_ids=[[1]], eos_token_id=eos_token_id)
# no processor list
scores = min_dist_proc(input_ids, scores)
scores = temp_dist_warp(input_ids, scores)
scores = rep_penalty_proc(input_ids, scores)
scores = top_k_warp(input_ids, scores)
scores = top_p_warp(input_ids, scores)
scores = no_repeat_proc(input_ids, scores)
scores = no_bad_words_dist_proc(input_ids, scores)
# with processor list
processor = LogitsProcessorList([
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)
# scores should be equal
self.assertTrue(torch.allclose(scores, scores_comp, atol=1e-3))
# input_ids should never be changed
self.assertListEqual(input_ids.tolist(), input_ids_comp.tolist())
def _get_logits_warper_list(num_beams, temperature, top_k, top_p):
warpers = LogitsProcessorList()
if temperature is not None and temperature != 1.0:
warpers.append(TemperatureLogitsWarper(temperature))
if top_k is not None and top_k != 0:
warpers.append(
TopKLogitsWarper(
top_k=top_k,
min_tokens_to_keep=(2 if num_beams > 1 else 1)))
if top_p is not None and top_p < 1.0:
warpers.append(
TopPLogitsWarper(
top_p=top_p,
min_tokens_to_keep=(2 if num_beams > 1 else 1)))
return warpers