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 = jax.nn.softmax(scores, axis=-1)
temp_dist_warper_sharper = FlaxTemperatureLogitsWarper(temperature=0.5)
temp_dist_warper_smoother = FlaxTemperatureLogitsWarper(temperature=1.3)
warped_prob_sharp = jax.nn.softmax(temp_dist_warper_sharper(input_ids, scores.copy(), cur_len=None), axis=-1)
warped_prob_smooth = jax.nn.softmax(temp_dist_warper_smoother(input_ids, scores.copy(), cur_len=None), axis=-1)
# uniform distribution stays uniform
self.assertTrue(jnp.allclose(probs[0, :], warped_prob_sharp[0, :], atol=1e-3))
self.assertTrue(jnp.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 test_processor_list(self):
batch_size = 4
sequence_length = 10
vocab_size = 15
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
input_ids_comp = input_ids.copy()
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = scores.copy()
# instantiate all dist processors
temp_dist_warp = FlaxTemperatureLogitsWarper(temperature=0.5)
top_k_warp = FlaxTopKLogitsWarper(3)
top_p_warp = FlaxTopPLogitsWarper(0.8)
# no processor list
scores = temp_dist_warp(input_ids, scores)
scores = top_k_warp(input_ids, scores)
scores = top_p_warp(input_ids, scores)
# with processor list
processor = FlaxLogitsProcessorList(
[temp_dist_warp, top_k_warp, top_p_warp])
scores_comp = processor(input_ids, scores_comp)
# scores should be equal
self.assertTrue(jnp.allclose(scores, scores_comp, atol=1e-3))
# input_ids should never be changed
self.assertListEqual(input_ids.tolist(), input_ids_comp.tolist())
def test_processor_list_jitted(self):
batch_size = 4
sequence_length = 10
vocab_size = 15
eos_token_id = 2
bos_token_id = 1
max_length = 15
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
input_ids_comp = input_ids.copy()
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = scores.copy()
# instantiate all dist processors
temp_dist_warp = FlaxTemperatureLogitsWarper(temperature=0.5)
top_k_warp = FlaxTopKLogitsWarper(3)
top_p_warp = FlaxTopPLogitsWarper(0.8)
# instantiate all logits processors
min_dist_proc = FlaxMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
bos_dist_proc = FlaxForcedBOSTokenLogitsProcessor(bos_token_id=bos_token_id)
eos_dist_proc = FlaxForcedEOSTokenLogitsProcessor(max_length=max_length, eos_token_id=eos_token_id)
cur_len = 10
# no processor list
def run_no_processor_list(input_ids, scores, cur_len):
scores = temp_dist_warp(input_ids, scores, cur_len=cur_len)
scores = top_k_warp(input_ids, scores, cur_len=cur_len)
scores = top_p_warp(input_ids, scores, cur_len=cur_len)
scores = min_dist_proc(input_ids, scores, cur_len=cur_len)
scores = bos_dist_proc(input_ids, scores, cur_len=cur_len)
scores = eos_dist_proc(input_ids, scores, cur_len=cur_len)
return scores
# with processor list
def run_processor_list(input_ids, scores, cur_len):
processor = FlaxLogitsProcessorList(
[temp_dist_warp, top_k_warp, top_p_warp, min_dist_proc, bos_dist_proc, eos_dist_proc]
)
scores = processor(input_ids, scores, cur_len=cur_len)
return scores
jitted_run_no_processor_list = jax.jit(run_no_processor_list)
jitted_run_processor_list = jax.jit(run_processor_list)
scores = jitted_run_no_processor_list(input_ids, scores, cur_len)
scores_comp = jitted_run_processor_list(input_ids, scores_comp, cur_len)
# scores should be equal
self.assertTrue(jnp.allclose(scores, scores_comp, atol=1e-3))
# input_ids should never be changed
self.assertListEqual(input_ids.tolist(), input_ids_comp.tolist())