SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Generate: TF compute_transition_scores (#21341)

Browse Source
This commit is contained in:
Joao Gante
2023-02-08 16:36:43 +00:00
committed by GitHub
parent d3046dad80
commit 1d9c26a4b8
5 changed files with 520 additions and 250 deletions
Download Patch File Download Diff File Expand all files Collapse all files

193
tests/generation/test_utils.py
View File

@@ -17,8 +17,6 @@
import inspect
import unittest
import numpy as np
from transformers import is_torch_available, pipeline
from transformers.testing_utils import require_torch, slow, torch_device
@@ -2220,165 +2218,6 @@ class GenerationIntegrationTests(unittest.TestCase, GenerationIntegrationTestsMi
self.assertListEqual(output_sequences_no_mask.tolist(), output_sequences_with_mask.tolist())
def test_transition_scores_greedy_search(self):
articles = ["Justin Timberlake", "Michael Phelps"]
tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
tokenizer.pad_token = tokenizer.eos_token
model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(
input_ids=input_ids,
max_new_tokens=5,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores)
expected_scores = np.array(
[
[0.3596273, 0.39646253, 0.46157718, 0.4594633, 0.44866616],
[0.34934354, 0.4935004, 0.6373219, 0.5173545, 0.57517034],
]
)
self.assertTrue(np.allclose(transition_scores.cpu().numpy(), expected_scores))
def test_transition_scores_greedy_search_normalized(self):
articles = ["Justin Timberlake", "Michael Phelps"]
tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
tokenizer.pad_token = tokenizer.eos_token
model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(
input_ids=input_ids,
max_new_tokens=5,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, normalize_logits=True)
expected_scores = np.array(
[
[-6.5532393, -6.5158753, -6.451863, -6.4527144, -6.459402],
[-6.5685124, -6.4277077, -6.282607, -6.399295, -6.340927],
]
)
self.assertTrue(np.allclose(transition_scores.cpu().numpy(), expected_scores))
def test_transition_scores_beam_search_encoder_decoder(self):
articles = [
"Justin Timberlake and Jessica Biel, welcome to parenthood.",
"Michael Phelps is arguably the most decorated Olympian of all time.",
]
tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
model = BartForConditionalGeneration.from_pretrained(
"hf-internal-testing/tiny-random-bart",
max_length=10,
num_beams=4,
num_return_sequences=2,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
length_penalty=0.0,
)
model = model.to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(input_ids=input_ids)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
transition_scores_sum = transition_scores.sum(-1)
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
def test_transition_scores_beam_search_encoder_decoder_with_eos(self):
articles = [
"Justin Timberlake and Jessica Biel, welcome to parenthood.",
"Michael Phelps is arguably the most decorated Olympian of all time.",
]
tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
model = BartForConditionalGeneration.from_pretrained(
"hf-internal-testing/tiny-random-bart",
max_length=10,
num_beams=4,
num_return_sequences=2,
return_dict_in_generate=True,
output_scores=True,
length_penalty=0.0,
)
model = model.to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(input_ids=input_ids)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
transition_scores_sum = transition_scores.sum(-1)
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
def test_transition_scores_beam_search_decoder_only(self):
articles = [
"Justin Timberlake",
"Michael Phelps",
]
tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
tokenizer.pad_token = tokenizer.eos_token
model = GPT2LMHeadModel.from_pretrained(
"hf-internal-testing/tiny-random-gpt2",
max_length=10,
num_beams=4,
num_return_sequences=2,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
length_penalty=0.0,
)
model = model.to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(input_ids=input_ids)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
transition_scores_sum = transition_scores.sum(-1)
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
def test_transition_scores_beam_sample_encoder_decoder(self):
articles = [
"Justin Timberlake and Jessica Biel, welcome to parenthood.",
"Michael Phelps is arguably the most decorated Olympian of all time.",
]
tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
model = BartForConditionalGeneration.from_pretrained(
"hf-internal-testing/tiny-random-bart",
do_sample=True,
max_length=10,
num_beams=4,
num_return_sequences=2,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
length_penalty=0.0,
)
model = model.to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(input_ids=input_ids)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
transition_scores_sum = transition_scores.sum(-1)
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
def test_transition_scores_group_beam_search_encoder_decoder(self):
articles = [
"Justin Timberlake and Jessica Biel, welcome to parenthood.",
@@ -2406,38 +2245,6 @@ class GenerationIntegrationTests(unittest.TestCase, GenerationIntegrationTestsMi
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
@slow
def test_transition_scores_early_stopping(self):
# This is an aggressive test that makes sure that `beam_search's`
# transition scores are computed correctly for varying `num_return_sequences`,
# `num_beams` and `batch_size > 1`
# 2 x input_ids for "question: How are you? \n context: I had a long day, "
input_ids = torch.tensor(2 * [[822, 10, 571, 33, 25, 58, 2625, 10, 27, 141, 3, 9, 307, 239, 6, 1]]).to(
torch_device
)
model = AutoModelForSeq2SeqLM.from_pretrained("t5-small").to(torch_device)
result = model.generate(
input_ids,
max_length=10,
return_dict_in_generate=True,
output_scores=True,
forced_eos_token_id=model.config.eos_token_id,
num_beams=4,
do_sample=False,
num_return_sequences=3,
length_penalty=0.0,
)
transition_scores = model.compute_transition_scores(
sequences=result.sequences, scores=result.scores, beam_indices=result.beam_indices
)
sum_transition_scores = torch.sum(transition_scores, dim=1)
self.assertListEqual(sum_transition_scores.cpu().tolist(), result.sequences_scores.cpu().tolist())
def test_log_scores_sample_decoder_only(self):
articles = ["I need input_ids to generate", "Short and"]
tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")