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

Generate: documented function to compute the transition scores (#21191)

Browse Source 
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
This commit is contained in:
Joao Gante
2023-01-20 12:50:01 +00:00
committed by GitHub
parent 91c2278b97
commit af37d183b3
3 changed files with 154 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
tests/generation/test_utils.py
View File

@@ -17,6 +17,8 @@
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
@@ -2485,6 +2487,58 @@ class GenerationIntegrationTests(unittest.TestCase):
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.",
@@ -2506,9 +2560,7 @@ class GenerationIntegrationTests(unittest.TestCase):
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_beam_scores(
outputs.sequences, outputs.scores, outputs.beam_indices
)
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))
@@ -2533,9 +2585,7 @@ class GenerationIntegrationTests(unittest.TestCase):
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_beam_scores(
outputs.sequences, outputs.scores, outputs.beam_indices
)
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))
@@ -2564,9 +2614,7 @@ class GenerationIntegrationTests(unittest.TestCase):
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_beam_scores(
outputs.sequences, outputs.scores, outputs.beam_indices
)
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))
@@ -2593,9 +2641,7 @@ class GenerationIntegrationTests(unittest.TestCase):
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_beam_scores(
outputs.sequences, outputs.scores, outputs.beam_indices
)
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))
@@ -2622,9 +2668,7 @@ class GenerationIntegrationTests(unittest.TestCase):
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_beam_scores(
outputs.sequences, outputs.scores, outputs.beam_indices
)
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))
@@ -2653,7 +2697,7 @@ class GenerationIntegrationTests(unittest.TestCase):
length_penalty=0.0,
)
transition_scores = model.compute_transition_beam_scores(
transition_scores = model.compute_transition_scores(
sequences=result.sequences, scores=result.scores, beam_indices=result.beam_indices
)