Spaces:
Sleeping
Sleeping
File size: 14,876 Bytes
6fc683c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import unittest
import torch
from fairseq.sequence_generator import SequenceGenerator
import tests.utils as test_utils
class TestSequenceGeneratorBase(unittest.TestCase):
def assertHypoTokens(self, hypo, tokens):
self.assertTensorEqual(hypo['tokens'], torch.LongTensor(tokens))
def assertHypoScore(self, hypo, pos_probs, normalized=True, lenpen=1.):
pos_scores = torch.FloatTensor(pos_probs).log()
self.assertAlmostEqual(hypo['positional_scores'], pos_scores)
self.assertEqual(pos_scores.numel(), hypo['tokens'].numel())
score = pos_scores.sum()
if normalized:
score /= pos_scores.numel()**lenpen
self.assertLess(abs(score - hypo['score']), 1e-6)
def assertAlmostEqual(self, t1, t2):
self.assertEqual(t1.size(), t2.size(), "size mismatch")
self.assertLess((t1 - t2).abs().max(), 1e-4)
def assertTensorEqual(self, t1, t2):
self.assertEqual(t1.size(), t2.size(), "size mismatch")
self.assertEqual(t1.ne(t2).long().sum(), 0)
class TestSequenceGenerator(TestSequenceGeneratorBase):
def setUp(self):
self.tgt_dict, self.w1, self.w2, src_tokens, src_lengths, self.model = (
test_utils.sequence_generator_setup()
)
self.sample = {
'net_input': {
'src_tokens': src_tokens, 'src_lengths': src_lengths,
},
}
def test_with_normalization(self):
generator = SequenceGenerator(self.tgt_dict, beam_size=2)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator(self.tgt_dict, beam_size=2, normalize_scores=False)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator(self.tgt_dict, beam_size=2, len_penalty=lenpen)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator(self.tgt_dict, beam_size=2, len_penalty=lenpen)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
def test_maxlen(self):
generator = SequenceGenerator(self.tgt_dict, beam_size=2, max_len_b=2)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.1, 0.6])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w2, w2, eos])
self.assertHypoScore(hypos[1][1], [0.3, 0.9, 0.01])
class TestDiverseBeamSearch(TestSequenceGeneratorBase):
def setUp(self):
# construct dummy dictionary
d = test_utils.dummy_dictionary(vocab_size=2)
self.assertEqual(d.pad(), 1)
self.assertEqual(d.eos(), 2)
self.assertEqual(d.unk(), 3)
self.eos = d.eos()
self.w1 = 4
self.w2 = 5
# construct source data
self.src_tokens = torch.LongTensor([
[self.w1, self.w2, self.eos],
[self.w1, self.w2, self.eos],
])
self.src_lengths = torch.LongTensor([2, 2])
args = argparse.Namespace()
unk = 0.
args.beam_probs = [
# step 0:
torch.FloatTensor([
# eos w1 w2
# sentence 1:
[0.0, unk, 0.9, 0.1], # beam 1
[0.0, unk, 0.9, 0.1], # beam 2
# sentence 2:
[0.0, unk, 0.7, 0.3],
[0.0, unk, 0.7, 0.3],
]),
# step 1:
torch.FloatTensor([
# eos w1 w2
# sentence 1:
[0.0, unk, 0.6, 0.4],
[0.0, unk, 0.6, 0.4],
# sentence 2:
[0.25, unk, 0.35, 0.4],
[0.25, unk, 0.35, 0.4],
]),
# step 2:
torch.FloatTensor([
# eos w1 w2
# sentence 1:
[1.0, unk, 0.0, 0.0],
[1.0, unk, 0.0, 0.0],
# sentence 2:
[0.9, unk, 0.1, 0.0],
[0.9, unk, 0.1, 0.0],
]),
]
task = test_utils.TestTranslationTask.setup_task(args, d, d)
self.model = task.build_model(args)
self.tgt_dict = task.target_dictionary
def test_diverse_beam_search(self):
generator = SequenceGenerator(
self.tgt_dict, beam_size=2, diverse_beam_groups=2, diverse_beam_strength=0.,
)
sample = {'net_input': {'src_tokens': self.src_tokens, 'src_lengths': self.src_lengths}}
hypos = generator.generate([self.model], sample)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 0.6, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.9])
class TestTopPSamplingSearch(TestSequenceGeneratorBase):
def setUp(self):
# construct dummy dictionary
d = test_utils.dummy_dictionary(vocab_size=2)
self.assertEqual(d.pad(), 1)
self.assertEqual(d.eos(), 2)
self.assertEqual(d.unk(), 3)
self.eos = d.eos()
self.w1 = 4
self.w2 = 5
# construct source data
self.src_tokens = torch.LongTensor([
[self.w1, self.w2, self.eos],
[self.w1, self.w2, self.eos],
])
self.src_lengths = torch.LongTensor([2, 2])
args = argparse.Namespace()
unk = 0.
# The minimal probability of top 2 tokens.
self.min_top2_prob = 0.75
# The minimal probability of the top 1 token.
self.min_top1_prob = 0.4
w1_prob = self.min_top1_prob
w2_prob = self.min_top2_prob - self.min_top1_prob
eos_prob = 1 - self.min_top2_prob
args.beam_probs = [
# step 0:
torch.FloatTensor([
# eos w1 w2
[0.0, unk, 1.0, 0.0],
[0.0, unk, 1.0, 0.0],
[0.0, unk, 1.0, 0.0],
[0.0, unk, 1.0, 0.0],
]),
# step 1:
torch.FloatTensor([
# eos w1 w2
[eos_prob, unk, w1_prob, w2_prob],
[eos_prob, unk, w1_prob, w2_prob],
[eos_prob, unk, w1_prob, w2_prob],
[eos_prob, unk, w1_prob, w2_prob],
]),
# step 2:
torch.FloatTensor([
# eos w1 w2
[1.0, unk, 0.0, 0.0],
[1.0, unk, 0.0, 0.0],
[1.0, unk, 0.0, 0.0],
[1.0, unk, 0.0, 0.0],
]),
]
task = test_utils.TestTranslationTask.setup_task(args, d, d)
self.model = task.build_model(args)
self.tgt_dict = task.target_dictionary
def test_topp_sampling_search_low_prob(self):
# Given a prob low enough to top-P sampling, we expect only the top
# 1 token to be sampled, which always results in the same output.
low_sampling_topp = self.min_top1_prob/2.0
generator = SequenceGenerator(
self.tgt_dict, beam_size=2, sampling=True,
sampling_topp=low_sampling_topp
)
sample = {
'net_input': {
'src_tokens': self.src_tokens,
'src_lengths': self.src_lengths
}
}
hypos = generator.generate([self.model], sample)
eos, w1 = self.eos, self.w1
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, w1, eos])
self.assertHypoScore(hypos[0][0], [1.0, 0.4, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, w1, eos])
self.assertHypoScore(hypos[0][1], [1.0, 0.4, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w1, eos])
self.assertHypoScore(hypos[1][0], [1.0, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w1, eos])
self.assertHypoScore(hypos[1][1], [1.0, 0.4, 1.0])
def test_topp_sampling_search_high_prob(self):
# Given a prob high enough to top-P sampling, any of the top 2
# tokens could be sampled. This can cause different outputs.
high_sampling_topp = (self.min_top1_prob+self.min_top2_prob)/2.0
generator = SequenceGenerator(
self.tgt_dict, beam_size=2, sampling=True,
sampling_topp=high_sampling_topp
)
sample = {
'net_input': {
'src_tokens': self.src_tokens,
'src_lengths': self.src_lengths
}
}
hypos = generator.generate([self.model], sample)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertTrue(self.hypoTokens(hypos[0][0], [w1, w1, eos]) or
self.hypoTokens(hypos[0][0], [w1, w2, eos]))
self.assertTrue(self.hypoScore(hypos[0][0], [1.0, 0.4, 1.0]) or
self.hypoScore(hypos[0][0], [1.0, 0.35, 1.0]))
# sentence 1, beam 2
self.assertTrue(self.hypoTokens(hypos[0][1], [w1, w1, eos]) or
self.hypoTokens(hypos[0][1], [w1, w2, eos]))
self.assertTrue(self.hypoScore(hypos[0][1], [1.0, 0.4, 1.0]) or
self.hypoScore(hypos[0][1], [1.0, 0.35, 1.0]))
# sentence 2, beam 1
self.assertTrue(self.hypoTokens(hypos[1][0], [w1, w1, eos]) or
self.hypoTokens(hypos[1][0], [w1, w2, eos]))
self.assertTrue(self.hypoScore(hypos[1][0], [1.0, 0.4, 1.0]) or
self.hypoScore(hypos[1][0], [1.0, 0.35, 1.0]))
# sentence 2, beam 2
self.assertTrue(self.hypoTokens(hypos[1][1], [w1, w1, eos]) or
self.hypoTokens(hypos[1][1], [w1, w2, eos]))
self.assertTrue(self.hypoScore(hypos[1][1], [1.0, 0.4, 1.0]) or
self.hypoScore(hypos[1][1], [1.0, 0.35, 1.0]))
def hypoTokens(self, hypo, tokens):
return self.tensorEqual(hypo['tokens'], torch.LongTensor(tokens))
def hypoScore(self, hypo, pos_probs, normalized=True, lenpen=1.):
pos_scores = torch.FloatTensor(pos_probs).log()
if not self.almostEqual(hypo['positional_scores'], pos_scores):
return False
if pos_scores.numel() != hypo['tokens'].numel():
return False
score = pos_scores.sum()
if normalized:
score /= pos_scores.numel() ** lenpen
return abs(score - hypo['score']) < 1e-6
def almostEqual(self, t1, t2):
return t1.size() == t2.size() and (t1 - t2).abs().max() < 1e-4
def tensorEqual(self, t1, t2):
return t1.size() == t2.size() and t1.ne(t2).long().sum() == 0
if __name__ == '__main__':
unittest.main()
|