-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtest_metrics.py
208 lines (190 loc) · 12.9 KB
/
test_metrics.py
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
import os
import unittest
from unittest import TestCase
import evaluate
import torch
from transformers import AutoTokenizer
from mbr import MetricRunner, MBRConfig
from mbr.metrics import metric_is_source_based
class MetricUtilsTestCase(TestCase):
def setUp(self):
self.mbr_config = MBRConfig(
metric="chrf",
metric_output_field="score",
num_samples=3,
num_references=2,
)
self.tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
self.tokenizer.pad_token = self.tokenizer.eos_token
self.metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
self.inputs = [ # shape: (batch_size,)
"This is an input sentence.",
"This is another input sentence.",
]
self.samples = [ # num_samples x batch_size
["This is a sample sentence.", "Something totally different."],
["This is a sample sentence.", "This a third sample sentence."],
["Something totally different.", "This is a fourth sample sentence."],
]
self.references = [ # num_references x batch_size
["This is a reference sentence.", "This is another reference sentence."],
["This is a reference sentence.", "This is a fourth reference sentence."],
]
self.input_ids = self.tokenizer(self.inputs, return_tensors="pt", padding=True).input_ids
self.sample_ids = tuple([self.tokenizer(sample, return_tensors="pt", padding=True).input_ids for sample in self.samples])
self.reference_ids = tuple([self.tokenizer(reference, return_tensors="pt", padding=True).input_ids for reference in self.references])
def test_is_source_based__chrf(self):
chrf = evaluate.load("chrf")
self.assertFalse(metric_is_source_based(chrf))
def test_is_source_based__comet(self):
comet = evaluate.load("comet", "eamt22-cometinho-da")
self.assertTrue(metric_is_source_based(comet))
@unittest.skipIf(os.getenv("SKIP_SLOW_TESTS", False), "Requires extra dependencies")
def test_is_source_based__bleurt(self):
bleurt = evaluate.load("bleurt")
self.assertFalse(metric_is_source_based(bleurt))
def test_load_metric(self):
self.mbr_config.metric = "chrf"
metric = self.metric_runner._load_metric()
self.assertIsInstance(metric, evaluate.Metric)
self.assertEqual(metric.name, "chr_f")
self.mbr_config.metric = evaluate.load("chrf")
metric = self.metric_runner._load_metric()
self.assertIsInstance(metric, evaluate.Metric)
self.assertEqual(metric.name, "chr_f")
def test_metric_config_name(self):
self.mbr_config.metric = "comet"
self.mbr_config.metric_config_name = "eamt22-cometinho-da"
self.mbr_config.metric_output_field = "mean_score"
metric = self.metric_runner._load_metric()
self.assertIsInstance(metric, evaluate.Metric)
self.assertEqual(metric.name, "comet")
# Test custom metric_config_name
self.assertEqual(metric.scorer.encoder.__class__.__name__, "MiniLMEncoder")
def test_compute_metric__chrf(self):
metric_output = self.metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertTrue(torch.is_floating_point(metric_output.scores_per_reference))
torch.testing.assert_close(metric_output.scores_per_reference.mean(dim=-1), metric_output.scores)
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
self.assertEqual(metric_output.scores_per_reference.shape, (2, 3, 2)) # batch_size x num_samples x num_references
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
torch.testing.assert_close(metric_output.scores_per_reference[0, 0, 0], metric_output.scores_per_reference[0, 1, 0])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])
def test_compute_metric__comet(self):
self.mbr_config.metric = evaluate.load("comet", "eamt22-cometinho-da")
self.mbr_config.metric.scorer.eval()
self.mbr_config.metric_output_field = "mean_score"
self.metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
self.assertEqual(self.metric_runner.metric.name, "comet")
metric_output = self.metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertTrue(torch.is_floating_point(metric_output.scores_per_reference))
torch.testing.assert_close(metric_output.scores_per_reference.mean(dim=-1), metric_output.scores)
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
self.assertEqual(metric_output.scores_per_reference.shape, (2, 3, 2)) # batch_size x num_samples x num_references
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
torch.testing.assert_close(metric_output.scores_per_reference[0, 0, 0], metric_output.scores_per_reference[0, 1, 0])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])
@unittest.skipIf(os.getenv("SKIP_SLOW_TESTS", False), "Requires extra dependencies")
def test_compute_metric__bleurt(self):
self.mbr_config.metric = evaluate.load("bleurt")
self.mbr_config.metric_output_field = "scores"
self.metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
self.assertEqual(self.metric_runner.metric.name, "bleurt")
metric_output = self.metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertTrue(torch.is_floating_point(metric_output.scores_per_reference))
torch.testing.assert_close(metric_output.scores_per_reference.mean(dim=-1), metric_output.scores)
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
self.assertEqual(metric_output.scores_per_reference.shape, (2, 3, 2)) # batch_size x num_samples x num_references
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
torch.testing.assert_close(metric_output.scores_per_reference[0, 0, 0], metric_output.scores_per_reference[0, 1, 0])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])
def test_comet_metric_runner(self):
from mbr.metrics.comet import CometMetricRunner
self.mbr_config.metric = evaluate.load("comet", "eamt22-cometinho-da")
self.mbr_config.metric.scorer.eval()
self.mbr_config.metric_output_field = "mean_score"
base_metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
self.assertEqual(base_metric_runner.metric.name, "comet")
self.assertFalse(base_metric_runner.metric.scorer.training)
comet_metric_runner = CometMetricRunner(self.mbr_config, self.tokenizer)
self.assertFalse(comet_metric_runner.metric.scorer.training)
# Output should be the same as the base MetricRunner
base_metric_scores = base_metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
metric_scores = comet_metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
torch.testing.assert_close(base_metric_scores, metric_scores)
def test_comet_metric_runner__cache(self):
"""Output should be identical irrespective of cache size"""
from mbr.metrics.comet import CometMetricRunner
self.mbr_config.metric = evaluate.load("comet", "eamt22-cometinho-da")
self.mbr_config.metric_output_field = "mean_score"
base_metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
base_metric_scores = base_metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertEqual(base_metric_runner.metric.name, "comet")
for cache_size in [1, 4, 8]:
self.mbr_config.metric_cache_size = cache_size
comet_metric_runner = CometMetricRunner(self.mbr_config, self.tokenizer)
metric_scores = comet_metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
torch.testing.assert_close(base_metric_scores, metric_scores)
def test_comet_metric_runner__aggregate(self):
from mbr.metrics.comet import AggregateCometMetricRunner
self.mbr_config.metric = evaluate.load("comet", "eamt22-cometinho-da")
self.mbr_config.metric.scorer.eval()
self.mbr_config.metric_output_field = "mean_score"
base_metric_runner = MetricRunner(self.mbr_config, self.tokenizer)
self.assertEqual(base_metric_runner.metric.name, "comet")
self.assertFalse(base_metric_runner.metric.scorer.training)
comet_metric_runner = AggregateCometMetricRunner(self.mbr_config, self.tokenizer)
self.assertFalse(comet_metric_runner.metric.scorer.training)
metric_output = comet_metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertIsNone(metric_output.scores_per_reference)
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])
def test_fastchrf_metric_runner__aggregate(self):
from mbr.metrics.fastchrf import FastChrfMetricRunner
metric_runner = FastChrfMetricRunner(self.mbr_config, self.tokenizer, compute_pairwise_average=False)
metric_output = metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertIsNone(metric_output.scores_per_reference)
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])
def test_fastchrf_metric_runner__pairwise(self):
from mbr.metrics.fastchrf import FastChrfMetricRunner
metric_runner = FastChrfMetricRunner(self.mbr_config, self.tokenizer, compute_pairwise_average=True)
metric_output = metric_runner(self.input_ids, self.sample_ids, self.reference_ids)
self.assertTrue(torch.is_floating_point(metric_output.scores))
self.assertTrue(torch.is_floating_point(metric_output.scores_per_reference))
self.assertEqual(metric_output.scores.shape, (2, 3)) # batch_size x num_samples
self.assertEqual(metric_output.scores_per_reference.shape, (2, 3, 2)) # batch_size x num_samples x num_references
# Duplicate samples should have the same scores
torch.testing.assert_close(metric_output.scores[0, 0], metric_output.scores[0, 1])
torch.testing.assert_close(metric_output.scores_per_reference[0, 0, 0], metric_output.scores_per_reference[0, 1, 0])
# The metric scores should rank as expected, given the test strings in self.samples and self.references
self.assertGreater(metric_output.scores[0, 0], metric_output.scores[0, 2])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 1])
self.assertLess(metric_output.scores[1, 0], metric_output.scores[1, 2])