-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluator.py
247 lines (203 loc) · 8.85 KB
/
evaluator.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
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
# standard imports
import math
import logging
from typing import Union, Tuple, List, Dict
# third-party imports
import sacrebleu
import youtokentome
import tensorflow as tf
from tqdm import tqdm
from colorama import Fore, init
# module imports
from model import Transformer
from dataloader import SequenceLoader
# logging house-keeping
init(autoreset=True)
logging.basicConfig(level=logging.INFO)
class Evaluator:
"""Utility class to evaluate a language model for the task of translation."""
def __init__(
self, model: Transformer, test_loader: SequenceLoader, bpe_model_path: str
):
"""
Initializes the Evaluator.
:param model: the Transformer model
:param test_loader: the sequence loader in test configuration
:param bpe_model_path: the path to the Byte-Pair Encoding model
"""
self.model = model
self.test_loader = test_loader
self.bpe_model = youtokentome.BPE(model=bpe_model_path)
def load_checkpoint(self, checkpoint_dir: str):
"""
Loads the model from the latest checkpoint.
:param checkpoint_dir: path to the directory containing the checkpoints
"""
checkpoint_path = tf.train.latest_checkpoint(checkpoint_dir)
if not checkpoint_path:
raise ValueError(f"No checkpoint found in {checkpoint_dir}")
# Dummy data to build the model
dummy_encoder_sequences = tf.zeros((1, 1), dtype=tf.int32)
dummy_decoder_sequences = tf.zeros((1, 1), dtype=tf.int32)
dummy_encoder_sequence_lengths = tf.zeros((1,), dtype=tf.int32)
dummy_decoder_sequence_lengths = tf.zeros((1,), dtype=tf.int32)
# Calling the model on the dummy data to build it
self.model(
encoder_sequences=dummy_encoder_sequences,
decoder_sequences=dummy_decoder_sequences,
encoder_sequence_lengths=dummy_encoder_sequence_lengths,
decoder_sequence_lengths=dummy_decoder_sequence_lengths,
training=False,
)
self.model.load_weights(checkpoint_path)
def evaluate(self, length_norm_coefficient: float = 0.6, k: int = 5):
"""
Evaluates the model using BLUE score.
:param length_norm_coefficient: coefficient for normalizing decoded sequences' scores by their lengths
:param k: beam size, when k = 1 the translation is equivalent to performing greedy decoding
:return:
"""
hypotheses = []
references = []
COLORS = [Fore.GREEN, Fore.YELLOW, Fore.CYAN, Fore.MAGENTA]
for _, (
source_seqs,
target_seqs,
source_seq_lengths,
target_seq_lengths,
) in enumerate(tqdm(self.test_loader, total=self.test_loader.n_batches)):
hypotheses.append(
self.translate(
source_sequence=source_seqs,
length_norm_coefficient=length_norm_coefficient,
k=k,
)[0]
)
references.extend(
self.test_loader.bpe_model.decode(
target_seqs.numpy().tolist(), ignore_ids=[0, 2, 3]
)
)
for i, (print_text, sacrebleu_text) in enumerate(
zip(
[
"13a tokenization, cased",
"13a tokenization, caseless",
"International tokenization, cased",
"International tokenization, caseless",
],
[
sacrebleu.corpus_bleu(hypotheses, [references]),
sacrebleu.corpus_bleu(hypotheses, [references], lowercase=True),
sacrebleu.corpus_bleu(hypotheses, [references], tokenize="intl"),
sacrebleu.corpus_bleu(
hypotheses, [references], tokenize="intl", lowercase=True
),
],
)
):
colored_log_message = f"{COLORS[i]}{print_text}"
colored_sacrebleu = f"{COLORS[i]}{sacrebleu_text}"
logging.info(colored_log_message)
logging.info(colored_sacrebleu)
def translate(
self,
source_sequence: Union[tf.Tensor, str],
length_norm_coefficient: float = 0.6,
k: int = 5,
) -> Tuple[str, List[Dict[str, Union[str, float]]]]:
"""
Translates a source language sequence into the target language, with beam search decoding.
:param source_sequence: the source language sequence, either a string or a Tensor of byte pair encoded indices
:param length_norm_coefficient: coefficient for normalizing decoded sequences' scores by their lengths
:param k: beam size, when k = 1 the translation is equivalent to performing greedy decoding
:return: the best hypothesis as well as all candidate hypotheses
"""
n_completed_hypotheses = min(k, 10) # minimum number of hypotheses to complete
vocab_size = self.bpe_model.vocab_size()
if isinstance(source_sequence, str):
encoder_sequences = self.bpe_model.encode(
source_sequence,
output_type=youtokentome.OutputType.ID,
bos=False,
eos=False,
)
encoder_sequences = tf.expand_dims(
encoder_sequences, 0
) # (1, source_sequence_length)
else:
encoder_sequences = source_sequence
encoder_sequence_lengths = tf.constant(
[encoder_sequences.shape[1]], dtype=tf.int32
) # (1)
encoder_sequences = self.model.encoder(
encoder_sequences=encoder_sequences,
encoder_sequence_lengths=encoder_sequence_lengths,
training=False,
) # (1, source_sequence_length, d_model)
hypotheses = tf.constant([[self.bpe_model.subword_to_id("<BOS>")]]) # (1, 1)
hypotheses_lengths = tf.constant([hypotheses.shape[1]], dtype=tf.int32) # (1)
hypotheses_scores = tf.zeros(1)
completed_hypotheses = []
completed_hypotheses_scores = []
step = 1
while True:
s = tf.shape(hypotheses)[0]
decoder_sequences = self.model.decoder(
decoder_sequences=hypotheses, # (s, step, vocab_size)
decoder_sequence_lengths=hypotheses_lengths,
encoder_sequences=tf.repeat(encoder_sequences, s, axis=0),
encoder_sequence_lengths=tf.repeat(encoder_sequence_lengths, s),
)
scores = decoder_sequences[:, -1, :] # (s, vocab_size)
scores = tf.nn.log_softmax(scores, axis=-1) # (s, vocab_size)
scores = tf.expand_dims(hypotheses_scores, 1) + scores # (s, vocab_size)
top_k_hypotheses_scores, unrolled_indices = tf.math.top_k(
tf.reshape(scores, [-1]), k
)
prev_word_indices = unrolled_indices // vocab_size # (k)
next_word_indices = unrolled_indices % vocab_size # (k)
top_k_hypotheses = tf.concat(
[
tf.gather(hypotheses, prev_word_indices),
tf.expand_dims(next_word_indices, 1),
],
axis=1,
) # (k, step + 1)
complete = tf.equal(
next_word_indices, self.bpe_model.subword_to_id("<EOS>")
) # (k)
completed_hypotheses.extend(
tf.boolean_mask(top_k_hypotheses, complete).numpy().tolist()
)
norm = math.pow(((5 + step) / (5 + 1)), length_norm_coefficient)
completed_hypotheses_scores.extend(
(tf.boolean_mask(top_k_hypotheses_scores, complete) / norm)
.numpy()
.tolist()
)
if len(completed_hypotheses) >= n_completed_hypotheses:
break
hypotheses = tf.boolean_mask(top_k_hypotheses, ~complete) # (s, step + 1)
hypotheses_scores = tf.boolean_mask(
top_k_hypotheses_scores, ~complete
) # (s)
hypotheses_lengths = tf.fill(
[tf.shape(hypotheses)[0]], tf.shape(hypotheses)[1]
) # (s)
if step > 100:
break
step += 1
if len(completed_hypotheses) == 0:
completed_hypotheses = hypotheses.numpy().tolist()
completed_hypotheses_scores = hypotheses_scores.numpy().tolist()
all_hypotheses = []
for i, h in enumerate(
self.bpe_model.decode(completed_hypotheses, ignore_ids=[0, 2, 3])
):
all_hypotheses.append(
{"hypothesis": h, "score": completed_hypotheses_scores[i]}
)
i = completed_hypotheses_scores.index(max(completed_hypotheses_scores))
best_hypothesis = all_hypotheses[i]["hypothesis"]
return best_hypothesis, all_hypotheses