-
Notifications
You must be signed in to change notification settings - Fork 2.3k
/
eval_neural_rescorer.py
330 lines (283 loc) · 13.8 KB
/
eval_neural_rescorer.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
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
# Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This script would evaluate a neural language model (Transformer) trained with
`examples/nlp/language_modeling/transformer_lm.py' as a rescorer for ASR systems.
Given a trained TransformerLMModel `.nemo` file, this script can be used to re-score the beams obtained from a beam
search decoder of an ASR model.
USAGE:
1. Obtain `.tsv` file with beams and their corresponding scores. Scores can be from a regular beam search decoder or
in fusion with an N-gram LM scores. For a given beam size `beam_size` and a number of examples
for evaluation `num_eval_examples`, it should contain (`beam_size` x `num_eval_examples`) lines of
form `beam_candidate_text \t score`. This file can be generated by `scripts/asr_language_modeling/ngram_lm/eval_beamsearch_ngram.py`.
2. Rescore the candidates:
python eval_neural_rescorer.py
--lm_model=[path to .nemo file of the LM]
--beams_file=[path to beams .tsv file]
--beam_size=[size of the beams]
--eval_manifest=[path to eval manifest .json file]
--batch_size=[batch size used for inference on the LM model]
--alpha=[the value for the parameter rescorer_alpha]
--beta=[the value for the parameter rescorer_beta]
You may find more info on how to use this script at:
https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html
"""
import contextlib
import inspect
import json
from argparse import ArgumentParser
import editdistance
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch
import tqdm
from transformers import AutoModelForCausalLM
from nemo.collections.nlp.models.language_modeling import TransformerLMModel
from nemo.collections.nlp.modules.common.tokenizer_utils import get_tokenizer
from nemo.utils import logging
class BeamScoresDataset(torch.utils.data.Dataset):
"""
Dataset to read the score file containing the beams and their score
Args:
data_path: path to the beams file
tokenizer: tokenizer of the LM model
manifest_path: manifest `.json` file which contains the ground truths transcripts
beam_size: the number of beams per sample
max_seq_length: the maximum length of sequences
"""
def __init__(self, data_path, tokenizer, manifest_path, beam_size=128, max_seq_length=256):
self.data = pd.read_csv(data_path, delimiter="\t", header=None)
self.tokenizer = tokenizer
self.ground_truths = []
with open(manifest_path, 'r', encoding='utf-8') as f_orig:
for line in f_orig:
item = json.loads(line)
self.ground_truths.append(item['text'])
self.beam_size = beam_size
self.max_seq_length = max_seq_length
if self.tokenizer.pad_id is not None:
self.pad_id = self.tokenizer.pad_id
elif self.tokenizer.eos_id is not None:
self.pad_id = self.tokenizer.eos_id
else:
logging.warning(f"Using 0 as pad_id as the tokenizer has no pad_id or eos_id.")
self.pad_id = 0
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
text = str(self.data[0][idx])
tokens = self.tokenizer.text_to_ids(text)
if self.tokenizer.bos_id is not None:
tokens = [self.tokenizer.bos_id] + tokens
if self.tokenizer.eos_id is not None:
tokens = tokens + [self.tokenizer.eos_id]
input_ids = [self.pad_id] * self.max_seq_length
input_ids[: len(tokens)] = tokens
input_ids = np.array(input_ids)
input_mask = np.zeros(self.max_seq_length)
input_mask[: len(tokens)] = 1
acoustic_score = self.data[1][idx]
dist = editdistance.eval(text.split(), self.ground_truths[idx // self.beam_size].split())
ref_len = len(self.ground_truths[idx // self.beam_size].split())
len_in_chars = len(str(self.data[0][idx]))
return input_ids, input_mask, acoustic_score, dist, ref_len, len_in_chars, idx
def linear_search_wer(
dists, scores1, scores2, total_len, coef_range=[0, 10], coef_steps=10000, param_name='parameter'
):
"""
performs linear search to find the best coefficient when two set of scores are getting linearly fused.
Args:
dists: Tesnor of the distances between the ground truth and the candidates with shape of [number of samples, beam size]
scores1: Tensor of the first set of scores with shape of [number of samples, beam size]
scores2: Tensor of the second set of scores with shape of [number of samples, beam size]
total_len: The total length of all samples
coef_range: the search range for the coefficient
coef_steps: the number of steps that the search range would get divided into
param_name: the name of the parameter to be used in the figure
Output:
(best coefficient found, best WER achieved)
"""
scale = scores1.mean().abs().item() / scores2.mean().abs().item()
left = coef_range[0] * scale
right = coef_range[1] * scale
coefs = np.linspace(left, right, coef_steps)
best_wer = 10000
best_coef = left
wers = []
for coef in coefs:
scores = scores1 + coef * scores2
wer = compute_wer(dists, scores, total_len)
wers.append(wer)
if wer < best_wer:
best_wer = wer
best_coef = coef
plt.plot(coefs, wers)
plt.title(f'WER% after rescoring with different values of {param_name}')
plt.ylabel('WER%')
plt.xlabel(param_name)
plt.show()
return best_coef, best_wer
def compute_wer(dists, scores, total_len):
"""
Sorts the candidates based on the scores and calculates the WER with the new top candidates.
Args:
dists: Tensor of the distances between the ground truth and the candidates with shape of [number of samples, beam size]
scores: Tensor of the scores for candidates with shape of [number of samples, beam size]
total_len: The total length of all samples
Output:
WER with the new scores
"""
indices = scores.max(dim=1, keepdim=True)[1]
wer = dists.gather(dim=1, index=indices).sum() / total_len
wer = wer.item()
return wer
def main():
parser = ArgumentParser()
parser.add_argument(
"--lm_model_file",
type=str,
required=True,
help="path to LM model .nemo file or the name of a HuggingFace pretrained models like 'transfo-xl-wt103' or 'gpt2'",
)
parser.add_argument("--beams_file", type=str, required=True, help="path to beams .tsv file")
parser.add_argument(
"--eval_manifest", type=str, required=True, help="path to the evaluation `.json` manifest file"
)
parser.add_argument("--beam_size", type=int, required=True, help="number of beams per candidate")
parser.add_argument("--batch_size", type=int, default=256, help="inference batch size")
parser.add_argument("--alpha", type=float, default=None, help="parameter alpha of the fusion")
parser.add_argument("--beta", type=float, default=None, help="parameter beta of the fusion")
parser.add_argument("--max_seq_length", default=512, help="Maximum sequence length (in tokens) for the input")
parser.add_argument(
"--scores_output_file", default=None, type=str, help="The optional path to store the rescored beams"
)
parser.add_argument(
"--device", default="cuda", type=str, help="The device to load the model onto to calculate the scores"
)
parser.add_argument(
"--use_amp", action="store_true", help="Whether to use AMP if available to calculate the scores"
)
args = parser.parse_args()
device = args.device
if device.startswith("cuda") and not torch.cuda.is_available():
logging.info(f"cuda is not available! switched to cpu.")
device = "cpu"
if args.lm_model_file.endswith(".nemo"):
nemo_model = True
logging.info("Attempting to initialize from .nemo file...")
model = TransformerLMModel.restore_from(
restore_path=args.lm_model_file, map_location=torch.device(device)
).eval()
model_tokenizer = model.tokenizer
else:
nemo_model = False
logging.info("Attempting to initialize from a pretrained model from HuggingFace...")
model = (
AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path=args.lm_model_file, is_decoder=True)
.to(device)
.eval()
)
model_tokenizer = get_tokenizer(tokenizer_name=args.lm_model_file)
max_seq_length = args.max_seq_length
dataset = BeamScoresDataset(args.beams_file, model_tokenizer, args.eval_manifest, args.beam_size, max_seq_length)
data_loader = torch.utils.data.DataLoader(dataset=dataset, batch_size=args.batch_size)
if args.use_amp:
if torch.cuda.is_available() and hasattr(torch.cuda, 'amp') and hasattr(torch.cuda.amp, 'autocast'):
logging.info("AMP is enabled!\n")
autocast = torch.cuda.amp.autocast
else:
@contextlib.contextmanager
def autocast():
yield
if "attention_mask" in inspect.getfullargspec(model.forward).args:
support_att_mask = True
else:
support_att_mask = False
logging.info(f"Rescoring with beam_size: {args.beam_size}")
logging.info("Calculating the scores...")
with autocast():
with torch.no_grad():
am_scores, lm_scores, dists, ref_lens, lens_in_chars = [], [], [], [], []
for batch in tqdm.tqdm(data_loader):
input_ids, input_mask, acoustic_score, dist, ref_len, len_in_chars, idx = batch
max_len_in_batch = input_mask.sum(dim=0).argmin().item()
input_ids, input_mask = input_ids[:, :max_len_in_batch], input_mask[:, :max_len_in_batch]
if torch.cuda.is_available():
input_ids, input_mask = input_ids.to(device), input_mask.to(device)
dist, acoustic_score, len_in_chars = (
dist.to(device),
acoustic_score.to(device),
len_in_chars.to(device),
)
# some models like Transformer-XL don't need attention_mask as input
if support_att_mask:
log_probs = model(input_ids=input_ids, attention_mask=input_mask)
else:
log_probs = model(input_ids=input_ids)
if not nemo_model:
log_probs = torch.nn.functional.log_softmax(log_probs.logits, dim=-1)
target_log_probs = log_probs[:, :-1].gather(2, input_ids[:, 1:].unsqueeze(2)).squeeze(2)
neural_lm_score = torch.sum(target_log_probs * input_mask[:, 1:], dim=-1)
am_scores.append(acoustic_score)
lm_scores.append(neural_lm_score)
dists.append(dist)
ref_lens.append(ref_len)
lens_in_chars.append(len_in_chars)
am_scores = torch.cat(am_scores).view(-1, args.beam_size)
lm_scores = torch.cat(lm_scores).view(-1, args.beam_size)
dists = torch.cat(dists).view(-1, args.beam_size)
ref_lens = torch.cat(ref_lens).view(-1, args.beam_size)
lens_in_chars = torch.cat(lens_in_chars).view(-1, args.beam_size).to(am_scores.dtype)
total_len = ref_lens[:, 0].sum()
model_wer = dists[:, 0].sum() / total_len
ideal_wer = dists.min(dim=1)[0].sum() / total_len
if args.alpha is None:
logging.info("Linear search for alpha...")
coef1, _ = linear_search_wer(
dists=dists, scores1=am_scores, scores2=lm_scores, total_len=total_len, param_name='alpha'
)
coef1 = np.round(coef1, 3)
logging.info(f"alpha={coef1} achieved the best WER.")
logging.info(f"------------------------------------------------")
else:
coef1 = args.alpha
scores = am_scores + coef1 * lm_scores
if args.beta is None:
logging.info("Linear search for beta...")
coef2, _ = linear_search_wer(
dists=dists, scores1=scores, scores2=lens_in_chars, total_len=total_len, param_name='beta'
)
coef2 = np.round(coef2, 3)
logging.info(f"beta={coef2} achieved the best WER.")
logging.info(f"------------------------------------------------")
else:
coef2 = args.beta
new_scores = am_scores + coef1 * lm_scores + coef2 * lens_in_chars
rescored_wer = compute_wer(dists, new_scores, total_len)
logging.info(f"Input beams WER: {np.round(model_wer.item() * 100, 2)}%")
logging.info(f"------------------------------------------------")
logging.info(f" +LM rescoring WER: {np.round(rescored_wer * 100, 2)}%")
logging.info(f" with alpha={coef1}, beta={coef2}")
logging.info(f"------------------------------------------------")
logging.info(f"Oracle WER: {np.round(ideal_wer.item() * 100, 2)}%")
logging.info(f"------------------------------------------------")
new_scores_flatten = new_scores.flatten()
if args.scores_output_file is not None:
logging.info(f'Saving the candidates with their new scores at `{args.scores_output_file}`...')
with open(args.scores_output_file, "w", encoding='utf-8') as fout:
for sample_id in range(len(dataset)):
fout.write(f"{dataset.data[0][sample_id]}\t{new_scores_flatten[sample_id]}\n")
if __name__ == '__main__':
main()