Commit
·
1b3874c
1
Parent(s):
347f330
add eval script
Browse files
README.md
CHANGED
@@ -11,7 +11,7 @@ tags:
|
|
11 |
datasets:
|
12 |
- mozilla-foundation/common_voice_9_0
|
13 |
model-index:
|
14 |
-
- name:
|
15 |
results:
|
16 |
- task:
|
17 |
name: Automatic Speech Recognition
|
@@ -55,28 +55,14 @@ model-index:
|
|
55 |
value: 11.09
|
56 |
---
|
57 |
|
58 |
-
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
59 |
-
should probably proofread and complete it, then remove this comment. -->
|
60 |
|
61 |
-
#
|
62 |
|
63 |
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - FR dataset.
|
64 |
It achieves the following results on the evaluation set:
|
65 |
- Loss: 0.1430
|
66 |
- Wer: 0.1245
|
67 |
|
68 |
-
## Model description
|
69 |
-
|
70 |
-
More information needed
|
71 |
-
|
72 |
-
## Intended uses & limitations
|
73 |
-
|
74 |
-
More information needed
|
75 |
-
|
76 |
-
## Training and evaluation data
|
77 |
-
|
78 |
-
More information needed
|
79 |
-
|
80 |
## Training procedure
|
81 |
|
82 |
### Training hyperparameters
|
@@ -171,6 +157,31 @@ The following hyperparameters were used during training:
|
|
171 |
| 0.1052 | 9.84 | 35500 | 0.1428 | 0.1247 |
|
172 |
| 0.1044 | 9.98 | 36000 | 0.1430 | 0.1245 |
|
173 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
|
175 |
### Framework versions
|
176 |
|
|
|
11 |
datasets:
|
12 |
- mozilla-foundation/common_voice_9_0
|
13 |
model-index:
|
14 |
+
- name: Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French
|
15 |
results:
|
16 |
- task:
|
17 |
name: Automatic Speech Recognition
|
|
|
55 |
value: 11.09
|
56 |
---
|
57 |
|
|
|
|
|
58 |
|
59 |
+
# Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French
|
60 |
|
61 |
This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - FR dataset.
|
62 |
It achieves the following results on the evaluation set:
|
63 |
- Loss: 0.1430
|
64 |
- Wer: 0.1245
|
65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
## Training procedure
|
67 |
|
68 |
### Training hyperparameters
|
|
|
157 |
| 0.1052 | 9.84 | 35500 | 0.1428 | 0.1247 |
|
158 |
| 0.1044 | 9.98 | 36000 | 0.1430 | 0.1245 |
|
159 |
|
160 |
+
## Evaluation
|
161 |
+
|
162 |
+
1. To evaluate on `mozilla-foundation/common_voice_9_0`
|
163 |
+
|
164 |
+
```bash
|
165 |
+
python eval.py \
|
166 |
+
--model_id "bhuang/wav2vec2-xls-r-1b-french" \
|
167 |
+
--dataset "mozilla-foundation/common_voice_9_0" \
|
168 |
+
--config "fr" \
|
169 |
+
--split "test" \
|
170 |
+
--log_outputs
|
171 |
+
```
|
172 |
+
|
173 |
+
2. To evaluate on `speech-recognition-community-v2/dev_data`
|
174 |
+
|
175 |
+
```bash
|
176 |
+
python eval.py \
|
177 |
+
--model_id "bhuang/wav2vec2-xls-r-1b-french" \
|
178 |
+
--dataset "speech-recognition-community-v2/dev_data" \
|
179 |
+
--config "fr" \
|
180 |
+
--split "validation" \
|
181 |
+
--chunk_length_s 5.0 \
|
182 |
+
--stride_length_s 1.0 \
|
183 |
+
--log_outputs
|
184 |
+
```
|
185 |
|
186 |
### Framework versions
|
187 |
|
eval.py
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
import argparse
|
4 |
+
import re
|
5 |
+
from typing import Dict
|
6 |
+
|
7 |
+
import torch
|
8 |
+
from datasets import Audio, Dataset, load_dataset, load_metric
|
9 |
+
|
10 |
+
from transformers import (
|
11 |
+
AutoConfig,
|
12 |
+
AutoFeatureExtractor,
|
13 |
+
AutoModelForCTC,
|
14 |
+
AutoTokenizer,
|
15 |
+
Wav2Vec2Processor,
|
16 |
+
Wav2Vec2ProcessorWithLM,
|
17 |
+
pipeline,
|
18 |
+
)
|
19 |
+
|
20 |
+
|
21 |
+
def log_results(result: Dataset, args: Dict[str, str]):
|
22 |
+
""" DO NOT CHANGE. This function computes and logs the result metrics. """
|
23 |
+
|
24 |
+
log_outputs = args.log_outputs
|
25 |
+
dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
|
26 |
+
|
27 |
+
# load metric
|
28 |
+
wer = load_metric("wer")
|
29 |
+
cer = load_metric("cer")
|
30 |
+
|
31 |
+
# compute metrics
|
32 |
+
wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
|
33 |
+
cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
|
34 |
+
|
35 |
+
# print & log results
|
36 |
+
result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
|
37 |
+
print(result_str)
|
38 |
+
|
39 |
+
with open(f"{dataset_id}_eval_results.txt", "w") as f:
|
40 |
+
f.write(result_str)
|
41 |
+
|
42 |
+
# log all results in text file. Possibly interesting for analysis
|
43 |
+
if log_outputs is not None:
|
44 |
+
pred_file = f"log_{dataset_id}_predictions.txt"
|
45 |
+
target_file = f"log_{dataset_id}_targets.txt"
|
46 |
+
|
47 |
+
with open(pred_file, "w") as p, open(target_file, "w") as t:
|
48 |
+
|
49 |
+
# mapping function to write output
|
50 |
+
def write_to_file(batch, i):
|
51 |
+
p.write(f"{i}" + "\n")
|
52 |
+
p.write(batch["prediction"] + "\n")
|
53 |
+
t.write(f"{i}" + "\n")
|
54 |
+
t.write(batch["target"] + "\n")
|
55 |
+
|
56 |
+
result.map(write_to_file, with_indices=True)
|
57 |
+
|
58 |
+
|
59 |
+
def normalize_text(text: str, invalid_chars_regex: str) -> str:
|
60 |
+
""" DO ADAPT FOR YOUR USE CASE. this function normalizes the target text. """
|
61 |
+
|
62 |
+
text = text.lower()
|
63 |
+
text = re.sub(r"’", "'", text)
|
64 |
+
text = re.sub(invalid_chars_regex, " ", text)
|
65 |
+
text = re.sub(r"\s+", " ", text).strip()
|
66 |
+
|
67 |
+
return text
|
68 |
+
|
69 |
+
|
70 |
+
def main(args):
|
71 |
+
# load dataset
|
72 |
+
dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
|
73 |
+
|
74 |
+
# for testing: only process the first two examples as a test
|
75 |
+
# dataset = dataset.select(range(10))
|
76 |
+
|
77 |
+
# load processor
|
78 |
+
if args.greedy:
|
79 |
+
processor = Wav2Vec2Processor.from_pretrained(args.model_id)
|
80 |
+
decoder = None
|
81 |
+
else:
|
82 |
+
processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
|
83 |
+
decoder = processor.decoder
|
84 |
+
|
85 |
+
feature_extractor = processor.feature_extractor
|
86 |
+
tokenizer = processor.tokenizer
|
87 |
+
sampling_rate = feature_extractor.sampling_rate
|
88 |
+
|
89 |
+
# resample audio
|
90 |
+
dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
|
91 |
+
|
92 |
+
# load eval pipeline
|
93 |
+
if args.device is None:
|
94 |
+
args.device = 0 if torch.cuda.is_available() else -1
|
95 |
+
|
96 |
+
config = AutoConfig.from_pretrained(args.model_id)
|
97 |
+
model = AutoModelForCTC.from_pretrained(args.model_id)
|
98 |
+
|
99 |
+
# asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
|
100 |
+
asr = pipeline(
|
101 |
+
"automatic-speech-recognition",
|
102 |
+
config=config,
|
103 |
+
model=model,
|
104 |
+
tokenizer=tokenizer,
|
105 |
+
feature_extractor=feature_extractor,
|
106 |
+
decoder=decoder,
|
107 |
+
device=args.device,
|
108 |
+
)
|
109 |
+
|
110 |
+
# build normalizer config
|
111 |
+
tokenizer = AutoTokenizer.from_pretrained(args.model_id)
|
112 |
+
tokens = [x for x in tokenizer.convert_ids_to_tokens(range(0, tokenizer.vocab_size))]
|
113 |
+
special_tokens = [
|
114 |
+
tokenizer.pad_token,
|
115 |
+
tokenizer.word_delimiter_token,
|
116 |
+
tokenizer.unk_token,
|
117 |
+
tokenizer.bos_token,
|
118 |
+
tokenizer.eos_token,
|
119 |
+
]
|
120 |
+
non_special_tokens = [x for x in tokens if x not in special_tokens]
|
121 |
+
invalid_chars_regex = f"[^\s{re.escape(''.join(set(non_special_tokens)))}]"
|
122 |
+
|
123 |
+
# normalize_to_lower = False
|
124 |
+
# for token in non_special_tokens:
|
125 |
+
# if token.isalpha() and token.islower():
|
126 |
+
# normalize_to_lower = True
|
127 |
+
# break
|
128 |
+
|
129 |
+
# map function to decode audio
|
130 |
+
def map_to_pred(batch):
|
131 |
+
prediction = asr(batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s)
|
132 |
+
|
133 |
+
batch["prediction"] = prediction["text"]
|
134 |
+
batch["target"] = normalize_text(batch["sentence"], invalid_chars_regex)
|
135 |
+
return batch
|
136 |
+
|
137 |
+
# run inference on all examples
|
138 |
+
result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
|
139 |
+
|
140 |
+
# filtering out empty targets
|
141 |
+
result = result.filter(lambda example: example["target"] != "")
|
142 |
+
|
143 |
+
# compute and log_results
|
144 |
+
# do not change function below
|
145 |
+
log_results(result, args)
|
146 |
+
|
147 |
+
|
148 |
+
if __name__ == "__main__":
|
149 |
+
parser = argparse.ArgumentParser()
|
150 |
+
|
151 |
+
parser.add_argument("--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers")
|
152 |
+
parser.add_argument(
|
153 |
+
"--dataset",
|
154 |
+
type=str,
|
155 |
+
required=True,
|
156 |
+
help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
|
157 |
+
)
|
158 |
+
parser.add_argument("--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice")
|
159 |
+
parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
|
160 |
+
parser.add_argument(
|
161 |
+
"--chunk_length_s",
|
162 |
+
type=float,
|
163 |
+
default=None,
|
164 |
+
help="Chunk length in seconds. Defaults to None. For long audio files a good value would be 5.0 seconds.",
|
165 |
+
)
|
166 |
+
parser.add_argument(
|
167 |
+
"--stride_length_s",
|
168 |
+
type=float,
|
169 |
+
default=None,
|
170 |
+
help="Stride of the audio chunks. Defaults to None. For long audio files a good value would be 1.0 seconds.",
|
171 |
+
)
|
172 |
+
parser.add_argument("--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis.")
|
173 |
+
parser.add_argument("--greedy", action="store_true", help="If defined, the LM will be ignored during inference.")
|
174 |
+
parser.add_argument(
|
175 |
+
"--device",
|
176 |
+
type=int,
|
177 |
+
default=None,
|
178 |
+
help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
|
179 |
+
)
|
180 |
+
args = parser.parse_args()
|
181 |
+
|
182 |
+
main(args)
|