geninhu commited on
Commit
556d907
β€’
1 Parent(s): 12f2205

Training in progress, step 16500

Browse files
.ipynb_checkpoints/eval-checkpoint.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ from datasets import Audio, Dataset, load_dataset, load_metric
7
+
8
+ from transformers import AutoFeatureExtractor, pipeline
9
+
10
+
11
+ def log_results(result: Dataset, args: Dict[str, str]):
12
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
13
+
14
+ log_outputs = args.log_outputs
15
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
16
+
17
+ # load metric
18
+ wer = load_metric("wer")
19
+ cer = load_metric("cer")
20
+
21
+ # compute metrics
22
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
23
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
24
+
25
+ # print & log results
26
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
27
+ print(result_str)
28
+
29
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
30
+ f.write(result_str)
31
+
32
+ # log all results in text file. Possibly interesting for analysis
33
+ if log_outputs is not None:
34
+ pred_file = f"log_{dataset_id}_predictions.txt"
35
+ target_file = f"log_{dataset_id}_targets.txt"
36
+
37
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
38
+
39
+ # mapping function to write output
40
+ def write_to_file(batch, i):
41
+ p.write(f"{i}" + "\n")
42
+ p.write(batch["prediction"] + "\n")
43
+ t.write(f"{i}" + "\n")
44
+ t.write(batch["target"] + "\n")
45
+
46
+ result.map(write_to_file, with_indices=True)
47
+
48
+
49
+ def normalize_text(text: str) -> str:
50
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
51
+
52
+ chars_to_ignore_regex = '[,?.!\-\;\:"β€œ%β€˜β€οΏ½β€”β€™β€¦β€“|]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
53
+
54
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
55
+
56
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
57
+ # note that order is important here!
58
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
59
+
60
+ for t in token_sequences_to_ignore:
61
+ text = " ".join(text.split(t))
62
+
63
+ return text
64
+
65
+
66
+ def main(args):
67
+ # load dataset
68
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
69
+
70
+ # for testing: only process the first two examples as a test
71
+ dataset = dataset.select(range(10))
72
+
73
+ # load processor
74
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
75
+ sampling_rate = feature_extractor.sampling_rate
76
+
77
+ # resample audio
78
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
79
+
80
+ # load eval pipeline
81
+ asr = pipeline("automatic-speech-recognition", model=args.model_id)
82
+
83
+ # map function to decode audio
84
+ def map_to_pred(batch):
85
+ prediction = asr(
86
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
87
+ )
88
+
89
+ batch["prediction"] = prediction["text"]
90
+ batch["target"] = normalize_text(batch["sentence"])
91
+ return batch
92
+
93
+ # run inference on all examples
94
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
95
+
96
+ # compute and log_results
97
+ # do not change function below
98
+ log_results(result, args)
99
+
100
+
101
+ if __name__ == "__main__":
102
+ parser = argparse.ArgumentParser()
103
+
104
+ parser.add_argument(
105
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with πŸ€— Transformers"
106
+ )
107
+ parser.add_argument(
108
+ "--dataset",
109
+ type=str,
110
+ required=True,
111
+ help="Dataset name to evaluate the `model_id`. Should be loadable with πŸ€— Datasets",
112
+ )
113
+ parser.add_argument(
114
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
115
+ )
116
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
117
+ parser.add_argument(
118
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
119
+ )
120
+ parser.add_argument(
121
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
122
+ )
123
+ parser.add_argument(
124
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
125
+ )
126
+ args = parser.parse_args()
127
+
128
+ main(args)
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b6d43a05236873c191a96ed3ff32aae1dd8d89f628dd55fd89bae680dcb26e91
3
  size 1264777329
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7562a5402132f6e0c7c8730ecf3e3d51fe8af1b0cb0bab26bdb438cb1e908708
3
  size 1264777329
runs/Jan26_00-02-24_job-fa775f5b-8438-4c3d-95fa-ed70ddaee577/events.out.tfevents.1643155382.job-fa775f5b-8438-4c3d-95fa-ed70ddaee577.14643.0 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:08d3678f679cc384b45ad08f6127111b80f571a47bb7a55a2faf62a828eddf4f
3
- size 9471
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2db8320f28dab2600fa328edc5fd31dacb58f2ca65b458a5ce6a7c8e67da81d2
3
+ size 9955