lucio commited on
Commit
6b9b55b
1 Parent(s): 79b8455

Training in progress, step 17500

Browse files
.ipynb_checkpoints/README-checkpoint.md ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - ug
4
+ license: apache-2.0
5
+ tags:
6
+ - automatic-speech-recognition
7
+ - mozilla-foundation/common_voice_8_0
8
+ - generated_from_trainer
9
+ datasets:
10
+ - common_voice
11
+ model-index:
12
+ - name: xls-r-uyghur-cv8
13
+ results: []
14
+ ---
15
+
16
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
17
+ should probably proofread and complete it, then remove this comment. -->
18
+
19
+ # xls-r-uyghur-cv8
20
+
21
+ This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - UG dataset.
22
+ It achieves the following results on the evaluation set:
23
+ - Loss: 0.2430
24
+ - Wer: 0.3804
25
+
26
+ ## Model description
27
+
28
+ More information needed
29
+
30
+ ## Intended uses & limitations
31
+
32
+ More information needed
33
+
34
+ ## Training and evaluation data
35
+
36
+ More information needed
37
+
38
+ ## Training procedure
39
+
40
+ ### Training hyperparameters
41
+
42
+ The following hyperparameters were used during training:
43
+ - learning_rate: 7.5e-05
44
+ - train_batch_size: 8
45
+ - eval_batch_size: 8
46
+ - seed: 42
47
+ - gradient_accumulation_steps: 4
48
+ - total_train_batch_size: 32
49
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
50
+ - lr_scheduler_type: linear
51
+ - lr_scheduler_warmup_steps: 2000
52
+ - num_epochs: 50.0
53
+ - mixed_precision_training: Native AMP
54
+
55
+ ### Training results
56
+
57
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
58
+ |:-------------:|:-----:|:----:|:---------------:|:------:|
59
+ | 3.6871 | 2.66 | 500 | 3.5374 | 1.0 |
60
+ | 3.1501 | 5.32 | 1000 | 3.1278 | 1.0 |
61
+ | 1.5843 | 7.97 | 1500 | 0.6358 | 0.6914 |
62
+ | 1.3378 | 10.64 | 2000 | 0.4422 | 0.5925 |
63
+ | 1.2595 | 13.3 | 2500 | 0.3921 | 0.5512 |
64
+ | 1.1643 | 15.95 | 3000 | 0.3507 | 0.5149 |
65
+ | 1.1352 | 18.61 | 3500 | 0.3351 | 0.5019 |
66
+ | 1.1113 | 21.28 | 4000 | 0.3153 | 0.4845 |
67
+ | 1.0914 | 23.93 | 4500 | 0.3050 | 0.4594 |
68
+ | 1.0468 | 26.59 | 5000 | 0.2890 | 0.4470 |
69
+ | 1.0473 | 29.25 | 5500 | 0.2755 | 0.4331 |
70
+ | 1.0065 | 31.91 | 6000 | 0.2718 | 0.4264 |
71
+ | 0.9794 | 34.57 | 6500 | 0.2646 | 0.4193 |
72
+ | 0.9849 | 37.23 | 7000 | 0.2610 | 0.4058 |
73
+ | 0.9496 | 39.89 | 7500 | 0.2522 | 0.3985 |
74
+ | 0.9367 | 42.55 | 8000 | 0.2514 | 0.3947 |
75
+ | 0.9295 | 45.21 | 8500 | 0.2458 | 0.3883 |
76
+ | 0.9187 | 47.87 | 9000 | 0.2439 | 0.3833 |
77
+
78
+
79
+ ### Framework versions
80
+
81
+ - Transformers 4.16.0.dev0
82
+ - Pytorch 1.10.1+cu102
83
+ - Datasets 1.18.2.dev0
84
+ - Tokenizers 0.11.0
.ipynb_checkpoints/added_tokens-checkpoint.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 42, "</s>": 43}
.ipynb_checkpoints/config-checkpoint.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 768,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1024,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 4096,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 24,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1024,
79
+ "pad_token_id": 41,
80
+ "proj_codevector_dim": 768,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.16.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 44,
106
+ "xvector_output_dim": 512
107
+ }
.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 = '[,?.!\-\;\:"“%‘”�—’…–]' # 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 🤗 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 🤗 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)
eval.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 = '[,?.!\-\;\:"“%‘”�—’…–]' # 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 🤗 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 🤗 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 sha256:6dfa20b9b180f61f11e224e9967ddacc1eb011986fa6d4ad303c97df6fcb1fa9
3
  size 1262104049
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3242dcd18b4783f24ee583f30a802f89c935921e0f0c840e9d9acd8c0ce65b9f
3
  size 1262104049
runs/Jan28_17-40-32_job-0074bb36-c67f-4775-b1b6-176eb09b0ba4/events.out.tfevents.1643391738.job-0074bb36-c67f-4775-b1b6-176eb09b0ba4.843454.0 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:190eb3efeee16b3e15d0bad09c0e6f494d9afe9f51a60280e1f4a24e91abba0b
3
- size 42262
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d0129fac457fac11ef82bb6a4b55ec2d16bbc10dd418883423335abda61ee12d
3
+ size 43386
vocab-Copy1.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"،": 1, "؛": 2, "؟": 3, "ئ": 4, "ا": 5, "ب": 6, "ت": 7, "ج": 8, "خ": 9, "د": 10, "ر": 11, "ز": 12, "س": 13, "ش": 14, "غ": 15, "ف": 16, "ق": 17, "ك": 18, "ل": 19, "م": 20, "ن": 21, "و": 22, "ى": 23, "ي": 24, "پ": 25, "چ": 26, "ژ": 27, "ڭ": 28, "گ": 29, "ھ": 30, "ۆ": 31, "ۇ": 32, "ۈ": 33, "ۋ": 34, "ې": 35, "ە": 36, "‹": 37, "›": 38, "−": 39, "|": 0, "[UNK]": 40, "[PAD]": 41}