Add language model

.gitignore

@@ -0,0 +1 @@
+checkpoint-*/

.ipynb_checkpoints/README-checkpoint.md

@@ -0,0 +1,90 @@
+---
+license: apache-2.0
+tags:
+- generated_from_trainer
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_7_0
+- fi
+- robust-speech-event
+datasets:
+- mozilla-foundation/common_voice_7_0
+model-index:
+- name: wav2vec2-xlsr-1b-finnish
+  results:
+
+
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# wav2vec2-xlsr-1b-finnish-lm
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 5e-05
+- train_batch_size: 32
+- eval_batch_size: 8
+- seed: 42
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 500
+- num_epochs: 5
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|
+| 0.968         | 0.18  | 500   | 0.4870          | 0.4720 |
+| 0.6557        | 0.36  | 1000  | 0.2450          | 0.2931 |
+| 0.647         | 0.54  | 1500  | 0.1818          | 0.2255 |
+| 0.5297        | 0.72  | 2000  | 0.1698          | 0.2354 |
+| 0.5802        | 0.9   | 2500  | 0.1581          | 0.2355 |
+| 0.6351        | 1.07  | 3000  | 0.1689          | 0.2336 |
+| 0.4626        | 1.25  | 3500  | 0.1719          | 0.3099 |
+| 0.4526        | 1.43  | 4000  | 0.1434          | 0.2069 |
+| 0.4692        | 1.61  | 4500  | 0.1645          | 0.2192 |
+| 0.4584        | 1.79  | 5000  | 0.1483          | 0.1987 |
+| 0.4234        | 1.97  | 5500  | 0.1499          | 0.2178 |
+| 0.4243        | 2.15  | 6000  | 0.1345          | 0.2070 |
+| 0.4108        | 2.33  | 6500  | 0.1383          | 0.1850 |
+| 0.4048        | 2.51  | 7000  | 0.1338          | 0.1811 |
+| 0.4085        | 2.69  | 7500  | 0.1290          | 0.1780 |
+| 0.4026        | 2.87  | 8000  | 0.1239          | 0.1650 |
+| 0.4033        | 3.04  | 8500  | 0.1346          | 0.1657 |
+| 0.3986        | 3.22  | 9000  | 0.1310          | 0.1850 |
+| 0.3867        | 3.4   | 9500  | 0.1273          | 0.1741 |
+| 0.3658        | 3.58  | 10000 | 0.1219          | 0.1672 |
+| 0.382         | 3.76  | 10500 | 0.1306          | 0.1698 |
+| 0.3847        | 3.94  | 11000 | 0.1230          | 0.1577 |
+| 0.3691        | 4.12  | 11500 | 0.1310          | 0.1615 |
+| 0.3593        | 4.3   | 12000 | 0.1296          | 0.1622 |
+| 0.3619        | 4.48  | 12500 | 0.1285          | 0.1601 |
+| 0.3361        | 4.66  | 13000 | 0.1261          | 0.1569 |
+| 0.3603        | 4.84  | 13500 | 0.1235          | 0.1533 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.11.0

README.md

@@ -1,3 +1,90 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+tags:
+- generated_from_trainer
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_7_0
+- fi
+- robust-speech-event
+datasets:
+- mozilla-foundation/common_voice_7_0
+model-index:
+- name: wav2vec2-xlsr-1b-finnish
+  results:
+
+
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# wav2vec2-xlsr-1b-finnish-lm
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b)
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 5e-05
+- train_batch_size: 32
+- eval_batch_size: 8
+- seed: 42
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 500
+- num_epochs: 5
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|
+| 0.968         | 0.18  | 500   | 0.4870          | 0.4720 |
+| 0.6557        | 0.36  | 1000  | 0.2450          | 0.2931 |
+| 0.647         | 0.54  | 1500  | 0.1818          | 0.2255 |
+| 0.5297        | 0.72  | 2000  | 0.1698          | 0.2354 |
+| 0.5802        | 0.9   | 2500  | 0.1581          | 0.2355 |
+| 0.6351        | 1.07  | 3000  | 0.1689          | 0.2336 |
+| 0.4626        | 1.25  | 3500  | 0.1719          | 0.3099 |
+| 0.4526        | 1.43  | 4000  | 0.1434          | 0.2069 |
+| 0.4692        | 1.61  | 4500  | 0.1645          | 0.2192 |
+| 0.4584        | 1.79  | 5000  | 0.1483          | 0.1987 |
+| 0.4234        | 1.97  | 5500  | 0.1499          | 0.2178 |
+| 0.4243        | 2.15  | 6000  | 0.1345          | 0.2070 |
+| 0.4108        | 2.33  | 6500  | 0.1383          | 0.1850 |
+| 0.4048        | 2.51  | 7000  | 0.1338          | 0.1811 |
+| 0.4085        | 2.69  | 7500  | 0.1290          | 0.1780 |
+| 0.4026        | 2.87  | 8000  | 0.1239          | 0.1650 |
+| 0.4033        | 3.04  | 8500  | 0.1346          | 0.1657 |
+| 0.3986        | 3.22  | 9000  | 0.1310          | 0.1850 |
+| 0.3867        | 3.4   | 9500  | 0.1273          | 0.1741 |
+| 0.3658        | 3.58  | 10000 | 0.1219          | 0.1672 |
+| 0.382         | 3.76  | 10500 | 0.1306          | 0.1698 |
+| 0.3847        | 3.94  | 11000 | 0.1230          | 0.1577 |
+| 0.3691        | 4.12  | 11500 | 0.1310          | 0.1615 |
+| 0.3593        | 4.3   | 12000 | 0.1296          | 0.1622 |
+| 0.3619        | 4.48  | 12500 | 0.1285          | 0.1601 |
+| 0.3361        | 4.66  | 13000 | 0.1261          | 0.1569 |
+| 0.3603        | 4.84  | 13500 | 0.1235          | 0.1533 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.11.0

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 33, "</s>": 34}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "'", "-", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e4", "\u00e5", "\u00f6", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

config.json

@@ -0,0 +1,108 @@
+{
+  "_name_or_path": "facebook/wav2vec2-xls-r-1b",
+  "activation_dropout": 0.055,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.094,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 1024,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": false,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.04,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.047,
+  "hidden_size": 1280,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.041,
+  "mask_feature_length": 10,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.082,
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 48,
+  "num_negatives": 100,
+  "output_hidden_size": 1280,
+  "pad_token_id": 32,
+  "proj_codevector_dim": 1024,
+  "tdnn_dilation": [
+    1,
+    2,
+    3,
+    1,
+    1
+  ],
+  "tdnn_dim": [
+    512,
+    512,
+    512,
+    512,
+    1500
+  ],
+  "tdnn_kernel": [
+    5,
+    3,
+    3,
+    1,
+    1
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.17.0.dev0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 35,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
+                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
+                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
+                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
+                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+    
+    chars_to_remove_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
+
+    text = re.sub(chars_to_remove_regex, "", text.lower())
+    text = re.sub("[-]", " ", text)
+
+    # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # note that order is important here!
+    token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    for t in token_sequences_to_ignore:
+        text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

language_model/.ipynb_checkpoints/attrs-checkpoint.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}

language_model/attrs.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}

language_model/kenlm_fi_model_correct.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14c745cf96cd5dc073bcd23e2968e4b5059af1154b4f4ee029725a136fd9a929
+size 16497522

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000,
+  "processor_class": "Wav2Vec2ProcessorWithLM"
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dce0dc75b7f73dfa3553f32ae2f994ab2ced33b6206bf23f1299e331d13c1d57
+size 3850492081

run_eval.sh

@@ -0,0 +1 @@
+python3 eval.py --dataset mozilla-foundation/common_voice_7_0 --config fi --model_id aapot/wav2vec2-xlsr-1b-finnish-lm --split test --log_outputs

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57e19b60b0278b22953f1ede3366cbebb37b013d017425e23165dba933554bf6
+size 3055

vocab.json

@@ -0,0 +1 @@
+{"'": 1, "a": 2, "b": 3, "c": 4, "d": 5, "e": 6, "f": 7, "g": 8, "h": 9, "i": 10, "j": 11, "k": 12, "l": 13, "m": 14, "n": 15, "o": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "u": 22, "v": 23, "w": 24, "x": 25, "y": 26, "z": 27, "ä": 28, "å": 29, "ö": 30, "|": 0, "[UNK]": 31, "[PAD]": 32}