add LM preprocessing

.ipynb_checkpoints/README-checkpoint.md

@@ -0,0 +1,86 @@
+---
+language: 
+- id
+
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- robust-speech-event
+datasets:
+- mozilla-foundation/common_voice_7_0
+metrics:
+- wer
+- cer
+model-index:
+- name: wav2vec2-large-xls-r-300m-Indonesian
+  results:
+  - task: 
+      type: automatic-speech-recognition  # Required. Example: automatic-speech-recognition
+      name: Speech Recognition  # Optional. Example: Speech Recognition
+    dataset:
+      type: mozilla-foundation/common_voice_7_0  # Required. Example: common_voice. Use dataset id from https://hf.co/datasets
+      name: Common Voice id  # Required. Example: Common Voice zh-CN
+      args: id         # Optional. Example: zh-CN
+    metrics:
+      - type: wer    # Required. Example: wer
+        value: 25.06 # Required. Example: 20.90
+        name: Test WER    # Optional. Example: Test WER
+    
+      - type: cer    # Required. Example: wer
+        value: 6.50  # Required. Example: 20.90
+        name: Test CER    # Optional. Example: Test WER
+     
+
+---
+
+<!-- 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-large-xls-r-300m-Indonesian
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the common_voice dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.4087
+- Wer: 0.2461
+- Cer: 0.0666
+
+
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0003
+- train_batch_size: 64
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 2
+- total_train_batch_size: 128
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 400
+- num_epochs: 50
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Wer    | Cer    |
+|:-------------:|:-----:|:----:|:---------------:|:------:|:------:|
+| 5.0788        | 4.26  | 200  | 2.9389          | 1.0    | 1.0    |
+| 2.8288        | 8.51  | 400  | 2.2535          | 1.0    | 0.8004 |
+| 0.907         | 12.77 | 600  | 0.4558          | 0.4243 | 0.1095 |
+| 0.4071        | 17.02 | 800  | 0.4013          | 0.3468 | 0.0913 |
+| 0.3           | 21.28 | 1000 | 0.4167          | 0.3075 | 0.0816 |
+| 0.2544        | 25.53 | 1200 | 0.4132          | 0.2835 | 0.0762 |
+| 0.2145        | 29.79 | 1400 | 0.3878          | 0.2693 | 0.0729 |
+| 0.1923        | 34.04 | 1600 | 0.4023          | 0.2623 | 0.0702 |
+| 0.1681        | 38.3  | 1800 | 0.3984          | 0.2581 | 0.0686 |
+| 0.1598        | 42.55 | 2000 | 0.3982          | 0.2493 | 0.0663 |
+| 0.1464        | 46.81 | 2200 | 0.4087          | 0.2461 | 0.0666 |
+
+
+### Framework versions
+
+- Transformers 4.17.0.dev0
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.2.dev0
+- Tokenizers 0.11.0

.ipynb_checkpoints/alphabet-checkpoint.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", "\u00e9", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

README.md

@@ -24,11 +24,11 @@ model-index:
     metrics:
       - type: wer    # Required. Example: wer
         value: 25.06 # Required. Example: 20.90
-        name: Test WER    # Optional. Example: Test WER
+        name: Test WER With LM   # Optional. Example: Test WER
     
       - type: cer    # Required. Example: wer
         value: 6.50  # Required. Example: 20.90
-        name: Test CER    # Optional. Example: Test WER
+        name: Test CER With LM   # Optional. Example: Test WER
      
 
 ---

eval.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+from datasets import load_dataset, load_metric, Audio, Dataset
+from transformers import pipeline, AutoFeatureExtractor
+import re
+import argparse
+import unicodedata
+from typing import Dict
+
+
+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_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\\\\&\—\！\，\’\(\)]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+    
+    text = text.lower()
+    # normalize non-standard (stylized) unicode characters
+    text = unicodedata.normalize('NFKC', text)
+
+    # Let's also make sure we split on all kinds of newlines, spaces, etc...
+    text = " ".join(text.split())
+
+    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
+    asr = pipeline("automatic-speech-recognition", model=args.model_id)
+
+    # 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 None. For long audio files a good value would be 5.0 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to None. For long audio files a good value would be 1.0 seconds."
+    )
+    parser.add_argument(
+        "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
+    )
+    args = parser.parse_args()
+
+    main(args)