add code for using kenlm

eval_lm.py

@@ -0,0 +1,229 @@
+#!/usr/bin/env python3
+import argparse
+import functools
+import re
+import string
+import unidecode
+from typing import Dict
+
+from datasets import Audio, Dataset, DatasetDict, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, AutoTokenizer, 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_regex = f'[{re.escape(string.punctuation)}]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(
+        chars_to_ignore_regex, 
+        "", 
+        re.sub("['`´]", "’",   # elsewhere probably meant as glottal stop
+               re.sub("([og])['`´]", "\g<1>‘",  # after o/g indicate modified char
+                      unidecode.unidecode(text).lower()
+                     )
+              )
+    ) + " "
+
+    # 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 create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token = None,
+    unk_token = None,
+    pad_token = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["test"].column_names,
+    )
+
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(vocabs["test"]["vocab"][0]))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def asr_pipeline_lm(model_id):
+    from transformers.models.auto.modeling_auto import AutoModelForCTC
+    from transformers.models.auto.processing_auto import AutoProcessor
+    import torch
+
+    model = AutoModelForCTC.from_pretrained(model_id)
+    processor = AutoProcessor.from_pretrained(model_id)
+
+    def asr(audio_array, sampling_rate):
+        input_values = processor(
+            audio_array, 
+            return_tensors="pt", 
+            sampling_rate=sampling_rate
+        ).input_values
+        with torch.no_grad():
+            logits = model(input_values).logits
+        return processor.batch_decode(logits.numpy()).text
+
+    return asr
+
+
+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))
+
+    if args.use_lm:
+        asr = asr_pipeline_lm(args.model_id)
+
+        def map_to_pred(batch):
+            prediction = asr(batch["audio"]["array"])
+            batch["prediction"] = prediction[0]
+            batch["target"] = normalize_text(batch["sentence"])
+            return batch
+
+    else:
+        # 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 args.check_vocab:
+        tokenizer = AutoTokenizer.from_pretrained(args.model_id)
+        unk_token = "[UNK]"
+        pad_token = "[PAD]"
+        word_delimiter_token = "|"
+        raw_datasets = DatasetDict({"test": result})
+        vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+        print(vocab_dict)
+        print("OOV chars:", set(vocab_dict) - set(tokenizer.get_vocab()))    
+
+        
+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(
+        "--check_vocab", action="store_true", help="Verify that normalized target text is within character set"
+    )
+    parser.add_argument(
+        "--use-lm", action="store_true", help="Use kenlm decoder"
+    )
+    args = parser.parse_args()
+
+    main(args)