baguette

added_tokens.json

@@ -1 +0,0 @@
-{"<s>": 216, "</s>": 217}

eval.py

@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+import argparse
+import re
+import unicodedata
+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)
+
+
+chars_to_remove_regex = r'[\,\?\.\!\-\_\;\:\"\“\%\‘\”\�\^]'
+
+def remove_accents(text):
+    nfkd_form = unicodedata.normalize('NFKD', text)
+    return u"".join([c for c in nfkd_form if not unicodedata.combining(c)])
+
+def remove_special_characters(text):
+    text = re.sub(chars_to_remove_regex, '', text).lower()
+    text = re.sub("ç", r'[cedille]', text)
+    text = re.sub("&", r'et', text)    
+    text = re.sub("%", r' pourcents', text)   
+    text = re.sub("([0-9]+)(,|.)([0-9+])", r'\1 virgule \3', text) 
+    text = re.sub("\$", r'dollar', text)
+    text = re.sub("\£", r'livre', text)
+    text = re.sub("\€", r'euro', text)
+    text = remove_accents(text)
+    text = re.sub(r"\[cedille\]", 'ç', text) + " "
+    return text
+    
+def normalize_text(text: str) -> str:
+    text = remove_special_characters(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(20))
+
+    # 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,,skip_special_tokens=True
+        )
+
+        batch["prediction"] = prediction["text"]# "".join(prediction["text"].split("<s>"))
+        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)

run.sh

@@ -6,8 +6,8 @@ python run_speech_recognition_ctc.py \
 --output_dir="./" \
 --overwrite_output_dir \
 --num_train_epochs="5" \
---per_device_train_batch_size="32" \
---per_device_eval_batch_size="32" \
+--per_device_train_batch_size="64" \
+--per_device_eval_batch_size="64" \
 --gradient_accumulation_steps="1" \
 --learning_rate="7e-5" \
 --warmup_steps="1500" \
@@ -22,9 +22,9 @@ python run_speech_recognition_ctc.py \
 --save_total_limit="3" \
 --freeze_feature_encoder \
 --feat_proj_dropout="0.0" \
---mask_time_prob="0.15" \
+--mask_time_prob="0.05" \
 --mask_time_length="10" \
---mask_feature_prob="0.3" \
+--mask_feature_prob="0.33" \
 --mask_feature_length="10" \
 --gradient_checkpointing \
 --report_to="wandb" \

run_speech_recognition_ctc.py

@@ -511,6 +511,8 @@ def main():
         tokenizer_kwargs = {
             "config": config if config.tokenizer_class is not None else None,
             "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "bos_token": None,
+            "eos_token": None,
             "unk_token": unk_token,
             "pad_token": pad_token,
             "word_delimiter_token": word_delimiter_token,
@@ -641,7 +643,7 @@ def main():
 
         pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
 
-        pred_str = tokenizer.batch_decode(pred_ids)
+        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
         # we do not want to group tokens when computing the metrics
         label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)