add eval script

README.md

@@ -11,7 +11,7 @@ tags:
 datasets:
 - mozilla-foundation/common_voice_9_0
 model-index:
-- name: wav2vec2-xls-r-1b-ft-cv9-fr
+- name: Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French
   results:
   - task:
       name: Automatic Speech Recognition
@@ -55,28 +55,14 @@ model-index:
       value: 11.09
 ---
 
-<!-- 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-xls-r-1b-ft-cv9-fr
+# Fine-tuned Wav2Vec2 XLS-R 1B model for ASR in French
 
 This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - FR dataset.
 It achieves the following results on the evaluation set:
 - Loss: 0.1430
 - Wer: 0.1245
 
-## Model description
-
-More information needed
-
-## Intended uses & limitations
-
-More information needed
-
-## Training and evaluation data
-
-More information needed
-
 ## Training procedure
 
 ### Training hyperparameters
@@ -171,6 +157,31 @@ The following hyperparameters were used during training:
 | 0.1052        | 9.84  | 35500 | 0.1428          | 0.1247 |
 | 0.1044        | 9.98  | 36000 | 0.1430          | 0.1245 |
 
+## Evaluation
+
+1. To evaluate on `mozilla-foundation/common_voice_9_0`
+
+```bash
+python eval.py \
+  --model_id "bhuang/wav2vec2-xls-r-1b-french" \
+  --dataset "mozilla-foundation/common_voice_9_0" \
+  --config "fr" \
+  --split "test" \
+  --log_outputs
+```
+
+2. To evaluate on `speech-recognition-community-v2/dev_data`
+
+```bash
+python eval.py \
+  --model_id "bhuang/wav2vec2-xls-r-1b-french" \
+  --dataset "speech-recognition-community-v2/dev_data" \
+  --config "fr" \
+  --split "validation" \
+  --chunk_length_s 5.0 \
+  --stride_length_s 1.0 \
+  --log_outputs
+```
 
 ### Framework versions
 

eval.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python
+
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoTokenizer,
+    Wav2Vec2Processor,
+    Wav2Vec2ProcessorWithLM,
+    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, invalid_chars_regex: str) -> str:
+    """ DO ADAPT FOR YOUR USE CASE. this function normalizes the target text. """
+
+    text = text.lower()
+    text = re.sub(r"’", "'", text)
+    text = re.sub(invalid_chars_regex, " ", text)
+    text = re.sub(r"\s+", " ", text).strip()
+
+    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
+    if args.greedy:
+        processor = Wav2Vec2Processor.from_pretrained(args.model_id)
+        decoder = None
+    else:
+        processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
+        decoder = processor.decoder
+
+    feature_extractor = processor.feature_extractor
+    tokenizer = processor.tokenizer
+    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
+
+    config = AutoConfig.from_pretrained(args.model_id)
+    model = AutoModelForCTC.from_pretrained(args.model_id)
+
+    # asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+    asr = pipeline(
+        "automatic-speech-recognition",
+        config=config,
+        model=model,
+        tokenizer=tokenizer,
+        feature_extractor=feature_extractor,
+        decoder=decoder,
+        device=args.device,
+    )
+
+    # build normalizer config
+    tokenizer = AutoTokenizer.from_pretrained(args.model_id)
+    tokens = [x for x in tokenizer.convert_ids_to_tokens(range(0, tokenizer.vocab_size))]
+    special_tokens = [
+        tokenizer.pad_token,
+        tokenizer.word_delimiter_token,
+        tokenizer.unk_token,
+        tokenizer.bos_token,
+        tokenizer.eos_token,
+    ]
+    non_special_tokens = [x for x in tokens if x not in special_tokens]
+    invalid_chars_regex = f"[^\s{re.escape(''.join(set(non_special_tokens)))}]"
+
+    # normalize_to_lower = False
+    # for token in non_special_tokens:
+    #     if token.isalpha() and token.islower():
+    #         normalize_to_lower = True
+    #         break
+
+    # 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"], invalid_chars_regex)
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # filtering out empty targets
+    result = result.filter(lambda example: example["target"] != "")
+
+    # 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.")
+    parser.add_argument("--greedy", action="store_true", help="If defined, the LM will be ignored during inference.")
+    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)