add eval script

.ipynb_checkpoints/README-checkpoint.md

@@ -1,12 +1,32 @@
 ---
+language:
+- ar
 license: apache-2.0
 tags:
+- automatic-speech-recognition
+- common_voice
 - generated_from_trainer
+- sv
+- robust-speech-event
 datasets:
 - common_voice
 model-index:
 - name: wav2vec2-xls-r-300m-ar
-  results: []
+  results:
+    - task: 
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: Robust Speech Event - Dev Data
+      type: speech-recognition-community-v2/dev_data
+      args: ar
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 1.0
+       - name: Test CER
+         type: cer
+         value: 1.0
 ---
 
 <!-- This model card has been generated automatically according to the information the Trainer had access to. You
@@ -14,7 +34,15 @@ should probably proofread and complete it, then remove this comment. -->
 
 # wav2vec2-xls-r-300m-ar
 
-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.
+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 - AR dataset.
+It achieves the following results on the evaluation set:
+- eval_loss: 3.0191
+- eval_wer: 1.0
+- eval_runtime: 252.2389
+- eval_samples_per_second: 30.217
+- eval_steps_per_second: 0.476
+- epoch: 1.0
+- step: 340
 
 ## Model description
 
@@ -49,3 +77,13 @@ The following hyperparameters were used during training:
 - Pytorch 1.10.2+cu102
 - Datasets 1.18.2.dev0
 - Tokenizers 0.11.0
+
+#### Evaluation Commands
+
+Please use the evaluation script `eval.py` included in the repo.
+
+1. To evaluate on `speech-recognition-community-v2/dev_data`
+
+```bash
+python eval.py --model_id nouamanetazi/wav2vec2-xls-r-300m-ar --dataset speech-recognition-community-v2/dev_data --config ar --split validation --chunk_length_s 5.0 --stride_length_s 1.0
+```

.ipynb_checkpoints/eval-checkpoint.py

@@ -0,0 +1,141 @@
+#!/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)
+
+
+# Normalize arabic
+def normalizeArabic(text):
+    # https://alraqmiyyat.github.io/2013/01-02.html
+    text = re.sub("[إأٱآا]", "ا", text)
+    text = re.sub("ى", "ي", text)
+    text = re.sub("ؤ", "ء", text)
+    text = re.sub("ئ", "ء", text)
+
+    # keep only characters which unicode \u0600-\u06FF and space
+    text = re.sub(r"[^\u0600-\u06FF ]", "", text)
+    return text
+
+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 = re.sub(chars_to_ignore_regex, "", 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))
+
+    text = normalizeArabic(text)
+
+    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)

README.md

@@ -6,11 +6,27 @@ tags:
 - automatic-speech-recognition
 - common_voice
 - generated_from_trainer
+- sv
+- robust-speech-event
 datasets:
 - common_voice
 model-index:
 - name: wav2vec2-xls-r-300m-ar
-  results: []
+  results:
+    - task: 
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: Robust Speech Event - Dev Data
+      type: speech-recognition-community-v2/dev_data
+      args: ar
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 1.0
+       - name: Test CER
+         type: cer
+         value: 1.0
 ---
 
 <!-- This model card has been generated automatically according to the information the Trainer had access to. You
@@ -61,3 +77,13 @@ The following hyperparameters were used during training:
 - Pytorch 1.10.2+cu102
 - Datasets 1.18.2.dev0
 - Tokenizers 0.11.0
+
+#### Evaluation Commands
+
+Please use the evaluation script `eval.py` included in the repo.
+
+1. To evaluate on `speech-recognition-community-v2/dev_data`
+
+```bash
+python eval.py --model_id nouamanetazi/wav2vec2-xls-r-300m-ar --dataset speech-recognition-community-v2/dev_data --config ar --split validation --chunk_length_s 5.0 --stride_length_s 1.0
+```

eval.py

@@ -0,0 +1,141 @@
+#!/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)
+
+
+# Normalize arabic
+def normalizeArabic(text):
+    # https://alraqmiyyat.github.io/2013/01-02.html
+    text = re.sub("[إأٱآا]", "ا", text)
+    text = re.sub("ى", "ي", text)
+    text = re.sub("ؤ", "ء", text)
+    text = re.sub("ئ", "ء", text)
+
+    # keep only characters which unicode \u0600-\u06FF and space
+    text = re.sub(r"[^\u0600-\u06FF ]", "", text)
+    return text
+
+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 = re.sub(chars_to_ignore_regex, "", 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))
+
+    text = normalizeArabic(text)
+
+    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)