Add eval WER and model card

README.md

@@ -0,0 +1,101 @@
+---
+language: 
+- te
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- openslr_SLR66
+- generated_from_trainer
+- robust-speech-event
+datasets:
+- openslr
+- SLR66
+metrics:
+- wer
+model-index:
+- name: xls-r-1B-te
+  results:
+  - task: 
+      type: automatic-speech-recognition
+      name: Speech Recognition
+    dataset:
+      type: openslr
+      name: Open SLR
+      args: SLR66
+    metrics:
+      - type: wer
+        value: 0.51
+        name: Test WER
+      - type: cer
+        value: 0.097
+        name: Test CER
+
+---
+
+<!-- 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. -->
+
+# 
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-2b](https://huggingface.co/facebook/wav2vec2-xls-r-2b) on the OPENSLR_SLR66 - NA dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.4253
+- Wer: 0.5109
+
+
+### Evaluation metrics
+
+| Metric | Split  | Decode with LM | Value     |
+|:------:|:------:|:--------------:|:---------:|
+| WER    | Train  | No             |      |
+| CER    | Train  | No             |      |
+| WER    | Test   | No             |      |
+| CER    | Test   | No             |      |
+| WER    | Train  | Yes            |      |
+| CER    | Train  | Yes            |      |
+| WER    | Test   | Yes            |      |
+| CER    | Test   | Yes            |      |
+
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 2e-05
+- train_batch_size: 4
+- eval_batch_size: 4
+- seed: 42
+- gradient_accumulation_steps: 12
+- total_train_batch_size: 64
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- learning_rate: 3e-6
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 2000
+- num_epochs: 150.0
+- hidden_dropout: 0.15
+- mixed_precision_training: Native AMP
+
+### Training results
+
+
+
+
+### Framework versions
+
+- Transformers 4.16.0.dev0
+- Pytorch 1.10.1+cu102
+- Datasets 1.17.1.dev0
+- Tokenizers 0.11.0

eval.py

@@ -0,0 +1,156 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+from datasets import Audio, Dataset, load_dataset, load_metric, DatasetDict
+
+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)
+
+
+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))
+
+    return text
+
+
+def get_telugu_dataset(validation_split=False):
+    dataset = load_dataset('openslr', 'SLR66')
+
+    seed=1242
+    
+    if validation_split:
+        train_testvalid = dataset['train'].train_test_split(test_size=0.2, seed=seed)
+        # Split the 10% test + valid in half test, half valid
+        test_valid = train_testvalid['test'].train_test_split(test_size=0.33, seed=seed)
+        # gather everyone if you want to have a single DatasetDict
+        out_dataset = DatasetDict({
+            'train': train_testvalid['train'],
+            'test': test_valid['test'],
+            'valid': test_valid['train']})
+    else:
+        train_testvalid = dataset['train'].train_test_split(test_size=0.25, seed=seed)
+        out_dataset = DatasetDict({
+            'train': train_testvalid['train'],
+            'test': train_testvalid['test']})
+    return out_dataset
+
+
+def main(args):
+    # load dataset
+    te_dataset = get_telugu_dataset(validation_split=False)
+    def load_te_dataset(split):
+        return te_dataset[split]
+    
+    # dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+    dataset = load_te_dataset(split=args.split)
+
+    # 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, device=0)
+
+    # 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__":
+    # python eval.py --model_id="xls-r-2B-te" --dataset="openslr" --config="te" --split="test" --log_outputs
+    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."
+    )
+    args = parser.parse_args()
+
+    main(args)