update evaluation

README.md

@@ -1,21 +1,25 @@
 ---
 language: pl
+license: apache-2.0
 datasets:
 - common_voice
+- mozilla-foundation/common_voice_6_0
 metrics:
 - wer
 - cer
 tags:
+- pl
 - audio
 - automatic-speech-recognition
 - speech
 - xlsr-fine-tuning-week
-license: apache-2.0
+- robust-speech-event
+- mozilla-foundation/common_voice_6_0
 model-index:
 - name: XLSR Wav2Vec2 Polish by Jonatas Grosman
   results:
   - task: 
-      name: Speech Recognition
+      name: Automatic Speech Recognition 
       type: automatic-speech-recognition
     dataset:
       name: Common Voice pl
@@ -24,10 +28,36 @@ model-index:
     metrics:
        - name: Test WER
          type: wer
-         value: 14.33
+         value: 14.21
+       - name: Test CER
+         type: cer
+         value: 3.49
+       - name: Test WER (+LM)
+         type: wer
+         value: 10.98
+       - name: Test CER (+LM)
+         type: cer
+         value: 2.93
+  - task: 
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: Robust Speech Event - Dev Data
+      type: speech-recognition-community-v2/dev_data
+      args: pl
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 33.18
        - name: Test CER
          type: cer
-         value: 3.36
+         value: 15.92
+       - name: Test WER (+LM)
+         type: wer
+         value: 29.31
+       - name: Test CER (+LM)
+         type: cer
+         value: 15.17
 ---
 
 # Wav2Vec2-Large-XLSR-53-Polish
@@ -109,76 +139,14 @@ for i, predicted_sentence in enumerate(predicted_sentences):
 
 ## Evaluation
 
-The model can be evaluated as follows on the Polish test data of Common Voice.
-
-```python
-import torch
-import re
-import librosa
-from datasets import load_dataset, load_metric
-from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
-
-LANG_ID = "pl"
-MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-polish"
-DEVICE = "cuda"
-
-CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
-                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
-                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
-                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
-                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
+1. To evaluate on `mozilla-foundation/common_voice_6_0` with split `test`
 
-test_dataset = load_dataset("common_voice", LANG_ID, split="test")
-
-wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
-cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py
-
-chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
-
-processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
-model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
-model.to(DEVICE)
-
-# Preprocessing the datasets.
-# We need to read the audio files as arrays
-def speech_file_to_array_fn(batch):
-    with warnings.catch_warnings():
-        warnings.simplefilter("ignore")
-        speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
-    batch["speech"] = speech_array
-    batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()
-    return batch
-
-test_dataset = test_dataset.map(speech_file_to_array_fn)
-
-# Preprocessing the datasets.
-# We need to read the audio files as arrays
-def evaluate(batch):
-    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
-
-    with torch.no_grad():
-        logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits
-
-    pred_ids = torch.argmax(logits, dim=-1)
-    batch["pred_strings"] = processor.batch_decode(pred_ids)
-    return batch
-
-result = test_dataset.map(evaluate, batched=True, batch_size=8)
-
-predictions = [x.upper() for x in result["pred_strings"]]
-references = [x.upper() for x in result["sentence"]]
-
-print(f"WER: {wer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
-print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")
+```bash
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset mozilla-foundation/common_voice_6_0 --config pl --split test
 ```
 
-**Test Result**:
-
-In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-04-22). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used.
+2. To evaluate on `speech-recognition-community-v2/dev_data`
 
-| Model | WER | CER |
-| ------------- | ------------- | ------------- |
-| jonatasgrosman/wav2vec2-large-xlsr-53-polish | **14.33%** | **3.36%** |
-| facebook/wav2vec2-large-xlsr-53-polish | 20.17% | 5.38% |
-| alexcleu/wav2vec2-large-xlsr-polish | 21.72% | 5.17% |
-| mbien/wav2vec2-large-xlsr-polish | 22.93% | 5.13% |
+```bash
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset speech-recognition-community-v2/dev_data --config pl --split validation --chunk_length_s 5.0 --stride_length_s 1.0
+```

eval.py

@@ -0,0 +1,164 @@
+#!/usr/bin/env python3
+from datasets import load_dataset, load_metric, Audio, Dataset
+from transformers import pipeline, AutoFeatureExtractor, AutoTokenizer, AutoConfig, AutoModelForCTC, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM
+import re
+import torch
+import argparse
+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)
+
+
+def normalize_text(text: str, invalid_chars_regex: str, to_lower: bool) -> str:
+    """ DO ADAPT FOR YOUR USE CASE. this function normalizes the target text. """
+
+    text = text.lower() if to_lower else text.upper()
+
+    text = re.sub(invalid_chars_regex, " ", text)
+
+    text = re.sub("\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
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=feature_extractor.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, args=args, asr=asr, invalid_chars_regex=invalid_chars_regex, normalize_to_lower=normalize_to_lower):
+        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, normalize_to_lower)
+        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)

full_eval.sh

@@ -0,0 +1,15 @@
+# CV - TEST
+
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset mozilla-foundation/common_voice_6_0 --config pl --split test --log_outputs --greedy
+mv log_mozilla-foundation_common_voice_6_0_pl_test_predictions.txt log_mozilla-foundation_common_voice_6_0_pl_test_predictions_greedy.txt
+mv mozilla-foundation_common_voice_6_0_pl_test_eval_results.txt mozilla-foundation_common_voice_6_0_pl_test_eval_results_greedy.txt
+
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset mozilla-foundation/common_voice_6_0 --config pl --split test --log_outputs
+
+# HF EVENT - DEV
+
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset speech-recognition-community-v2/dev_data --config pl --split validation --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs --greedy
+mv log_speech-recognition-community-v2_dev_data_pl_validation_predictions.txt log_speech-recognition-community-v2_dev_data_pl_validation_predictions_greedy.txt
+mv speech-recognition-community-v2_dev_data_pl_validation_eval_results.txt speech-recognition-community-v2_dev_data_pl_validation_eval_results_greedy.txt
+
+python eval.py --model_id jonatasgrosman/wav2vec2-large-xlsr-53-polish --dataset speech-recognition-community-v2/dev_data --config pl --split validation --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs

mozilla-foundation_common_voice_6_0_pl_test_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.10980770869429375
+CER: 0.029387492481781432

mozilla-foundation_common_voice_6_0_pl_test_eval_results_greedy.txt

@@ -0,0 +1,2 @@
+WER: 0.14213082625614223
+CER: 0.03491250891695691

speech-recognition-community-v2_dev_data_pl_validation_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.2931926751592357
+CER: 0.15175189393939395

speech-recognition-community-v2_dev_data_pl_validation_eval_results_greedy.txt

@@ -0,0 +1,2 @@
+WER: 0.33187367303609344
+CER: 0.15924530632411066