up

eval.py

@@ -1,45 +1,126 @@
 #!/usr/bin/env python3
-from datasets import load_dataset, load_metric, Audio
-from transformers import AutoModelForCTC, AutoProcessor, Wav2Vec2Processor
-import torch
+from datasets import load_dataset, load_metric, Audio, Dataset
+from transformers import pipeline, AutoFeatureExtractor
 import re
+import argparse
+from typing import Dict
 
-lang = "sv-SE"
-model_id = "./xls-r-300m-sv"
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
+def log_results(result: Dataset, args: Dict[str, str]):
+    """ DO NOT CHANGE. This function computes and logs the result metrics. """
 
-dataset = load_dataset("mozilla-foundation/common_voice_7_0", lang, split="test", use_auth_token=True)
-wer = load_metric("wer")
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
 
-dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
 
-model = AutoModelForCTC.from_pretrained(model_id).to(device)
-processor = AutoProcessor.from_pretrained(model_id)
+    # 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)
 
-chars_to_ignore_regex = '[,?.!\-\;\:\"“%‘”�—’…–]'  # change to the ignored characters of your fine-tuned model
+    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"
 
-def map_to_pred(batch):
-    input_values = processor(batch["audio"]["array"], return_tensors="pt", padding="longest", sampling_rate=16_000).input_values
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
 
-    with torch.no_grad():
-        logits = model(input_values.to(device)).logits
+            # 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")
 
-    if processor.__class__.__name__ == "Wav2Vec2Processor":
-        predicted_ids = torch.argmax(logits, dim=-1)
-        transcription = processor.batch_decode(predicted_ids)[0]
-    else:
-        transcription = processor.batch_decode(logits.cpu().numpy()).text[0]
+            result.map(write_to_file, with_indices=True)
 
-    batch["transcription"] = transcription
-    batch["text"] = re.sub(chars_to_ignore_regex, "", batch["sentence"].lower())
-    return batch
 
+def normalize_text(text: str) -> str:
+    """ DO ADAPT FOR YOUR USE CASE. this function normalizes the target text. """
 
-result = dataset.map(map_to_pred, remove_columns=["audio"])
+    chars_to_ignore_regex = '[,?.!\-\;\:\"“%‘”�—’…–]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
 
-wer_result = wer.compute(references=result["text"], predictions=result["transcription"])
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
 
-print("WER", wer_result)
+    # 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(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 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)

log_mozilla-foundation_common_voice_7_0_sv-SE_test_predictions.txt

@@ -0,0 +1,20 @@
+0
+jag lämnade grovjobbet åt honom
+1
+ja för att åter få ett stulet föremål
+2
+har du fortfarande samma nummer
+3
+det räcker inte
+4
+där är om
+5
+vill jag se dig död skulle du var det
+6
+vi är oemottagliga för din utstrålning
+7
+jag vet att du pratar med honom
+8
+dra åt helvete här u
+9
+hon fick bra betyg för att hon pluggade på kvällen

log_mozilla-foundation_common_voice_7_0_sv-SE_test_targets.txt

@@ -0,0 +1,20 @@
+0
+jag lämnade grovjobbet åt honom
+1
+ja för att återfå ett stulet föremål
+2
+har du fortfarande samma nummer
+3
+det räcker inte
+4
+där är de
+5
+ville jag se dig död skulle du vara det
+6
+vi är oemottagliga för din utstrålning
+7
+jag vet att du pratar med honom
+8
+dra åt helvete harry
+9
+hon fick bra betyg för att hon pluggade på kvällen

mozilla-foundation_common_voice_7_0_sv-SE_test_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.11864406779661017
+CER: 0.02666666666666667

run_cv_eval.sh

@@ -0,0 +1 @@
+./eval.py --model_id hf-test/xls-r-300m-sv --dataset mozilla-foundation/common_voice_7_0 --config sv-SE --split test --log_outputs

run_real_eval.sh

@@ -0,0 +1 @@
+./eval.py --model_id hf-test/xls-r-300m-sv --dataset speech-recognition-community-internal/tedx_manual_dev_test --config sv --split validation --chunk_length_s 5.0 --stride_length_s 1.0 --log_outputs