New model results

README.md

@@ -23,7 +23,7 @@ model-index:
     metrics:
        - name: Test WER
          type: wer
-         value: 22.36
+         value: 18.44
 ---
 
 # Wav2Vec2-Large-XLSR-53-Russian
@@ -49,17 +49,17 @@ model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-russian")
 resampler = torchaudio.transforms.Resample(48_000, 16_000)
 
 # Preprocessing the datasets.
-# We need to read the aduio files as arrays
+# We need to read the audio files as arrays
 def speech_file_to_array_fn(batch):
-\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
-\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
-\treturn batch
+    speech_array, sampling_rate = torchaudio.load(batch["path"])
+    batch["speech"] = resampler(speech_array).squeeze().numpy()
+    return batch
 
 test_dataset = test_dataset.map(speech_file_to_array_fn)
 inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
 
 with torch.no_grad():
-\tlogits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
+\\tlogits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
 
 predicted_ids = torch.argmax(logits, dim=-1)
 
@@ -70,67 +70,74 @@ print("Reference:", test_dataset["sentence"][:2])
 
 ## Evaluation
 
-The model can be evaluated as follows on the {language} test data of Common Voice.  # TODO: replace #TODO: replace language with your {language}, *e.g.* French
+The model can be evaluated as follows on the Russian test data of Common Voice.
 
 
 ```python
 import torch
 import torchaudio
-from datasets import load_dataset, load_metric
+import urllib.request
+import tarfile
+import pandas as pd
+from tqdm.auto import tqdm
+from datasets import load_metric
 from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
-import re
 
-test_dataset = load_dataset("common_voice", "ru", split="test")
+# Download the raw data instead of using HF datasets to save disk space 
+data_url = "https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-6.1-2020-12-11/ru.tar.gz"
+filestream = urllib.request.urlopen(data_url)
+data_file = tarfile.open(fileobj=filestream, mode="r|gz")
+data_file.extractall()
+
 wer = load_metric("wer")
 
 processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-russian")
 model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-russian")
 model.to("cuda")
 
-resampler = torchaudio.transforms.Resample(48_000, 16_000)
+cv_test = pd.read_csv("cv-corpus-6.1-2020-12-11/ru/test.tsv", sep='\t')
+clips_path = "cv-corpus-6.1-2020-12-11/ru/clips/"
 
 def clean_sentence(sent):
-\tsent = sent.lower()
-\t# replace non-alphanumeric characters with space ("какой-то, вот" -> "какой то  вот")
-\tsent = "".join(ch if ch.isalnum() else " " for ch in sent)
-\t# remove repeated spaces
-\tsent = " ".join(sent.split())
-\t# these letters are considered equivalent in written Russian
-\tsent = sent.replace('ё', 'е')
-\treturn sent
+    sent = sent.lower()
+    # these letters are considered equivalent in written Russian
+    sent = sent.replace('ё', 'е')
+    # replace non-alpha characters with space
+    sent = "".join(ch if ch.isalpha() else " " for ch in sent)
+    # remove repeated spaces
+    sent = " ".join(sent.split())
+    return sent
 
-# Preprocessing the datasets.
-# We need to read the aduio files as arrays
-def speech_file_to_array_fn(batch):
-\tbatch["sentence"] = clean_sentence(batch["sentence"])
-\tspeech_array, sampling_rate = torchaudio.load(batch["path"])
-\tbatch["speech"] = resampler(speech_array).squeeze().numpy()
-\treturn batch
+targets = []
+preds = []
 
-test_dataset = test_dataset.map(speech_file_to_array_fn)
+for i, row in tqdm(cv_test.iterrows(), total=cv_test.shape[0]):
+    row["sentence"] = clean_sentence(row["sentence"])
+    speech_array, sampling_rate = torchaudio.load(clips_path + row["path"])
+    resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)
+    row["speech"] = resampler(speech_array).squeeze().numpy()
 
-# Preprocessing the datasets.
-# We need to read the aduio files as arrays
-def evaluate(batch):
-\tinputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
+    inputs = processor(row["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
 
-\twith torch.no_grad():
-\t\tlogits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
+    with torch.no_grad():
+        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
 
-\tpred_ids = torch.argmax(logits, dim=-1)
-\tbatch["pred_strings"] = processor.batch_decode(pred_ids)
-\treturn batch
+    pred_ids = torch.argmax(logits, dim=-1)
 
-result = test_dataset.map(evaluate, batched=True, batch_size=8)
+    targets.append(row["sentence"])
+    preds.append(processor.batch_decode(pred_ids)[0])
 
-print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
+# free up some memory
+del model
+del processor
+del cv_test
+
+print("WER: {:2f}".format(100 * wer.compute(predictions=preds, references=targets)))
 ```
 
-**Test Result**: 22.36 %  
+**Test Result**: 18.44 %  
 
 
 ## Training
 
 The Common Voice `train` and `validation` datasets were used for training.
-
-The script used for training can be found [here](github.com)