Updated the README
Browse files
README.md
CHANGED
@@ -88,7 +88,81 @@ print("Reference:", test_dataset["sentence"][:2])
|
|
88 |
|
89 |
## Evaluation
|
90 |
|
91 |
-
The model can be evaluated as follows on the
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
|
93 |
|
94 |
```python
|
|
|
88 |
|
89 |
## Evaluation
|
90 |
|
91 |
+
The model can be evaluated as follows on the following two datasets:
|
92 |
+
1. Custom dataset created from 20% of Indic, IIITH and CV (test)
|
93 |
+
2. CommonVoice Hindi test dataset
|
94 |
+
|
95 |
+
```python
|
96 |
+
import torch
|
97 |
+
import torchaudio
|
98 |
+
from datasets import load_dataset, load_metric
|
99 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
|
100 |
+
import re
|
101 |
+
|
102 |
+
## Load the datasets
|
103 |
+
test_dataset = load_dataset("common_voice", "hi", split="test")
|
104 |
+
|
105 |
+
indic = load_dataset("csv", data_files= {'train':"/workspace/data/hi2/indic_train_full.csv",
|
106 |
+
"test": "/workspace/data/hi2/indic_test_full.csv"}, download_mode="force_redownload")
|
107 |
+
iiith = load_dataset("csv", data_files= {"train": "/workspace/data/hi2/iiit_hi_train.csv",
|
108 |
+
"test": "/workspace/data/hi2/iiit_hi_test.csv"}, download_mode="force_redownload")
|
109 |
+
|
110 |
+
## Pre-process datasets and concatenate to create test dataset
|
111 |
+
# Drop columns of common_voice
|
112 |
+
split = ['train', 'test', 'validation', 'other', 'invalidated']
|
113 |
+
|
114 |
+
for sp in split:
|
115 |
+
common_voice[sp] = common_voice[sp].remove_columns(['client_id', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'])
|
116 |
+
|
117 |
+
common_voice = common_voice.rename_column('path', 'audio_path')
|
118 |
+
common_voice = common_voice.rename_column('sentence', 'target_text')
|
119 |
+
|
120 |
+
train_dataset = datasets.concatenate_datasets([indic['train'], iiith['train'], common_voice['train']])
|
121 |
+
test_dataset = datasets.concatenate_datasets([indic['test'], iiith['test'], common_voice['test'], common_voice['validation']])
|
122 |
+
|
123 |
+
## Load model from HF hub
|
124 |
+
|
125 |
+
wer = load_metric("wer")
|
126 |
+
|
127 |
+
processor = Wav2Vec2Processor.from_pretrained("skylord/wav2vec2-large-xlsr-hindi")
|
128 |
+
model = Wav2Vec2ForCTC.from_pretrained("skylord/wav2vec2-large-xlsr-hindi")
|
129 |
+
model.to("cuda")
|
130 |
+
|
131 |
+
chars_to_ignore_regex = '[\,\?\.\!\-\'\;\:\"\“\%\‘\”\�Utrnle\_]'
|
132 |
+
unicode_ignore_regex = r'[dceMaWpmFui\xa0\u200d]' # Some unwanted unicode chars
|
133 |
+
resampler = torchaudio.transforms.Resample(48_000, 16_000)
|
134 |
+
|
135 |
+
# Preprocessing the datasets.
|
136 |
+
# We need to read the aduio files as arrays
|
137 |
+
|
138 |
+
def speech_file_to_array_fn(batch):
|
139 |
+
batch["target_text"] = re.sub(chars_to_ignore_regex, '', batch["target_text"])
|
140 |
+
batch["target_text"] = re.sub(unicode_ignore_regex, '', batch["target_text"])
|
141 |
+
|
142 |
+
speech_array, sampling_rate = torchaudio.load(batch["audio_path"])
|
143 |
+
batch["speech"] = resampler(speech_array).squeeze().numpy()
|
144 |
+
return batch
|
145 |
+
|
146 |
+
test_dataset = test_dataset.map(speech_file_to_array_fn)
|
147 |
+
|
148 |
+
# Preprocessing the datasets.
|
149 |
+
# We need to read the aduio files as arrays
|
150 |
+
|
151 |
+
def evaluate(batch):
|
152 |
+
inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
|
153 |
+
with torch.no_grad():
|
154 |
+
logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
|
155 |
+
pred_ids = torch.argmax(logits, dim=-1)
|
156 |
+
batch["pred_strings"] = processor.batch_decode(pred_ids)
|
157 |
+
return batch
|
158 |
+
|
159 |
+
result = test_dataset.map(evaluate, batched=True, batch_size=8)
|
160 |
+
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))
|
161 |
+
```
|
162 |
+
|
163 |
+
**Test Result on custom dataset**: 19.xx %
|
164 |
+
|
165 |
+
|
166 |
|
167 |
|
168 |
```python
|