def test_inference_ctc_batched(self):
model = HubertForCTC.from_pretrained(
"facebook/hubert-large-ls960-ft",
torch_dtype=torch.float16).to(torch_device)
processor = Wav2Vec2Processor.from_pretrained(
"facebook/hubert-large-ls960-ft", do_lower_case=True)
input_speech = self._load_datasamples(2)
inputs = processor(input_speech, return_tensors="pt", padding=True)
input_values = inputs.input_values.half().to(torch_device)
attention_mask = inputs.attention_mask.to(torch_device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
predicted_trans = processor.batch_decode(predicted_ids)
EXPECTED_TRANSCRIPTIONS = [
"a man said to the universe sir i exist",
"sweat covered brion's body trickling into the tight loin cloth that was the only garment he wore",
]
self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
# only load the relevant featuresets for featurization to save memory
if 'allosaurus_features' in feature_sets:
import allosaurus_features
if 'audioset_features' in feature_sets:
import audioset_features
if 'audiotext_features' in feature_sets:
import audiotext_features
if 'hubert_features' in feature_sets:
import hubert_features
import torch
from transformers import HubertModel, HubertConfig
from transformers import Wav2Vec2Processor, HubertForCTC
import soundfile as sf
hubert_processor_ = Wav2Vec2Processor.from_pretrained("facebook/hubert-large-ls960-ft")
hubert_model_ = HubertForCTC.from_pretrained("facebook/hubert-large-ls960-ft")
else:
hubert_model_ = ''
hubert_processor_ =''
if 'librosa_features' in feature_sets:
import librosa_features
if 'loudness_features' in feature_sets:
import loudness_features
if 'meta_features' in feature_sets:
import meta_features
os.system('pip3 install scikit-learn==0.19.1')
if 'mixed_features' in feature_sets:
import mixed_features
if 'multispeaker_features' in feature_sets:
import multispeaker_features
if 'myprosody_features' in feature_sets:
sampling_rate = 16000
channels = 1
batch_size = 1
my_dataset = LPAudioSet(os.path.join(
os.path.dirname(os.path.abspath(__file__)), '..', 'data', 'audio'),
sr=sampling_rate,
channels=channels)
train_loader = torch.utils.data.DataLoader(my_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1,
drop_last=True,
collate_fn=LPAudioSet.collate_fn)
for idx, audio in enumerate(train_loader):
print(idx, audio.shape)
#sys.exit(0)
processor = Wav2Vec2Processor.from_pretrained(
"facebook/hubert-xlarge-ls960-ft",
cache_dir=os.getenv("cache_dir", "../../models"))
model = HubertForCTC.from_pretrained("facebook/hubert-xlarge-ls960-ft",
cache_dir=os.getenv(
"cache_dir", "../../models"))
for idx, audio in enumerate(train_loader):
input_values = processor(audio,
sampling_rate=sampling_rate,
return_tensors="pt").input_values # Batch size 1
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.decode(predicted_ids[0])
print(transcription)
