Feature Extraction
Transformers
Safetensors
finelap
audio grounding
audio-text retrieval
sound-event-detection
multimodal
clap
custom_code
Instructions to use AndreasXi/FineLAP with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use AndreasXi/FineLAP with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("feature-extraction", model="AndreasXi/FineLAP", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("AndreasXi/FineLAP", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
add batch infer
Browse files- modeling_finelap.py +41 -19
modeling_finelap.py
CHANGED
|
@@ -5,6 +5,8 @@ import torchaudio
|
|
| 5 |
from transformers import PreTrainedModel, RobertaModel, RobertaTokenizer
|
| 6 |
from .configuration_finelap import FineLAPConfig
|
| 7 |
from .modeling_eat import EATModel
|
|
|
|
|
|
|
| 8 |
|
| 9 |
class FineLAPModel(PreTrainedModel):
|
| 10 |
config_class = FineLAPConfig
|
|
@@ -42,26 +44,45 @@ class FineLAPModel(PreTrainedModel):
|
|
| 42 |
|
| 43 |
def load_audio(self, audio_path, device=None):
|
| 44 |
device = device or self.device
|
| 45 |
-
|
| 46 |
-
if
|
| 47 |
-
|
| 48 |
-
if sr != 16000:
|
| 49 |
-
wav = torchaudio.functional.resample(wav, sr, 16000)
|
| 50 |
-
wav = wav.squeeze(0)
|
| 51 |
-
wav = wav - wav.mean()
|
| 52 |
-
|
| 53 |
-
mel = torchaudio.compliance.kaldi.fbank(
|
| 54 |
-
wav.unsqueeze(0), htk_compat=True, sample_frequency=16000,
|
| 55 |
-
use_energy=False, window_type='hanning', num_mel_bins=128,
|
| 56 |
-
dither=0.0, frame_shift=10
|
| 57 |
-
)
|
| 58 |
-
target_len = 1024
|
| 59 |
-
if mel.shape[0] < target_len:
|
| 60 |
-
mel = F.pad(mel, (0, 0, 0, target_len - mel.shape[0]))
|
| 61 |
else:
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
|
| 66 |
def encode_audio(self, audio_path):
|
| 67 |
audio_mel = self.load_audio(audio_path)
|
|
@@ -97,6 +118,7 @@ class FineLAPModel(PreTrainedModel):
|
|
| 97 |
|
| 98 |
@torch.no_grad()
|
| 99 |
def get_frame_level_score(self, audio_path, text_labels, device=None):
|
|
|
|
| 100 |
device = device or self.device
|
| 101 |
self.to(device)
|
| 102 |
self.eval()
|
|
|
|
| 5 |
from transformers import PreTrainedModel, RobertaModel, RobertaTokenizer
|
| 6 |
from .configuration_finelap import FineLAPConfig
|
| 7 |
from .modeling_eat import EATModel
|
| 8 |
+
import os
|
| 9 |
+
|
| 10 |
|
| 11 |
class FineLAPModel(PreTrainedModel):
|
| 12 |
config_class = FineLAPConfig
|
|
|
|
| 44 |
|
| 45 |
def load_audio(self, audio_path, device=None):
|
| 46 |
device = device or self.device
|
| 47 |
+
|
| 48 |
+
if isinstance(audio_path, (str, os.PathLike)):
|
| 49 |
+
audio_path = [str(audio_path)]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
else:
|
| 51 |
+
audio_path = [str(p) for p in audio_path] # list/tuple/iterable of paths
|
| 52 |
+
|
| 53 |
+
mels = []
|
| 54 |
+
target_len = 1024
|
| 55 |
+
|
| 56 |
+
for audio_path in audio_path:
|
| 57 |
+
wav, sr = torchaudio.load(audio_path)
|
| 58 |
+
if wav.shape[0] > 1:
|
| 59 |
+
wav = wav.mean(dim=0, keepdim=True)
|
| 60 |
+
if sr != 16000:
|
| 61 |
+
wav = torchaudio.functional.resample(wav, sr, 16000)
|
| 62 |
+
wav = wav.squeeze(0)
|
| 63 |
+
wav = wav - wav.mean()
|
| 64 |
+
|
| 65 |
+
mel = torchaudio.compliance.kaldi.fbank(
|
| 66 |
+
wav.unsqueeze(0),
|
| 67 |
+
htk_compat=True,
|
| 68 |
+
sample_frequency=16000,
|
| 69 |
+
use_energy=False,
|
| 70 |
+
window_type="hanning",
|
| 71 |
+
num_mel_bins=128,
|
| 72 |
+
dither=0.0,
|
| 73 |
+
frame_shift=10,
|
| 74 |
+
) # [T, 128]
|
| 75 |
+
|
| 76 |
+
if mel.shape[0] < target_len:
|
| 77 |
+
mel = F.pad(mel, (0, 0, 0, target_len - mel.shape[0]))
|
| 78 |
+
else:
|
| 79 |
+
mel = mel[:target_len, :]
|
| 80 |
+
|
| 81 |
+
mel = (mel - (-4.268)) / (4.569 * 2) # [T, 128]
|
| 82 |
+
mels.append(mel)
|
| 83 |
+
|
| 84 |
+
mel_batch = torch.stack(mels, dim=0).unsqueeze(1).to(device) # [B, 1, 1024, 128]
|
| 85 |
+
return mel_batch
|
| 86 |
|
| 87 |
def encode_audio(self, audio_path):
|
| 88 |
audio_mel = self.load_audio(audio_path)
|
|
|
|
| 118 |
|
| 119 |
@torch.no_grad()
|
| 120 |
def get_frame_level_score(self, audio_path, text_labels, device=None):
|
| 121 |
+
assert isinstance(audio_path, str) or (isinstance(audio_path, (list, tuple)) and len(audio_path) == 1)
|
| 122 |
device = device or self.device
|
| 123 |
self.to(device)
|
| 124 |
self.eval()
|