Update usage of new APIs (#5331)

PaddleAudio/examples/panns/audio_tag.py

@@ -20,9 +20,11 @@ from typing import List
 import numpy as np
 import paddle
 from paddleaudio.backends import load as load_audio
-from paddleaudio.features import melspectrogram
 from paddleaudio.models.panns import cnn14
-from paddleaudio.utils import logger
+from paddleaudio.transforms import LogMelSpectrogram
+from paddleaudio.utils import Timer, get_logger
+
+logger = get_logger()
 
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)
@@ -57,10 +59,7 @@ def batchify(data: List[List[float]], sample_rate: int, batch_size: int,
     """
     Extract features from waveforms and create batches.
     """
-    examples = []
-    for waveform in data:
-        feats = melspectrogram(waveform, sample_rate, **kwargs).transpose()
-        examples.append(feats)
+    examples = data
 
     # Seperates data into some batches.
     one_batch = []
@@ -74,6 +73,7 @@ def batchify(data: List[List[float]], sample_rate: int, batch_size: int,
 
 
 def predict(model,
+            feature_extractor,
             data: List[List[float]],
             sample_rate: int,
             batch_size: int = 1):
@@ -83,10 +83,8 @@ def predict(model,
     batches = batchify(data, sample_rate, batch_size)
     results = None
     model.eval()
-    for batch in batches:
-        feats = paddle.to_tensor(batch).unsqueeze(1)  \
-            # (batch_size, num_frames, num_melbins) -> (batch_size, 1, num_frames, num_melbins)
-
+    for waveforms in batches:
+        feats = feature_extractor(paddle.to_tensor(waveforms))
         audioset_scores = model(feats)
         if results is None:
             results = audioset_scores.numpy()
@@ -98,11 +96,13 @@ def predict(model,
 
 if __name__ == '__main__':
     paddle.set_device(args.device)
+    feature_extractor = LogMelSpectrogram(
+        sr=16000, n_fft=512, hop_length=320, n_mels=64, f_min=50)
     model = cnn14(pretrained=True, extract_embedding=False)
     waveform, sr = load_audio(args.wav, sr=None)
     time, data = split(waveform, int(args.sample_duration * sr),
                        int(args.hop_duration * sr))
-    results = predict(model, data, sr, batch_size=8)
+    results = predict(model, feature_extractor, data, sr, batch_size=8)
 
     if not os.path.exists(args.output_dir):
         os.makedirs(args.output_dir)

PaddleAudio/examples/sound_classification/predict.py

@@ -22,8 +22,8 @@ import paddle.nn.functional as F
 from model import SoundClassifier
 from paddleaudio.backends import load as load_audio
 from paddleaudio.datasets import ESC50
-from paddleaudio.features import melspectrogram
 from paddleaudio.models.panns import cnn14
+from paddleaudio.transforms import LogMelSpectrogram
 
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)
@@ -34,24 +34,19 @@ parser.add_argument("--checkpoint", type=str, required=True, help="Checkpoint of
 args = parser.parse_args()
 # yapf: enable
 
-
-def extract_features(file: str, **kwargs):
-    waveform, sr = load_audio(args.wav, sr=None)
-    feats = melspectrogram(waveform, sr, **kwargs).transpose()
-    return feats
-
-
 if __name__ == '__main__':
     paddle.set_device(args.device)
 
-    model = SoundClassifier(backbone=cnn14(pretrained=False,
-                                           extract_embedding=True),
-                            num_class=len(ESC50.label_list))
+    feature_extractor = LogMelSpectrogram(
+        sr=16000, n_fft=512, hop_length=320, n_mels=64, f_min=50)
+    model = SoundClassifier(
+        backbone=cnn14(pretrained=False, extract_embedding=True),
+        num_class=len(ESC50.label_list))
     model.set_state_dict(paddle.load(args.checkpoint))
     model.eval()
 
-    feats = extract_features(args.wav)
-    feats = paddle.to_tensor(np.expand_dims(feats, 0))
+    waveform, sr = load_audio(args.wav)
+    feats = feature_extractor(paddle.to_tensor(waveform).unsqueeze(0))
     logits = model(feats)
     probs = F.softmax(logits, axis=1).numpy()
 

PaddleAudio/examples/sound_classification/train.py

@@ -21,7 +21,10 @@ import paddle.nn.functional as F
 from model import SoundClassifier
 from paddleaudio.datasets import ESC50
 from paddleaudio.models.panns import cnn14
-from paddleaudio.utils import Timer, logger
+from paddleaudio.transforms import LogMelSpectrogram
+from paddleaudio.utils import Timer, get_logger
+
+logger = get_logger()
 
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)
@@ -41,14 +44,16 @@ if __name__ == "__main__":
     nranks = paddle.distributed.get_world_size()
     local_rank = paddle.distributed.get_rank()
 
+    feature_extractor = LogMelSpectrogram(
+        sr=16000, n_fft=512, hop_length=320, n_mels=64, f_min=50)
     backbone = cnn14(pretrained=True, extract_embedding=True)
     model = SoundClassifier(backbone, num_class=len(ESC50.label_list))
-    optimizer = paddle.optimizer.Adam(learning_rate=args.learning_rate,
-                                      parameters=model.parameters())
+    optimizer = paddle.optimizer.Adam(
+        learning_rate=args.learning_rate, parameters=model.parameters())
     criterion = paddle.nn.loss.CrossEntropyLoss()
 
-    train_ds = ESC50(mode='train', feat_type='melspectrogram')
-    dev_ds = ESC50(mode='dev', feat_type='melspectrogram')
+    train_ds = ESC50(mode='train')
+    dev_ds = ESC50(mode='dev')
 
     train_sampler = paddle.io.DistributedBatchSampler(
         train_ds, batch_size=args.batch_size, shuffle=True, drop_last=False)
@@ -71,7 +76,8 @@ if __name__ == "__main__":
         num_corrects = 0
         num_samples = 0
         for batch_idx, batch in enumerate(train_loader):
-            feats, labels = batch
+            waveforms, labels = batch
+            feats = feature_extractor(waveforms)
             logits = model(feats)
 
             loss = criterion(logits, labels)
@@ -110,10 +116,11 @@ if __name__ == "__main__":
                 num_samples = 0
 
         if epoch % args.save_freq == 0 and batch_idx + 1 == steps_per_epoch and local_rank == 0:
-            dev_sampler = paddle.io.BatchSampler(dev_ds,
-                                                 batch_size=args.batch_size,
-                                                 shuffle=False,
-                                                 drop_last=False)
+            dev_sampler = paddle.io.BatchSampler(
+                dev_ds,
+                batch_size=args.batch_size,
+                shuffle=False,
+                drop_last=False)
             dev_loader = paddle.io.DataLoader(
                 dev_ds,
                 batch_sampler=dev_sampler,
@@ -126,7 +133,8 @@ if __name__ == "__main__":
             num_samples = 0
             with logger.processing('Evaluation on validation dataset'):
                 for batch_idx, batch in enumerate(dev_loader):
-                    feats, labels = batch
+                    waveforms, labels = batch
+                    feats = feature_extractor(waveforms)
                     logits = model(feats)
 
                     preds = paddle.argmax(logits, axis=1)