Initial commit

app.py

@@ -97,10 +97,11 @@ if __name__ == "__main__":
     # model = download_whisper()
     # extract_and_save_encoder(model)\
 
-    if torch.cuda.is_available():
-        device = "cuda"
-    else:
-        device = "cpu"
+    # if torch.cuda.is_available():
+    #     device = "cuda"
+    # else:
+    #     device = "cpu"
+    device = 'cpu'
 
     with open('config.yaml', "r") as f:
         config = yaml.safe_load(f)

evaluate_models.py

@@ -189,9 +189,9 @@ def inference(
     y_pred_label = torch.Tensor([]).to(device)
 
     preds  = []
-
+    model = model.to(device)
     for i, (batch_x, _, batch_y, metadata) in enumerate(test_loader):
-        model.eval()
+        model = model.eval()
         _, path, _, _ = metadata
         if i % 10 == 0:
             print(f"Batch [{i}/{batches_number}]")
@@ -201,6 +201,9 @@ def inference(
             batch_y = batch_y.to(device)
             num_total += batch_x.size(0)
 
+            print('batch device', batch_x)
+            print('model device', model)
+
             batch_pred = model(batch_x).squeeze(1)
             batch_pred = torch.sigmoid(batch_pred)
             batch_pred_label = (batch_pred + 0.5).int()

src/datasets/base_dataset.py

@@ -84,9 +84,6 @@ class SimpleAudioFakeDataset(Dataset):
             path, label, attack_type = self.samples[index]
 
         waveform, sample_rate = torchaudio.load(path, normalize=APPLY_NORMALIZATION)
-        import librosa
-        # waveform, sample_rate = librosa.load(path, sr=SAMPLING_RATE)
-        # waveform = torch.tensor(waveform)
         print('waveform', waveform)
         real_sec_length = len(waveform[0]) / sample_rate
 

src/frontends.py

@@ -7,7 +7,8 @@ SAMPLING_RATE = 16_000
 win_length = 400  # int((25 / 1_000) * SAMPLING_RATE)
 hop_length = 160  # int((10 / 1_000) * SAMPLING_RATE)
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
+# device = "cuda" if torch.cuda.is_available() else "cpu"
+device = 'cpu'
 
 MFCC_FN = torchaudio.transforms.MFCC(
     sample_rate=SAMPLING_RATE,
@@ -39,7 +40,7 @@ MEL_SCALE_FN = torchaudio.transforms.MelScale(
 delta_fn = torchaudio.transforms.ComputeDeltas(
     win_length=400,
     mode="replicate",
-)
+).to(device)
 
 
 def get_frontend(
@@ -65,6 +66,7 @@ def prepare_lfcc_double_delta(input):
 def prepare_mfcc_double_delta(input):
     if input.ndim < 4:
         input = input.unsqueeze(1)  # (bs, 1, n_lfcc, frames)
+    input.to(device)
     x = MFCC_FN(input)
     delta = delta_fn(x)
     double_delta = delta_fn(delta)