Fixed normalization

app.py

@@ -43,7 +43,7 @@ def waveformer(audio, label_choices):
     if fs != 44100:
         raise ValueError(fs)
     mixture = torch.from_numpy(
-        mixture).unsqueeze(0).unsqueeze(0).to(torch.float)
+        mixture).unsqueeze(0).unsqueeze(0).to(torch.float) / (2.0 ** 15)
 
     # Construct the query vector
     if len(label_choices) == 0:
@@ -53,7 +53,7 @@ def waveformer(audio, label_choices):
         query[0, TARGETS.index(t)] = 1.
 
     with torch.no_grad():
-        output = model(mixture, query)
+        output = (2.0 ** 15) * model(mixture, query)
 
     return fs, output.squeeze(0).squeeze(0).to(torch.short).numpy()