upload model

app.py

@@ -2,11 +2,14 @@ import streamlit as st
 import torch
 import numpy as np
 import matplotlib.pyplot as plt
-from pydub import AudioSegment
+
 import pretty_midi as pm
+
 from VAE import VAE
-from midi2audio import FluidSynth
+
 import pretty_midi as pm
+from scipy.io.wavfile import write
+
 
 
 # Define device
@@ -70,25 +73,23 @@ def reconstruct(right, left, model):
     input_tensor = torch.cat([right_tensor, left_tensor], dim=0) 
     input_tensor = input_tensor.unsqueeze(0) 
     
-    print(input_tensor.shape)
+
     with torch.no_grad():
         recon_data, _, _, _ = model(input_tensor)
 
     return recon_data.squeeze(0).cpu().numpy()
 
 
+def midi_to_wav(midi_file, wav_file="output.wav", volume_increase_db=17):
+    midi_data = pm.PrettyMIDI(midi_file)
+    audio_data = midi_data.synthesize(fs=44100)  
 
-def midi_to_wav(midi_file, wav_file="output.wav", sound_font_path="soundfont.sf2", volume_increase_db=17):
-    fs = FluidSynth(sound_font_path)  
-    fs.midi_to_audio(midi_file, wav_file)  
+    audio_data = np.int16(audio_data / np.max(np.abs(audio_data)) * 32767 * 0.9) 
 
-    audio = AudioSegment.from_wav(wav_file)
-    louder_audio = audio + volume_increase_db 
-    
-    louder_audio.export(wav_file, format="wav")
-    
+    write(wav_file, 44100, audio_data)
     return wav_file
 
+
 # Create a MIDI stream from piano roll data
 def create_midi_from_piano_roll(right_hand, left_hand, fs=8):
     pm_obj = pm.PrettyMIDI()
@@ -142,6 +143,7 @@ def convert_to_midi(right_hand, left_hand, file_name="output.mid", fs=8):
 # Streamlit interface
 st.title("GRU-VAE Reconstruction Demo")
 model = load_model()
+    
 
 # File upload
 uploaded_file = st.file_uploader("Upload a MIDI file", type=["mid", "midi"])

requirements.txt

@@ -1,6 +1,5 @@
 streamlit==1.16.0
 torch==1.11.0
 pretty_midi
-midi2audio
 scipy
 pydub