Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,45 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import argparse
|
| 2 |
import base64
|
| 3 |
import io
|
| 4 |
-
import os
|
| 5 |
-
import random
|
| 6 |
import numpy as np
|
| 7 |
-
import gradio as gr
|
| 8 |
-
import rtmidi
|
| 9 |
-
import onnxruntime as rt
|
| 10 |
from huggingface_hub import hf_hub_download
|
|
|
|
| 11 |
import MIDI
|
| 12 |
from midi_synthesizer import MidiSynthesizer
|
| 13 |
from midi_tokenizer import MIDITokenizer
|
| 14 |
|
| 15 |
-
#
|
| 16 |
-
|
| 17 |
-
IN_SPACE = os.getenv("SYSTEM") == "spaces"
|
| 18 |
-
MAX_LENGTH = 1024 # Maximum tokens for generation
|
| 19 |
|
| 20 |
-
# MIDI Device Manager
|
| 21 |
class MIDIDeviceManager:
|
|
|
|
| 22 |
def __init__(self):
|
| 23 |
self.midiout = rtmidi.MidiOut()
|
| 24 |
self.midiin = rtmidi.MidiIn()
|
| 25 |
|
| 26 |
def get_device_info(self):
|
|
|
|
| 27 |
out_ports = self.midiout.get_ports() or ["No MIDI output devices"]
|
| 28 |
in_ports = self.midiin.get_ports() or ["No MIDI input devices"]
|
| 29 |
return f"Output Devices:\n{'\n'.join(out_ports)}\n\nInput Devices:\n{'\n'.join(in_ports)}"
|
| 30 |
|
| 31 |
def close(self):
|
|
|
|
| 32 |
if self.midiout.is_port_open():
|
| 33 |
self.midiout.close_port()
|
| 34 |
if self.midiin.is_port_open():
|
| 35 |
self.midiin.close_port()
|
| 36 |
del self.midiout, self.midiin
|
| 37 |
|
| 38 |
-
# MIDI Processor with ONNX Generation
|
| 39 |
class MIDIManager:
|
|
|
|
| 40 |
def __init__(self):
|
| 41 |
-
|
| 42 |
-
self.
|
|
|
|
| 43 |
self.tokenizer = self._load_tokenizer("skytnt/midi-model")
|
| 44 |
self.model_base = rt.InferenceSession(
|
| 45 |
hf_hub_download(repo_id="skytnt/midi-model", filename="onnx/model_base.onnx"),
|
|
@@ -49,10 +50,10 @@ class MIDIManager:
|
|
| 49 |
hf_hub_download(repo_id="skytnt/midi-model", filename="onnx/model_token.onnx"),
|
| 50 |
providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
|
| 51 |
)
|
| 52 |
-
self.generated_files = []
|
| 53 |
-
self.is_playing = False
|
| 54 |
|
| 55 |
def _load_tokenizer(self, repo_id):
|
|
|
|
| 56 |
config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
|
| 57 |
with open(config_path, "r") as f:
|
| 58 |
config = json.load(f)
|
|
@@ -61,39 +62,32 @@ class MIDIManager:
|
|
| 61 |
return tokenizer
|
| 62 |
|
| 63 |
def load_midi(self, file_path):
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
except Exception as e:
|
| 67 |
-
raise ValueError(f"Failed to load MIDI file: {e}")
|
| 68 |
|
| 69 |
-
def
|
| 70 |
-
|
| 71 |
mid_seq = self.tokenizer.tokenize(MIDI.midi2score(midi_data))
|
| 72 |
input_tensor = np.array([mid_seq], dtype=np.int64)
|
| 73 |
cur_len = input_tensor.shape[1]
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
probs = softmax(logits / temp, axis=-1)
|
| 82 |
-
next_token = sample_top_p_k(probs, top_p, top_k, generator)
|
| 83 |
input_tensor = np.concatenate([input_tensor, next_token], axis=1)
|
| 84 |
cur_len += 1
|
| 85 |
-
|
| 86 |
-
# Detokenize and save as MIDI
|
| 87 |
new_seq = input_tensor[0].tolist()
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
MIDI.
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
return midi_data
|
| 94 |
|
| 95 |
def play_midi(self, midi_data):
|
| 96 |
-
|
| 97 |
midi_bytes = base64.b64decode(midi_data)
|
| 98 |
midi_file = MIDI.load(io.BytesIO(midi_bytes))
|
| 99 |
audio = io.BytesIO()
|
|
@@ -101,91 +95,103 @@ class MIDIManager:
|
|
| 101 |
audio.seek(0)
|
| 102 |
return audio
|
| 103 |
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
probs_sort[:, k:] = 0.0 # Top-k filtering
|
| 119 |
-
probs_sort /= probs_sort.sum(axis=-1, keepdims=True)
|
| 120 |
-
next_token = generator.choice(probs.shape[-1], p=probs_sort[0])
|
| 121 |
-
return np.array([[next_token]])
|
| 122 |
-
|
| 123 |
-
# UI Functions
|
| 124 |
-
def process_midi_upload(files):
|
| 125 |
if not files:
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
|
|
|
|
|
|
| 129 |
midi_data = midi_processor.load_midi(file.name)
|
| 130 |
-
generated_midi = midi_processor.
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
if __name__ == "__main__":
|
| 155 |
-
parser = argparse.ArgumentParser(description="MIDI Composer
|
| 156 |
-
parser.add_argument("--port", type=int, default=7860)
|
| 157 |
-
parser.add_argument("--share", action="store_true")
|
| 158 |
-
|
| 159 |
|
| 160 |
device_manager = MIDIDeviceManager()
|
| 161 |
midi_processor = MIDIManager()
|
| 162 |
|
| 163 |
-
with gr.Blocks(
|
| 164 |
-
|
| 165 |
-
|
|
|
|
| 166 |
with gr.Tabs():
|
| 167 |
# MIDI Prompt Tab
|
| 168 |
with gr.Tab("MIDI Prompt"):
|
| 169 |
-
midi_upload = gr.File(label="Upload MIDI File",
|
| 170 |
-
|
| 171 |
status = gr.Textbox(label="Status", value="Ready", interactive=False)
|
| 172 |
-
midi_upload.change(
|
| 173 |
-
process_midi_upload,
|
| 174 |
-
inputs=[midi_upload],
|
| 175 |
-
outputs=[audio_output, status, gr.HTML(elem_id="downloads")]
|
| 176 |
-
)
|
| 177 |
|
| 178 |
-
#
|
| 179 |
-
with gr.Tab("
|
| 180 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 181 |
|
| 182 |
# Devices Tab
|
| 183 |
with gr.Tab("Devices"):
|
| 184 |
-
device_info = gr.Textbox(label="MIDI Devices", value=device_manager.get_device_info(), interactive=False)
|
| 185 |
refresh_btn = gr.Button("Refresh Devices")
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 189 |
|
| 190 |
-
|
|
|
|
| 191 |
device_manager.close()
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import json
|
| 3 |
+
import rtmidi
|
| 4 |
+
import os
|
| 5 |
import argparse
|
| 6 |
import base64
|
| 7 |
import io
|
|
|
|
|
|
|
| 8 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
| 9 |
from huggingface_hub import hf_hub_download
|
| 10 |
+
import onnxruntime as rt
|
| 11 |
import MIDI
|
| 12 |
from midi_synthesizer import MidiSynthesizer
|
| 13 |
from midi_tokenizer import MIDITokenizer
|
| 14 |
|
| 15 |
+
# Match the JavaScript constant
|
| 16 |
+
MIDI_OUTPUT_BATCH_SIZE = 4
|
|
|
|
|
|
|
| 17 |
|
|
|
|
| 18 |
class MIDIDeviceManager:
|
| 19 |
+
"""Manages MIDI input/output devices."""
|
| 20 |
def __init__(self):
|
| 21 |
self.midiout = rtmidi.MidiOut()
|
| 22 |
self.midiin = rtmidi.MidiIn()
|
| 23 |
|
| 24 |
def get_device_info(self):
|
| 25 |
+
"""Returns a string listing available MIDI devices."""
|
| 26 |
out_ports = self.midiout.get_ports() or ["No MIDI output devices"]
|
| 27 |
in_ports = self.midiin.get_ports() or ["No MIDI input devices"]
|
| 28 |
return f"Output Devices:\n{'\n'.join(out_ports)}\n\nInput Devices:\n{'\n'.join(in_ports)}"
|
| 29 |
|
| 30 |
def close(self):
|
| 31 |
+
"""Closes open MIDI ports."""
|
| 32 |
if self.midiout.is_port_open():
|
| 33 |
self.midiout.close_port()
|
| 34 |
if self.midiin.is_port_open():
|
| 35 |
self.midiin.close_port()
|
| 36 |
del self.midiout, self.midiin
|
| 37 |
|
|
|
|
| 38 |
class MIDIManager:
|
| 39 |
+
"""Handles MIDI processing, generation, and playback."""
|
| 40 |
def __init__(self):
|
| 41 |
+
# Load soundfont and models from Hugging Face
|
| 42 |
+
self.soundfont_path = hf_hub_download(repo_id="skytnt/midi-model", filename="soundfont.sf2")
|
| 43 |
+
self.synthesizer = MidiSynthesizer(self.soundfont_path)
|
| 44 |
self.tokenizer = self._load_tokenizer("skytnt/midi-model")
|
| 45 |
self.model_base = rt.InferenceSession(
|
| 46 |
hf_hub_download(repo_id="skytnt/midi-model", filename="onnx/model_base.onnx"),
|
|
|
|
| 50 |
hf_hub_download(repo_id="skytnt/midi-model", filename="onnx/model_token.onnx"),
|
| 51 |
providers=['CUDAExecutionProvider', 'CPUExecutionProvider']
|
| 52 |
)
|
| 53 |
+
self.generated_files = []
|
|
|
|
| 54 |
|
| 55 |
def _load_tokenizer(self, repo_id):
|
| 56 |
+
"""Loads the MIDI tokenizer configuration."""
|
| 57 |
config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
|
| 58 |
with open(config_path, "r") as f:
|
| 59 |
config = json.load(f)
|
|
|
|
| 62 |
return tokenizer
|
| 63 |
|
| 64 |
def load_midi(self, file_path):
|
| 65 |
+
"""Loads a MIDI file from the given path."""
|
| 66 |
+
return MIDI.load(file_path)
|
|
|
|
|
|
|
| 67 |
|
| 68 |
+
def generate_onnx(self, midi_data):
|
| 69 |
+
"""Generates a MIDI variation using ONNX models."""
|
| 70 |
mid_seq = self.tokenizer.tokenize(MIDI.midi2score(midi_data))
|
| 71 |
input_tensor = np.array([mid_seq], dtype=np.int64)
|
| 72 |
cur_len = input_tensor.shape[1]
|
| 73 |
+
max_len = 1024
|
| 74 |
+
while cur_len < max_len:
|
| 75 |
+
inputs = {"x": input_tensor[:, -1:]}
|
| 76 |
+
hidden = self.model_base.run(None, inputs)[0]
|
| 77 |
+
logits = self.model_token.run(None, {"hidden": hidden})[0]
|
| 78 |
+
probs = self._softmax(logits, axis=-1)
|
| 79 |
+
next_token = self._sample_top_p_k(probs, 0.98, 20)
|
|
|
|
|
|
|
| 80 |
input_tensor = np.concatenate([input_tensor, next_token], axis=1)
|
| 81 |
cur_len += 1
|
|
|
|
|
|
|
| 82 |
new_seq = input_tensor[0].tolist()
|
| 83 |
+
generated_midi = self.tokenizer.detokenize(new_seq)
|
| 84 |
+
# Store base64-encoded MIDI data for downloads
|
| 85 |
+
midi_bytes = MIDI.save(generated_midi)
|
| 86 |
+
self.generated_files.append(base64.b64encode(midi_bytes).decode('utf-8'))
|
| 87 |
+
return generated_midi
|
|
|
|
| 88 |
|
| 89 |
def play_midi(self, midi_data):
|
| 90 |
+
"""Renders MIDI data to audio bytes."""
|
| 91 |
midi_bytes = base64.b64decode(midi_data)
|
| 92 |
midi_file = MIDI.load(io.BytesIO(midi_bytes))
|
| 93 |
audio = io.BytesIO()
|
|
|
|
| 95 |
audio.seek(0)
|
| 96 |
return audio
|
| 97 |
|
| 98 |
+
@staticmethod
|
| 99 |
+
def _softmax(x, axis):
|
| 100 |
+
"""Computes softmax probabilities."""
|
| 101 |
+
exp_x = np.exp(x - np.max(x, axis=axis, keepdims=True))
|
| 102 |
+
return exp_x / np.sum(exp_x, axis=axis, keepdims=True)
|
| 103 |
+
|
| 104 |
+
@staticmethod
|
| 105 |
+
def _sample_top_p_k(probs, p, k):
|
| 106 |
+
"""Samples a token using top-p and top-k sampling (simplified)."""
|
| 107 |
+
# Placeholder: replace with actual sampling logic if needed
|
| 108 |
+
return np.array([[np.random.choice(len(probs[0]))]])
|
| 109 |
+
|
| 110 |
+
def process_midi(files):
|
| 111 |
+
"""Processes uploaded MIDI files and yields updates for Gradio components."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
if not files:
|
| 113 |
+
yield [gr.update()] * (1 + 2 * MIDI_OUTPUT_BATCH_SIZE)
|
| 114 |
+
return
|
| 115 |
+
|
| 116 |
+
for idx, file in enumerate(files):
|
| 117 |
+
output_idx = idx % MIDI_OUTPUT_BATCH_SIZE
|
| 118 |
midi_data = midi_processor.load_midi(file.name)
|
| 119 |
+
generated_midi = midi_processor.generate_onnx(midi_data)
|
| 120 |
+
|
| 121 |
+
# Placeholder for MIDI events; in practice, extract from generated_midi
|
| 122 |
+
# Expected format: ["note", delta_time, track, channel, pitch, velocity, duration]
|
| 123 |
+
events = [
|
| 124 |
+
["note", 0, 0, 0, 60, 100, 1000], # Example event
|
| 125 |
+
# Add logic to convert generated_midi to events using tokenizer
|
| 126 |
+
]
|
| 127 |
+
|
| 128 |
+
# Prepare updates list: [js_msg, audio0, midi0, audio1, midi1, ...]
|
| 129 |
+
updates = [gr.update()] * (1 + 2 * MIDI_OUTPUT_BATCH_SIZE)
|
| 130 |
+
|
| 131 |
+
# Clear visualizer
|
| 132 |
+
updates[0] = js_msg.update(value=json.dumps([{"name": "visualizer_clear", "data": [output_idx, "v2"]}]))
|
| 133 |
+
yield updates
|
| 134 |
+
|
| 135 |
+
# Send MIDI events
|
| 136 |
+
updates[0] = js_msg.update(value=json.dumps([{"name": "visualizer_append", "data": [output_idx, events]}]))
|
| 137 |
+
yield updates
|
| 138 |
+
|
| 139 |
+
# Finalize visualizer and update audio/MIDI outputs
|
| 140 |
+
audio_update = midi_processor.play_midi(generated_midi)
|
| 141 |
+
midi_update = gr.File.update(value=generated_midi, label=f"Generated MIDI {output_idx}")
|
| 142 |
+
updates[0] = js_msg.update(value=json.dumps([{"name": "visualizer_end", "data": output_idx}]))
|
| 143 |
+
updates[1 + 2 * output_idx] = audio_update # Audio component
|
| 144 |
+
updates[2 + 2 * output_idx] = midi_update # MIDI file component
|
| 145 |
+
yield updates
|
| 146 |
+
|
| 147 |
+
# Final yield to ensure all components are in a stable state
|
| 148 |
+
yield [gr.update()] * (1 + 2 * MIDI_OUTPUT_BATCH_SIZE)
|
| 149 |
+
|
| 150 |
if __name__ == "__main__":
|
| 151 |
+
parser = argparse.ArgumentParser(description="MIDI Composer App")
|
| 152 |
+
parser.add_argument("--port", type=int, default=7860, help="Server port")
|
| 153 |
+
parser.add_argument("--share", action="store_true", help="Share the app publicly")
|
| 154 |
+
opt = parser.parse_args()
|
| 155 |
|
| 156 |
device_manager = MIDIDeviceManager()
|
| 157 |
midi_processor = MIDIManager()
|
| 158 |
|
| 159 |
+
with gr.Blocks(theme=gr.themes.Soft()) as app:
|
| 160 |
+
# Hidden textbox for sending messages to JS
|
| 161 |
+
js_msg = gr.Textbox(visible=False, elem_id="msg_receiver")
|
| 162 |
+
|
| 163 |
with gr.Tabs():
|
| 164 |
# MIDI Prompt Tab
|
| 165 |
with gr.Tab("MIDI Prompt"):
|
| 166 |
+
midi_upload = gr.File(label="Upload MIDI File(s)", file_count="multiple")
|
| 167 |
+
generate_btn = gr.Button("Generate")
|
| 168 |
status = gr.Textbox(label="Status", value="Ready", interactive=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 169 |
|
| 170 |
+
# Outputs Tab
|
| 171 |
+
with gr.Tab("Outputs"):
|
| 172 |
+
output_audios = []
|
| 173 |
+
output_midis = []
|
| 174 |
+
for i in range(MIDI_OUTPUT_BATCH_SIZE):
|
| 175 |
+
with gr.Column():
|
| 176 |
+
gr.Markdown(f"## Output {i+1}")
|
| 177 |
+
gr.HTML(elem_id=f"midi_visualizer_container_{i}")
|
| 178 |
+
output_audio = gr.Audio(label="Generated Audio", type="bytes", autoplay=True, elem_id=f"midi_audio_{i}")
|
| 179 |
+
output_midi = gr.File(label="Generated MIDI", file_types=[".mid"])
|
| 180 |
+
output_audios.append(output_audio)
|
| 181 |
+
output_midis.append(output_midi)
|
| 182 |
|
| 183 |
# Devices Tab
|
| 184 |
with gr.Tab("Devices"):
|
| 185 |
+
device_info = gr.Textbox(label="Connected MIDI Devices", value=device_manager.get_device_info(), interactive=False)
|
| 186 |
refresh_btn = gr.Button("Refresh Devices")
|
| 187 |
+
refresh_btn.click(fn=lambda: device_manager.get_device_info(), outputs=[device_info])
|
| 188 |
+
|
| 189 |
+
# Define output components for event handling
|
| 190 |
+
outputs = [js_msg] + output_audios + output_midis
|
| 191 |
+
|
| 192 |
+
# Bind the generate button to the processing function
|
| 193 |
+
generate_btn.click(fn=process_midi, inputs=[midi_upload], outputs=outputs)
|
| 194 |
|
| 195 |
+
# Launch the app
|
| 196 |
+
app.launch(server_port=opt.port, share=opt.share, inbrowser=True)
|
| 197 |
device_manager.close()
|