hephaestusonfire commited on
Commit
0968d2e
·
verified ·
1 Parent(s): 057df1e

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +25 -46
app.py CHANGED
@@ -1,7 +1,7 @@
1
  """
2
  Gradio app for classical music generation with a grid-based Transformer.
3
- Fully optimized with a client-side HTML/JS matrix, robust hidden JSON state bridges,
4
- and an absolute in-memory FluidSynth audio synthesis engine for Hugging Face Spaces.
5
  """
6
 
7
  import os
@@ -10,9 +10,8 @@ import tempfile
10
  from typing import Tuple
11
  import gradio as gr
12
  import numpy as np
13
- import fluidsynth
14
  import pretty_midi
15
- from scipy.io import wavfile
16
 
17
  import config
18
  from inference import generate_midi_file
@@ -135,7 +134,6 @@ function() {
135
  let cellClass = `btn-${family}-${row}-${col}`;
136
  let buttons = document.querySelectorAll('.' + cellClass);
137
 
138
- // Flip internal visual arrays tracking memory
139
  let currentVal = window.studioGrids[family][row][col];
140
  let newVal = currentVal === 0 ? 1 : 0;
141
  window.studioGrids[family][row][col] = newVal;
@@ -149,7 +147,6 @@ function() {
149
  }
150
  });
151
 
152
- // Push the fully synchronized multi-dimensional array directly to hidden state modules
153
  let wrapperId = `json_hidden_${family}`;
154
  let container = document.getElementById(wrapperId);
155
  if (container) {
@@ -194,12 +191,11 @@ def execute_pipeline(strings_json: str, woodwinds_json: str, brass_json: str, pe
194
  midi_path = os.path.join(tmpdir, "generated.mid")
195
  audio_path = os.path.join(tmpdir, "generated.wav")
196
 
197
- # Safe parsing loops to guarantee structural grid sizing matching core layouts
198
  def load_safe_grid(json_str, target_rows):
199
  if not json_str or json_str.strip() == "":
200
- return np.zeros((target_rows, config.INPUT_STEPS, 1), dtype=np.float32)
201
  try:
202
- parsed = np.array(json.loads(json_str), dtype=np.float32) # [Rows, Steps]
203
  if parsed.shape != (target_rows, config.INPUT_STEPS):
204
  return np.zeros((target_rows, config.INPUT_STEPS), dtype=np.float32)
205
  return parsed
@@ -213,7 +209,7 @@ def execute_pipeline(strings_json: str, woodwinds_json: str, brass_json: str, pe
213
 
214
  def to_backend_layout(numpy_grid, r_dims):
215
  reshaped = numpy_grid.reshape(r_dims, config.INPUT_BARS, config.STEPS_PER_BAR)
216
- return np.transpose(reshaped, (1, 2, 0)) # Transposes back to expected format [Bars, Steps, Features]
217
 
218
  s_seed = to_backend_layout(s_arr, config.PITCHES_MELODIC)
219
  w_seed = to_backend_layout(w_arr, config.PITCHES_MELODIC)
@@ -226,35 +222,18 @@ def execute_pipeline(strings_json: str, woodwinds_json: str, brass_json: str, pe
226
  midi_out_path=midi_path, device="cpu", primary_family=lead
227
  )
228
 
229
- # ABSOLUTE IN-MEMORY FLUIDSYNTH SYNTHESIS ENGINE
230
  try:
231
- pm = pretty_midi.PrettyMIDI(midi_path)
232
- sample_rate = 44100
233
-
234
- # Instantiate FluidSynth core instance context block
235
- fs = fluidsynth.Synth()
236
- sfid = fs.sfload("soundfont.sf2")
237
- fs.program_select(0, sfid, 0, 0)
238
-
239
- # Calculate maximum sample array size limits safely based on total generated track time
240
- duration = pm.get_end_time()
241
- total_samples = int(duration * sample_rate) + (sample_rate * 2)
242
-
243
- # Allocate flat block arrays
244
- synth_buffer = np.zeros(total_samples, dtype=np.float32)
245
-
246
- # Instruct FluidSynth to process sequential MIDI events directly into data loops
247
- # This bypasses all external platform sound card system layers
248
- audio_data = pm.fluidsynth(fs=sample_rate)
249
 
250
- if len(audio_data) > 0:
251
- # Enforce track volume signal scaling safety limits
252
- audio_data = audio_data / np.max(np.abs(audio_data))
253
- wavfile.write(audio_path, sample_rate, (audio_data * 32767).astype(np.int16))
254
- else:
255
- wavfile.write(audio_path, 44100, np.zeros(44100, dtype=np.int16))
256
  except Exception as e:
257
- print(f"[Synth Rendering Fail Exception Context]: {e}")
 
 
258
  wavfile.write(audio_path, 44100, np.zeros(44100, dtype=np.int16))
259
 
260
  return midi_path, (audio_path if os.path.exists(audio_path) else None)
@@ -265,6 +244,12 @@ def execute_pipeline(strings_json: str, woodwinds_json: str, brass_json: str, pe
265
 
266
  with gr.Blocks(css=CUSTOM_DAW_CSS, title="Classical Music Grid Generator") as demo:
267
 
 
 
 
 
 
 
268
  gr.HTML("""
269
  <div style="text-align:center; padding:15px 0;">
270
  <h1 style="font-size:2.2rem; font-weight:800; background:linear-gradient(135deg, #63b3ed 0%, #9b5de5 50%, #fc814a 100%); -webkit-background-clip:text; -webkit-text-fill-color:transparent; margin:0 0 5px 0;">
@@ -294,12 +279,6 @@ with gr.Blocks(css=CUSTOM_DAW_CSS, title="Classical Music Grid Generator") as de
294
  </div>
295
  """)
296
 
297
- # Hidden textareas acting as continuous 2D structured JSON bridges to Python
298
- strings_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_strings")
299
- woodwinds_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_woodwinds")
300
- brass_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_brass")
301
- perc_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_perc")
302
-
303
  # Interactive Grid Displays Mapping Sections
304
  with gr.Row():
305
  with gr.Column(scale=1):
@@ -312,11 +291,11 @@ with gr.Blocks(css=CUSTOM_DAW_CSS, title="Classical Music Grid Generator") as de
312
  with gr.Row():
313
  with gr.Column(scale=3):
314
  with gr.Row():
315
- target_minutes = gr.Slider(minimum=1.0, maximum=5.0, value=2.0, step=0.5, label="🕐 Target Length (Minutes)")
316
- temperature = gr.Slider(minimum=0.7, maximum=1.3, value=1.0, step=0.05, label="🌡️ Creativity Temperature")
317
  with gr.Row():
318
- top_k = gr.Slider(minimum=2, maximum=12, value=6, step=1, label="🎲 Top-K Distribution Filter")
319
- primary_family = gr.Dropdown(choices=["strings", "woodwinds", "brass"], value="strings", label="🎼 Primary Orchestration Lead")
320
 
321
  generate_button = gr.Button("🎵 Execute Transformer Music Generation Pipeline", variant="primary")
322
 
 
1
  """
2
  Gradio app for classical music generation with a grid-based Transformer.
3
+ Utilizes a native HTML/JS matrix selection console coupled with a stable
4
+ midi2audio system path renderer to guarantee perfect playback on Hugging Face.
5
  """
6
 
7
  import os
 
10
  from typing import Tuple
11
  import gradio as gr
12
  import numpy as np
 
13
  import pretty_midi
14
+ from midi2audio import FluidSynth
15
 
16
  import config
17
  from inference import generate_midi_file
 
134
  let cellClass = `btn-${family}-${row}-${col}`;
135
  let buttons = document.querySelectorAll('.' + cellClass);
136
 
 
137
  let currentVal = window.studioGrids[family][row][col];
138
  let newVal = currentVal === 0 ? 1 : 0;
139
  window.studioGrids[family][row][col] = newVal;
 
147
  }
148
  });
149
 
 
150
  let wrapperId = `json_hidden_${family}`;
151
  let container = document.getElementById(wrapperId);
152
  if (container) {
 
191
  midi_path = os.path.join(tmpdir, "generated.mid")
192
  audio_path = os.path.join(tmpdir, "generated.wav")
193
 
 
194
  def load_safe_grid(json_str, target_rows):
195
  if not json_str or json_str.strip() == "":
196
+ return np.zeros((target_rows, config.INPUT_STEPS), dtype=np.float32)
197
  try:
198
+ parsed = np.array(json.loads(json_str), dtype=np.float32)
199
  if parsed.shape != (target_rows, config.INPUT_STEPS):
200
  return np.zeros((target_rows, config.INPUT_STEPS), dtype=np.float32)
201
  return parsed
 
209
 
210
  def to_backend_layout(numpy_grid, r_dims):
211
  reshaped = numpy_grid.reshape(r_dims, config.INPUT_BARS, config.STEPS_PER_BAR)
212
+ return np.transpose(reshaped, (1, 2, 0))
213
 
214
  s_seed = to_backend_layout(s_arr, config.PITCHES_MELODIC)
215
  w_seed = to_backend_layout(w_arr, config.PITCHES_MELODIC)
 
222
  midi_out_path=midi_path, device="cpu", primary_family=lead
223
  )
224
 
225
+ # HIGH-PERFORMANCE LINUX CONTAINER SYNTHESIS PIPELINE
226
  try:
227
+ # Absolute structural path pointing straight to the global packages.txt system soundfont
228
+ soundfont_path = "/usr/share/sounds/sf2/FluidR3_GM.sf2"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
229
 
230
+ # Instantiate FluidSynth wrapper targeting the shared library context directly
231
+ fs = FluidSynth(soundfont_path=soundfont_path)
232
+ fs.midi_to_audio(midi_path, audio_path)
 
 
 
233
  except Exception as e:
234
+ print(f"[FluidSynth Wrapper Failure Exception]: {e}")
235
+ # Build safety empty wave chunk if rendering fails
236
+ from scipy.io import wavfile
237
  wavfile.write(audio_path, 44100, np.zeros(44100, dtype=np.int16))
238
 
239
  return midi_path, (audio_path if os.path.exists(audio_path) else None)
 
244
 
245
  with gr.Blocks(css=CUSTOM_DAW_CSS, title="Classical Music Grid Generator") as demo:
246
 
247
+ # Hidden textareas acting as continuous 2D structured JSON bridges to Python
248
+ strings_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_strings")
249
+ woodwinds_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_woodwinds")
250
+ brass_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_brass")
251
+ perc_json = gr.Textbox(value="", visible=False, elem_id="json_hidden_perc")
252
+
253
  gr.HTML("""
254
  <div style="text-align:center; padding:15px 0;">
255
  <h1 style="font-size:2.2rem; font-weight:800; background:linear-gradient(135deg, #63b3ed 0%, #9b5de5 50%, #fc814a 100%); -webkit-background-clip:text; -webkit-text-fill-color:transparent; margin:0 0 5px 0;">
 
279
  </div>
280
  """)
281
 
 
 
 
 
 
 
282
  # Interactive Grid Displays Mapping Sections
283
  with gr.Row():
284
  with gr.Column(scale=1):
 
291
  with gr.Row():
292
  with gr.Column(scale=3):
293
  with gr.Row():
294
+ target_minutes = gr.Slider(minimum=1.0, maximum=5.0, value=2.0, step=0.5, label="Target Playback Length (Minutes)")
295
+ temperature = gr.Slider(minimum=0.7, maximum=1.3, value=1.0, step=0.05, label="Creativity Temperature")
296
  with gr.Row():
297
+ top_k = gr.Slider(minimum=2, maximum=12, value=6, step=1, label="Top-K Distribution Filter")
298
+ primary_family = gr.Dropdown(choices=["strings", "woodwinds", "brass"], value="strings", label="Melodic Ensemble Driver Lead")
299
 
300
  generate_button = gr.Button("🎵 Execute Transformer Music Generation Pipeline", variant="primary")
301