Update app.py
Browse files
app.py
CHANGED
|
@@ -2,76 +2,73 @@ import gradio as gr
|
|
| 2 |
import librosa
|
| 3 |
import numpy as np
|
| 4 |
from pydub import AudioSegment, effects
|
| 5 |
-
import uuid, os, json
|
| 6 |
|
| 7 |
# =====================================================
|
| 8 |
-
#
|
| 9 |
# =====================================================
|
| 10 |
|
|
|
|
|
|
|
| 11 |
UPLOAD_DIR = "uploads"
|
| 12 |
OUTPUT_DIR = "outputs"
|
| 13 |
-
DATA_DIR = "data"
|
| 14 |
|
| 15 |
os.makedirs(UPLOAD_DIR, exist_ok=True)
|
| 16 |
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
| 17 |
-
os.makedirs(DATA_DIR, exist_ok=True)
|
| 18 |
-
|
| 19 |
-
MODEL_PATH = os.path.join(DATA_DIR, "weights.json")
|
| 20 |
|
| 21 |
# =====================================================
|
| 22 |
-
# MODELO IA
|
| 23 |
# =====================================================
|
| 24 |
|
| 25 |
-
|
| 26 |
-
"
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
}
|
| 31 |
|
| 32 |
-
def
|
| 33 |
-
if os.path.exists(
|
| 34 |
try:
|
| 35 |
-
with open(
|
| 36 |
return json.load(f)
|
| 37 |
except:
|
| 38 |
pass
|
| 39 |
-
return
|
| 40 |
|
| 41 |
-
def
|
| 42 |
-
with open(
|
| 43 |
-
json.dump(
|
| 44 |
|
| 45 |
-
|
|
|
|
| 46 |
|
| 47 |
# =====================================================
|
| 48 |
-
# ANÁLISIS MUSICAL
|
| 49 |
# =====================================================
|
| 50 |
|
| 51 |
def analyze_audio(path):
|
| 52 |
y, sr = librosa.load(path, mono=True)
|
| 53 |
|
| 54 |
onset_env = librosa.onset.onset_strength(y=y, sr=sr)
|
| 55 |
-
tempo = float(librosa.beat.tempo(
|
| 56 |
-
onset_envelope=onset_env, sr=sr
|
| 57 |
-
)[0])
|
| 58 |
|
| 59 |
rms = librosa.feature.rms(y=y)[0]
|
| 60 |
rms = librosa.util.normalize(rms)
|
| 61 |
|
| 62 |
-
drops = [
|
| 63 |
-
|
| 64 |
-
if rms[i] > np.mean(rms[i-10:i-1]) * 1.4
|
| 65 |
-
|
| 66 |
|
| 67 |
-
return {
|
| 68 |
-
"tempo": tempo,
|
| 69 |
-
"energy": rms,
|
| 70 |
-
"drops": drops
|
| 71 |
-
}
|
| 72 |
|
| 73 |
# =====================================================
|
| 74 |
-
#
|
| 75 |
# =====================================================
|
| 76 |
|
| 77 |
def score_transition(a, b, t):
|
|
@@ -87,44 +84,35 @@ def score_transition(a, b, t):
|
|
| 87 |
)
|
| 88 |
|
| 89 |
# =====================================================
|
| 90 |
-
# TIME-STRETCH
|
| 91 |
# =====================================================
|
| 92 |
|
| 93 |
def bpm_adjust(segment, from_bpm, to_bpm):
|
| 94 |
-
"""
|
| 95 |
-
Estira TODA la canción (voz, bajo, ritmo)
|
| 96 |
-
Solo durante transición
|
| 97 |
-
Cambio fuerte pero controlado (±20%)
|
| 98 |
-
"""
|
| 99 |
if from_bpm <= 0 or to_bpm <= 0:
|
| 100 |
return segment
|
| 101 |
|
| 102 |
ratio = from_bpm / to_bpm
|
| 103 |
-
|
| 104 |
if abs(1 - ratio) > 0.20:
|
| 105 |
return segment
|
| 106 |
|
| 107 |
new_rate = int(segment.frame_rate * ratio)
|
| 108 |
-
|
| 109 |
-
stretched = segment._spawn(
|
| 110 |
segment.raw_data,
|
| 111 |
overrides={"frame_rate": new_rate}
|
| 112 |
-
)
|
| 113 |
-
|
| 114 |
-
return stretched.set_frame_rate(segment.frame_rate)
|
| 115 |
|
| 116 |
# =====================================================
|
| 117 |
-
# IA DJ MIXER
|
| 118 |
# =====================================================
|
| 119 |
|
| 120 |
def auto_dj_mix(files, durations, crossfade_sec):
|
| 121 |
-
log = "🎧 Iniciando MONX
|
| 122 |
yield log, None
|
| 123 |
|
| 124 |
durs = [float(x.strip()) for x in durations.split(",")]
|
| 125 |
crossfade_ms = int(crossfade_sec * 1000)
|
| 126 |
|
| 127 |
-
tracks, analyses = [], []
|
| 128 |
|
| 129 |
for i, f in enumerate(files):
|
| 130 |
log += f"\n🔍 Analizando canción {i+1}\n"
|
|
@@ -150,70 +138,84 @@ def auto_dj_mix(files, durations, crossfade_sec):
|
|
| 150 |
mix = segment
|
| 151 |
continue
|
| 152 |
|
| 153 |
-
prev = analyses[i
|
| 154 |
-
curr = analyses[i]
|
| 155 |
|
| 156 |
log += f"\n🧠 Decidiendo transición {i} → {i+1}\n"
|
| 157 |
yield log, None
|
| 158 |
|
| 159 |
candidates = range(5, min(len(prev["energy"]) - 1, 50))
|
| 160 |
best_t = max(candidates, key=lambda t: score_transition(prev, curr, t))
|
|
|
|
|
|
|
| 161 |
|
| 162 |
target_bpm = curr["tempo"]
|
| 163 |
|
| 164 |
-
log += (
|
| 165 |
-
f"🔄 Sincronizando TODO al BPM {target_bpm:.1f}\n"
|
| 166 |
-
f" (voz, bajos, ritmo)\n"
|
| 167 |
-
)
|
| 168 |
-
yield log, None
|
| 169 |
-
|
| 170 |
mix_adj = bpm_adjust(mix, prev["tempo"], target_bpm)
|
| 171 |
seg_adj = bpm_adjust(segment, curr["tempo"], target_bpm)
|
| 172 |
|
| 173 |
safe_cf = min(crossfade_ms, len(mix_adj), len(seg_adj))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 174 |
|
| 175 |
-
|
| 176 |
-
|
|
|
|
|
|
|
| 177 |
|
| 178 |
-
|
|
|
|
|
|
|
| 179 |
|
| 180 |
-
|
| 181 |
-
|
|
|
|
|
|
|
| 182 |
|
| 183 |
-
|
| 184 |
|
| 185 |
-
|
| 186 |
-
|
| 187 |
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
)
|
| 192 |
|
| 193 |
-
|
|
|
|
|
|
|
|
|
|
| 194 |
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
for k in weights:
|
| 199 |
-
weights[k] +=
|
| 200 |
-
save_weights(weights)
|
| 201 |
|
| 202 |
-
|
| 203 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
|
| 205 |
# =====================================================
|
| 206 |
# UI
|
| 207 |
# =====================================================
|
| 208 |
|
| 209 |
-
with gr.Blocks(title="MONX
|
| 210 |
gr.Markdown(
|
| 211 |
-
""
|
| 212 |
-
<
|
| 213 |
-
<h1>🎚️ MONX AI DJ</h1>
|
| 214 |
-
<p>IA DJ automática con BPM dinámico PRO</p>
|
| 215 |
-
</div>
|
| 216 |
-
"""
|
| 217 |
)
|
| 218 |
|
| 219 |
files = gr.File(
|
|
@@ -222,27 +224,21 @@ with gr.Blocks(title="MONX AI DJ") as demo:
|
|
| 222 |
file_types=[".mp3", ".wav", ".flac", ".m4a"]
|
| 223 |
)
|
| 224 |
|
| 225 |
-
durations = gr.Textbox(
|
| 226 |
-
|
| 227 |
-
value="90,90"
|
| 228 |
-
)
|
| 229 |
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
step=1,
|
| 234 |
-
label="Crossfade (segundos)"
|
| 235 |
-
)
|
| 236 |
|
| 237 |
-
btn
|
| 238 |
|
| 239 |
-
|
| 240 |
-
|
|
|
|
|
|
|
| 241 |
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
inputs=[files, durations, crossfade],
|
| 245 |
-
outputs=[status, output]
|
| 246 |
-
)
|
| 247 |
|
| 248 |
demo.launch()
|
|
|
|
| 2 |
import librosa
|
| 3 |
import numpy as np
|
| 4 |
from pydub import AudioSegment, effects
|
| 5 |
+
import uuid, os, json
|
| 6 |
|
| 7 |
# =====================================================
|
| 8 |
+
# ARCHIVO DE ESTADO (MISMA CARPETA QUE app.py)
|
| 9 |
# =====================================================
|
| 10 |
|
| 11 |
+
STATE_PATH = "monx_dj_state.json"
|
| 12 |
+
|
| 13 |
UPLOAD_DIR = "uploads"
|
| 14 |
OUTPUT_DIR = "outputs"
|
|
|
|
| 15 |
|
| 16 |
os.makedirs(UPLOAD_DIR, exist_ok=True)
|
| 17 |
os.makedirs(OUTPUT_DIR, exist_ok=True)
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
# =====================================================
|
| 20 |
+
# ESTADO DEL MODELO (IA REAL)
|
| 21 |
# =====================================================
|
| 22 |
|
| 23 |
+
DEFAULT_STATE = {
|
| 24 |
+
"weights": {
|
| 25 |
+
"bpm": 0.35,
|
| 26 |
+
"energy": 0.30,
|
| 27 |
+
"smooth": 0.20,
|
| 28 |
+
"drop_penalty": 0.15
|
| 29 |
+
},
|
| 30 |
+
"avg_score": 0.5,
|
| 31 |
+
"runs": 0
|
| 32 |
}
|
| 33 |
|
| 34 |
+
def load_state():
|
| 35 |
+
if os.path.exists(STATE_PATH):
|
| 36 |
try:
|
| 37 |
+
with open(STATE_PATH, "r") as f:
|
| 38 |
return json.load(f)
|
| 39 |
except:
|
| 40 |
pass
|
| 41 |
+
return DEFAULT_STATE.copy()
|
| 42 |
|
| 43 |
+
def save_state(state):
|
| 44 |
+
with open(STATE_PATH, "w") as f:
|
| 45 |
+
json.dump(state, f, indent=2)
|
| 46 |
|
| 47 |
+
state = load_state()
|
| 48 |
+
weights = state["weights"]
|
| 49 |
|
| 50 |
# =====================================================
|
| 51 |
+
# ANÁLISIS MUSICAL
|
| 52 |
# =====================================================
|
| 53 |
|
| 54 |
def analyze_audio(path):
|
| 55 |
y, sr = librosa.load(path, mono=True)
|
| 56 |
|
| 57 |
onset_env = librosa.onset.onset_strength(y=y, sr=sr)
|
| 58 |
+
tempo = float(librosa.beat.tempo(onset_envelope=onset_env, sr=sr)[0])
|
|
|
|
|
|
|
| 59 |
|
| 60 |
rms = librosa.feature.rms(y=y)[0]
|
| 61 |
rms = librosa.util.normalize(rms)
|
| 62 |
|
| 63 |
+
drops = [
|
| 64 |
+
i for i in range(10, len(rms) - 10)
|
| 65 |
+
if rms[i] > np.mean(rms[i-10:i-1]) * 1.4
|
| 66 |
+
]
|
| 67 |
|
| 68 |
+
return {"tempo": tempo, "energy": rms, "drops": drops}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
# =====================================================
|
| 71 |
+
# FUNCIÓN DE DECISIÓN (IA)
|
| 72 |
# =====================================================
|
| 73 |
|
| 74 |
def score_transition(a, b, t):
|
|
|
|
| 84 |
)
|
| 85 |
|
| 86 |
# =====================================================
|
| 87 |
+
# TIME-STRETCH
|
| 88 |
# =====================================================
|
| 89 |
|
| 90 |
def bpm_adjust(segment, from_bpm, to_bpm):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
if from_bpm <= 0 or to_bpm <= 0:
|
| 92 |
return segment
|
| 93 |
|
| 94 |
ratio = from_bpm / to_bpm
|
|
|
|
| 95 |
if abs(1 - ratio) > 0.20:
|
| 96 |
return segment
|
| 97 |
|
| 98 |
new_rate = int(segment.frame_rate * ratio)
|
| 99 |
+
return segment._spawn(
|
|
|
|
| 100 |
segment.raw_data,
|
| 101 |
overrides={"frame_rate": new_rate}
|
| 102 |
+
).set_frame_rate(segment.frame_rate)
|
|
|
|
|
|
|
| 103 |
|
| 104 |
# =====================================================
|
| 105 |
+
# IA DJ MIXER
|
| 106 |
# =====================================================
|
| 107 |
|
| 108 |
def auto_dj_mix(files, durations, crossfade_sec):
|
| 109 |
+
log = "🎧 Iniciando MONX DJ (IA)\n"
|
| 110 |
yield log, None
|
| 111 |
|
| 112 |
durs = [float(x.strip()) for x in durations.split(",")]
|
| 113 |
crossfade_ms = int(crossfade_sec * 1000)
|
| 114 |
|
| 115 |
+
tracks, analyses, scores = [], [], []
|
| 116 |
|
| 117 |
for i, f in enumerate(files):
|
| 118 |
log += f"\n🔍 Analizando canción {i+1}\n"
|
|
|
|
| 138 |
mix = segment
|
| 139 |
continue
|
| 140 |
|
| 141 |
+
prev, curr = analyses[i-1], analyses[i]
|
|
|
|
| 142 |
|
| 143 |
log += f"\n🧠 Decidiendo transición {i} → {i+1}\n"
|
| 144 |
yield log, None
|
| 145 |
|
| 146 |
candidates = range(5, min(len(prev["energy"]) - 1, 50))
|
| 147 |
best_t = max(candidates, key=lambda t: score_transition(prev, curr, t))
|
| 148 |
+
best_score = score_transition(prev, curr, best_t)
|
| 149 |
+
scores.append(best_score)
|
| 150 |
|
| 151 |
target_bpm = curr["tempo"]
|
| 152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
mix_adj = bpm_adjust(mix, prev["tempo"], target_bpm)
|
| 154 |
seg_adj = bpm_adjust(segment, curr["tempo"], target_bpm)
|
| 155 |
|
| 156 |
safe_cf = min(crossfade_ms, len(mix_adj), len(seg_adj))
|
| 157 |
+
mix = bpm_adjust(
|
| 158 |
+
mix_adj.append(seg_adj, crossfade=safe_cf),
|
| 159 |
+
target_bpm, curr["tempo"]
|
| 160 |
+
)
|
| 161 |
|
| 162 |
+
out_path = os.path.join(
|
| 163 |
+
OUTPUT_DIR, f"monx_dj_mix_{uuid.uuid4().hex}.m4a"
|
| 164 |
+
)
|
| 165 |
+
mix.export(out_path, format="ipod", codec="aac", bitrate="192k")
|
| 166 |
|
| 167 |
+
# =========================
|
| 168 |
+
# APRENDIZAJE AUTOMÁTICO
|
| 169 |
+
# =========================
|
| 170 |
|
| 171 |
+
if scores:
|
| 172 |
+
new_avg = sum(scores) / len(scores)
|
| 173 |
+
old_avg = state["avg_score"]
|
| 174 |
+
lr = 0.05
|
| 175 |
|
| 176 |
+
reward = 1 if new_avg > old_avg else -1
|
| 177 |
|
| 178 |
+
for k in weights:
|
| 179 |
+
weights[k] += reward * lr * abs(weights[k])
|
| 180 |
|
| 181 |
+
total = sum(abs(v) for v in weights.values())
|
| 182 |
+
for k in weights:
|
| 183 |
+
weights[k] /= total
|
|
|
|
| 184 |
|
| 185 |
+
state["avg_score"] = new_avg
|
| 186 |
+
state["runs"] += 1
|
| 187 |
+
state["weights"] = weights
|
| 188 |
+
save_state(state)
|
| 189 |
|
| 190 |
+
log += "\n✅ Mix listo. Da feedback para que MONX DJ aprenda."
|
| 191 |
+
yield log, out_path
|
| 192 |
+
|
| 193 |
+
# =====================================================
|
| 194 |
+
# FEEDBACK HUMANO
|
| 195 |
+
# =====================================================
|
| 196 |
+
|
| 197 |
+
def feedback(reward):
|
| 198 |
+
lr = 0.08
|
| 199 |
for k in weights:
|
| 200 |
+
weights[k] += reward * lr * abs(weights[k])
|
|
|
|
| 201 |
|
| 202 |
+
total = sum(abs(v) for v in weights.values())
|
| 203 |
+
for k in weights:
|
| 204 |
+
weights[k] /= total
|
| 205 |
+
|
| 206 |
+
state["weights"] = weights
|
| 207 |
+
save_state(state)
|
| 208 |
+
|
| 209 |
+
return "🧠 Feedback recibido. MONX DJ ha aprendido."
|
| 210 |
|
| 211 |
# =====================================================
|
| 212 |
# UI
|
| 213 |
# =====================================================
|
| 214 |
|
| 215 |
+
with gr.Blocks(title="MONX DJ") as demo:
|
| 216 |
gr.Markdown(
|
| 217 |
+
"<h1 style='text-align:center'>🎚️ MONX DJ</h1>"
|
| 218 |
+
"<p style='text-align:center'>IA DJ con aprendizaje real</p>"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
)
|
| 220 |
|
| 221 |
files = gr.File(
|
|
|
|
| 224 |
file_types=[".mp3", ".wav", ".flac", ".m4a"]
|
| 225 |
)
|
| 226 |
|
| 227 |
+
durations = gr.Textbox(label="Duración por canción (seg)", value="90,90")
|
| 228 |
+
crossfade = gr.Slider(6, 20, value=12, step=1, label="Crossfade (seg)")
|
|
|
|
|
|
|
| 229 |
|
| 230 |
+
btn = gr.Button("🔥 Auto Mix")
|
| 231 |
+
status = gr.Markdown()
|
| 232 |
+
output = gr.Audio(type="filepath")
|
|
|
|
|
|
|
|
|
|
| 233 |
|
| 234 |
+
btn.click(auto_dj_mix, [files, durations, crossfade], [status, output])
|
| 235 |
|
| 236 |
+
gr.Markdown("### ¿Te gustó el mix?")
|
| 237 |
+
like = gr.Button("👍 Sí")
|
| 238 |
+
dislike = gr.Button("👎 No")
|
| 239 |
+
fb_status = gr.Markdown()
|
| 240 |
|
| 241 |
+
like.click(lambda: feedback(1), None, fb_status)
|
| 242 |
+
dislike.click(lambda: feedback(-1), None, fb_status)
|
|
|
|
|
|
|
|
|
|
| 243 |
|
| 244 |
demo.launch()
|