Spaces:
Runtime error
Runtime error
fix: errors
Browse files- .gitattributes +1 -0
- .gitignore +3 -0
- GeneralUserGS.sf2 +3 -0
- app.py +65 -37
- requirements.txt +2 -0
- script.py +103 -0
.gitattributes
CHANGED
@@ -40,3 +40,4 @@ model/saved_model.pb filter=lfs diff=lfs merge=lfs -text
|
|
40 |
model/variables filter=lfs diff=lfs merge=lfs -text
|
41 |
model/variables/variables.data-00000-of-00001 filter=lfs diff=lfs merge=lfs -text
|
42 |
model/variables/variables.index filter=lfs diff=lfs merge=lfs -text
|
|
|
|
40 |
model/variables filter=lfs diff=lfs merge=lfs -text
|
41 |
model/variables/variables.data-00000-of-00001 filter=lfs diff=lfs merge=lfs -text
|
42 |
model/variables/variables.index filter=lfs diff=lfs merge=lfs -text
|
43 |
+
GeneralUserGS.sf2 filter=lfs diff=lfs merge=lfs -text
|
.gitignore
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
generated_music.mid
|
3 |
+
output.wav
|
GeneralUserGS.sf2
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f45b6b4a68b6bf3d792fcbb6d7de24dc701a0f89c5900a21ef3aaece993b839a
|
3 |
+
size 31281186
|
app.py
CHANGED
@@ -3,31 +3,9 @@ from tensorflow.keras.models import load_model
|
|
3 |
import numpy as np
|
4 |
import pretty_midi
|
5 |
import tensorflow as tf
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
def mse_with_positive_pressure(y_true, y_pred):
|
10 |
-
# 通常のMSE
|
11 |
-
mse = tf.reduce_mean(tf.square(y_true - y_pred))
|
12 |
-
|
13 |
-
return mse
|
14 |
-
# # 予測が真の値よりも大きい場合のペナルティ
|
15 |
-
# positive_pressure = tf.reduce_mean(tf.maximum(0.0, y_pred - y_true))
|
16 |
-
|
17 |
-
# # 通常のMSEにペナルティを加える
|
18 |
-
# return mse + positive_pressure
|
19 |
-
|
20 |
-
# model = load_model('model/', custom_objects={'mse_with_positive_pressure': mse_with_positive_pressure})
|
21 |
-
|
22 |
-
# model = load_model('model/')
|
23 |
-
|
24 |
-
# model = load_model('yiwv/music-gen-kit-model/')
|
25 |
-
|
26 |
-
# from huggingface_hub import hf_hub_download
|
27 |
-
|
28 |
-
# model_dir = hf_hub_download(repo_id="yiwv/music-gen-kit-model", revision="main") #, subfolder="model")
|
29 |
-
# model = tf.keras.models.load_model(model_dir)
|
30 |
-
|
31 |
from huggingface_hub import from_pretrained_keras
|
32 |
|
33 |
|
@@ -60,38 +38,88 @@ def predict_next_note(notes, keras_model, temperature=1.0):
|
|
60 |
return int(pitch.numpy()), float(step.numpy()), float(duration.numpy())
|
61 |
|
62 |
|
63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
pm = pretty_midi.PrettyMIDI()
|
65 |
instrument = pretty_midi.Instrument(program=pretty_midi.instrument_name_to_program(instrument_name))
|
66 |
|
|
|
67 |
for note_data in notes:
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
69 |
instrument.notes.append(note)
|
|
|
70 |
|
71 |
pm.instruments.append(instrument)
|
72 |
pm.write(out_file)
|
73 |
return out_file
|
74 |
|
75 |
|
76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
input_sequence = np.fromstring(input_text, sep=',')
|
78 |
temperature = 2.0
|
79 |
num_predictions = 120
|
|
|
|
|
|
|
|
|
80 |
generated_notes = []
|
81 |
prev_start = 0
|
82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
for _ in range(num_predictions):
|
84 |
-
pitch, step, duration = predict_next_note(
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
|
91 |
-
|
92 |
-
notes_to_midi(generated_notes, output_file_name)
|
93 |
|
94 |
-
return output_file_name
|
95 |
|
96 |
|
97 |
# Gradioインターフェースの定義
|
|
|
3 |
import numpy as np
|
4 |
import pretty_midi
|
5 |
import tensorflow as tf
|
6 |
+
import soundfile as sf
|
7 |
+
import fluidsynth
|
8 |
+
import subprocess
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
from huggingface_hub import from_pretrained_keras
|
10 |
|
11 |
|
|
|
38 |
return int(pitch.numpy()), float(step.numpy()), float(duration.numpy())
|
39 |
|
40 |
|
41 |
+
|
42 |
+
def convert_midi_to_wav(midi_path, wav_path, soundfont_path):
|
43 |
+
cmd = ["fluidsynth", "-ni", soundfont_path, midi_path, "-F", wav_path, "-r", "44100"]
|
44 |
+
subprocess.run(cmd)
|
45 |
+
|
46 |
+
|
47 |
+
def notes_to_midi(notes, out_file, instrument_name):
|
48 |
pm = pretty_midi.PrettyMIDI()
|
49 |
instrument = pretty_midi.Instrument(program=pretty_midi.instrument_name_to_program(instrument_name))
|
50 |
|
51 |
+
prev_start = 0
|
52 |
for note_data in notes:
|
53 |
+
pitch = int(note_data[0])
|
54 |
+
step = note_data[1]
|
55 |
+
duration = note_data[2]
|
56 |
+
start = prev_start + step
|
57 |
+
end = start + duration
|
58 |
+
note = pretty_midi.Note(velocity=100, pitch=pitch, start=start, end=end)
|
59 |
instrument.notes.append(note)
|
60 |
+
prev_start = start
|
61 |
|
62 |
pm.instruments.append(instrument)
|
63 |
pm.write(out_file)
|
64 |
return out_file
|
65 |
|
66 |
|
67 |
+
|
68 |
+
def convert_midi_to_wav_with_pretty_midi(midi_path, wav_path):
|
69 |
+
# MIDIファイルを読み込む
|
70 |
+
midi_data = pretty_midi.PrettyMIDI(midi_path)
|
71 |
+
# オーディオ波形を取得する
|
72 |
+
audio_data = midi_data.synthesize()
|
73 |
+
# WAVファイルとして保存する
|
74 |
+
sf.write(wav_path, audio_data.T, 44100)
|
75 |
+
|
76 |
+
|
77 |
+
def generate_music(input_text, instrument_name="Acoustic Grand Piano"):
|
78 |
+
"""
|
79 |
+
input_text example: 60,0.5,0.5,62,0.5,0.5,64,0.5,0.5
|
80 |
+
"""
|
81 |
input_sequence = np.fromstring(input_text, sep=',')
|
82 |
temperature = 2.0
|
83 |
num_predictions = 120
|
84 |
+
|
85 |
+
seq_length = 25
|
86 |
+
vocab_size = 128
|
87 |
+
|
88 |
generated_notes = []
|
89 |
prev_start = 0
|
90 |
|
91 |
+
# 入力テキストを数値のリストに変換
|
92 |
+
input_values = [float(val) for val in input_text.split(",")]
|
93 |
+
|
94 |
+
# 入力データの形状を動的に調整
|
95 |
+
num_notes = len(input_values) // 3
|
96 |
+
if num_notes > 25:
|
97 |
+
input_data = np.zeros((num_notes, 3))
|
98 |
+
else:
|
99 |
+
input_data = np.zeros((25, 3))
|
100 |
+
|
101 |
+
input_data[-num_notes:] = np.array(input_values).reshape(-1, 3)
|
102 |
+
|
103 |
+
# 音楽生成のループ
|
104 |
+
generated_notes = []
|
105 |
for _ in range(num_predictions):
|
106 |
+
pitch, step, duration = predict_next_note(input_data[-25:], model, temperature)
|
107 |
+
generated_notes.append((pitch, step, duration))
|
108 |
+
new_note = np.array([[pitch, step, duration]])
|
109 |
+
input_data = np.vstack([input_data, new_note])
|
110 |
+
|
111 |
+
# 生成されたノートをMIDIファイルに変換
|
112 |
+
generated_notes_array = np.array(generated_notes)
|
113 |
+
output_file_name = "generated_music.mid"
|
114 |
+
notes_to_midi(generated_notes_array, output_file_name, instrument_name)
|
115 |
+
|
116 |
+
try:
|
117 |
+
convert_midi_to_wav(output_file_name, 'output.wav', "GeneralUserGS.sf2")
|
118 |
+
except:
|
119 |
+
convert_midi_to_wav_with_pretty_midi(output_file_name, 'output.wav')
|
120 |
|
121 |
+
return 'output.wav'
|
|
|
122 |
|
|
|
123 |
|
124 |
|
125 |
# Gradioインターフェースの定義
|
requirements.txt
CHANGED
@@ -3,3 +3,5 @@ pyFluidSynth==1.3.2
|
|
3 |
pretty-midi==0.2.10
|
4 |
tensorflow==2.13.0
|
5 |
keras==2.13.1
|
|
|
|
|
|
3 |
pretty-midi==0.2.10
|
4 |
tensorflow==2.13.0
|
5 |
keras==2.13.1
|
6 |
+
huggingface_hub
|
7 |
+
soundfile
|
script.py
ADDED
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from tensorflow.keras.models import load_model
|
3 |
+
import numpy as np
|
4 |
+
import pretty_midi
|
5 |
+
import tensorflow as tf
|
6 |
+
|
7 |
+
from huggingface_hub import from_pretrained_keras
|
8 |
+
|
9 |
+
|
10 |
+
def mse_with_positive_pressure(y_true: tf.Tensor, y_pred: tf.Tensor):
|
11 |
+
mse = (y_true - y_pred) ** 2
|
12 |
+
positive_pressure = 10 * tf.maximum(-y_pred, 0.0)
|
13 |
+
return tf.reduce_mean(mse + positive_pressure)
|
14 |
+
|
15 |
+
|
16 |
+
model = from_pretrained_keras("yiwv/music-gen-kit-model", custom_objects={'mse_with_positive_pressure': mse_with_positive_pressure})
|
17 |
+
|
18 |
+
|
19 |
+
def predict_next_note(notes, keras_model, temperature=1.0):
|
20 |
+
assert temperature > 0
|
21 |
+
inputs = tf.expand_dims(notes, 0)
|
22 |
+
predictions = model.predict(inputs)
|
23 |
+
pitch_logits = predictions['pitch']
|
24 |
+
step = predictions['step']
|
25 |
+
duration = predictions['duration']
|
26 |
+
|
27 |
+
pitch_logits /= temperature
|
28 |
+
pitch = tf.random.categorical(pitch_logits, num_samples=1)
|
29 |
+
pitch = tf.squeeze(pitch, axis=-1)
|
30 |
+
duration = tf.squeeze(duration, axis=-1)
|
31 |
+
step = tf.squeeze(step, axis=-1)
|
32 |
+
|
33 |
+
step = tf.maximum(0, step)
|
34 |
+
duration = tf.maximum(0, duration)
|
35 |
+
|
36 |
+
return int(pitch.numpy()), float(step.numpy()), float(duration.numpy())
|
37 |
+
|
38 |
+
|
39 |
+
def notes_to_midi(notes, out_file, instrument_name):
|
40 |
+
pm = pretty_midi.PrettyMIDI()
|
41 |
+
instrument = pretty_midi.Instrument(program=pretty_midi.instrument_name_to_program(instrument_name))
|
42 |
+
|
43 |
+
prev_start = 0
|
44 |
+
for note_data in notes:
|
45 |
+
pitch = int(note_data[0])
|
46 |
+
step = note_data[1]
|
47 |
+
duration = note_data[2]
|
48 |
+
start = prev_start + step
|
49 |
+
end = start + duration
|
50 |
+
note = pretty_midi.Note(velocity=100, pitch=pitch, start=start, end=end)
|
51 |
+
instrument.notes.append(note)
|
52 |
+
prev_start = start
|
53 |
+
|
54 |
+
pm.instruments.append(instrument)
|
55 |
+
pm.write(out_file)
|
56 |
+
return out_file
|
57 |
+
|
58 |
+
|
59 |
+
|
60 |
+
def generate_music(input_text, instrument_name="Acoustic Grand Piano"):
|
61 |
+
"""
|
62 |
+
input_text example: 60,0.5,0.5,62,0.5,0.5,64,0.5,0.5
|
63 |
+
"""
|
64 |
+
input_sequence = np.fromstring(input_text, sep=',')
|
65 |
+
temperature = 2.0
|
66 |
+
num_predictions = 120
|
67 |
+
|
68 |
+
seq_length = 25
|
69 |
+
vocab_size = 128
|
70 |
+
|
71 |
+
generated_notes = []
|
72 |
+
prev_start = 0
|
73 |
+
|
74 |
+
# 入力テキストを数値のリストに変換
|
75 |
+
input_values = [float(val) for val in input_text.split(",")]
|
76 |
+
|
77 |
+
|
78 |
+
# 入力データの形状を動的に調整
|
79 |
+
num_notes = len(input_values) // 3
|
80 |
+
if num_notes > 25:
|
81 |
+
input_data = np.zeros((num_notes, 3))
|
82 |
+
else:
|
83 |
+
input_data = np.zeros((25, 3))
|
84 |
+
|
85 |
+
input_data[-num_notes:] = np.array(input_values).reshape(-1, 3)
|
86 |
+
|
87 |
+
# 音楽生成のループ
|
88 |
+
generated_notes = []
|
89 |
+
for _ in range(num_predictions):
|
90 |
+
pitch, step, duration = predict_next_note(input_data[-25:], model, temperature)
|
91 |
+
generated_notes.append((pitch, step, duration))
|
92 |
+
new_note = np.array([[pitch, step, duration]])
|
93 |
+
input_data = np.vstack([input_data, new_note])
|
94 |
+
|
95 |
+
# 生成されたノートをMIDIファイルに変換
|
96 |
+
generated_notes_array = np.array(generated_notes)
|
97 |
+
output_file_name = "generated_music.mid"
|
98 |
+
notes_to_midi(generated_notes_array, output_file_name, instrument_name)
|
99 |
+
|
100 |
+
return output_file_name
|
101 |
+
|
102 |
+
|
103 |
+
generate_music("60,0.5,0.5,62,0.5,0.5,64,0.5,0.5")
|