Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -48,7 +48,7 @@ def mode_dur(seq):
|
|
48 |
def mode_pitch(seq):
|
49 |
return statistics.mode([t % 128 for t in seq if 256 < t < 512])
|
50 |
|
51 |
-
|
52 |
|
53 |
train_data = []
|
54 |
|
@@ -58,12 +58,12 @@ for m in tqdm.tqdm(melody_chords_f):
|
|
58 |
|
59 |
for tv in range(-3, 3):
|
60 |
|
61 |
-
|
62 |
-
|
63 |
|
64 |
score = [t+tv if 256 < t < 512 else t for t in m[5]]
|
65 |
|
66 |
-
seq = [916] + [
|
67 |
|
68 |
seq += score
|
69 |
|
@@ -144,61 +144,6 @@ def Generate_POP_Section(input_parsons_code,
|
|
144 |
|
145 |
#===============================================================================
|
146 |
|
147 |
-
print('Instantiating Parsons Code Melody Transformer model...')
|
148 |
-
|
149 |
-
SEQ_LEN = 322
|
150 |
-
PAD_IDX = 392
|
151 |
-
|
152 |
-
model = TransformerWrapper(
|
153 |
-
num_tokens = PAD_IDX+1,
|
154 |
-
max_seq_len = SEQ_LEN,
|
155 |
-
attn_layers = Decoder(dim = 1024,
|
156 |
-
depth = 4,
|
157 |
-
heads = 8,
|
158 |
-
rotary_pos_emb = True,
|
159 |
-
attn_flash = True
|
160 |
-
)
|
161 |
-
)
|
162 |
-
|
163 |
-
model = AutoregressiveWrapper(model, ignore_index = PAD_IDX, pad_value=PAD_IDX)
|
164 |
-
|
165 |
-
print('=' * 70)
|
166 |
-
print('Loading model checkpoint...')
|
167 |
-
|
168 |
-
model_path = 'Parsons_Code_Melody_Transformer_Trained_Model_13786_steps_0.3058_loss_0.8819_acc.pth'
|
169 |
-
|
170 |
-
model.load_state_dict(torch.load(model_path, map_location='cpu'))
|
171 |
-
|
172 |
-
model.cpu()
|
173 |
-
model.eval()
|
174 |
-
|
175 |
-
dtype = torch.bfloat16
|
176 |
-
|
177 |
-
ctx = torch.amp.autocast(device_type='cpu', dtype=dtype)
|
178 |
-
|
179 |
-
print('Done!')
|
180 |
-
print('=' * 70)
|
181 |
-
|
182 |
-
#===============================================================================
|
183 |
-
|
184 |
-
print('Prepping Parsons code string...')
|
185 |
-
|
186 |
-
td_str = re.sub('[^*DRU]', '', input_parsons_code)
|
187 |
-
|
188 |
-
print(len(td_str))
|
189 |
-
print('=' * 70)
|
190 |
-
|
191 |
-
if '*' in td_str and len(td_str) > 1:
|
192 |
-
code_mult = (64 // len(td_str[1:]))+1
|
193 |
-
mult_code = ('*' + (td_str[1:] * code_mult))[:64]
|
194 |
-
|
195 |
-
else:
|
196 |
-
mult_code = '*UUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUUDDDDDDDUUUUUUU'
|
197 |
-
|
198 |
-
pcode = parsons_code_to_tokens(mult_code)
|
199 |
-
|
200 |
-
print('Done!')
|
201 |
-
print('=' * 70)
|
202 |
|
203 |
#===============================================================================
|
204 |
|
@@ -244,31 +189,42 @@ def Generate_POP_Section(input_parsons_code,
|
|
244 |
song_f = []
|
245 |
|
246 |
time = 0
|
247 |
-
dur =
|
248 |
vel = 90
|
249 |
-
pitch =
|
250 |
channel = 0
|
251 |
-
|
252 |
for ss in song:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
|
254 |
-
|
255 |
-
|
256 |
-
time += ss * 32
|
257 |
-
|
258 |
-
if 128 <= ss < 256:
|
259 |
-
|
260 |
-
dur = (ss-128) * 32
|
261 |
-
|
262 |
-
if 256 <= ss < 384:
|
263 |
-
|
264 |
-
pitch = ss-256
|
265 |
-
|
266 |
-
song_f.append(['note', time, dur, channel, pitch, vel, 0])
|
267 |
-
|
268 |
-
fn1 = 'Parsons-Code-Melody-Transformer-Composition'
|
269 |
|
270 |
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
|
271 |
-
output_signature = '
|
272 |
output_file_name = fn1,
|
273 |
track_name='Project Los Angeles'
|
274 |
)
|
@@ -324,15 +280,14 @@ if __name__ == "__main__":
|
|
324 |
|
325 |
with app:
|
326 |
|
327 |
-
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>
|
328 |
-
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Generate unique
|
329 |
gr.Markdown(
|
330 |
-
"
|
331 |
-
"
|
332 |
-
"Check out [Tegridy MIDI Dataset](https://github.com/asigalov61/Tegridy-MIDI-Dataset) on GitHub!\n\n"
|
333 |
)
|
334 |
|
335 |
-
gr.Markdown("##
|
336 |
|
337 |
input_parsons_code = gr.Textbox(label="Parsons code",
|
338 |
info="Make sure your Parsons code starts with *",
|
@@ -340,13 +295,6 @@ if __name__ == "__main__":
|
|
340 |
value="*"
|
341 |
)
|
342 |
|
343 |
-
clr_btn = gr.ClearButton(components=input_parsons_code)
|
344 |
-
|
345 |
-
def reset_pcode():
|
346 |
-
return '*'
|
347 |
-
|
348 |
-
clr_btn.click(reset_pcode, outputs=input_parsons_code)
|
349 |
-
|
350 |
gr.Markdown("## Select generation options:")
|
351 |
|
352 |
input_first_note_duration = gr.Slider(1, 127, value=15, step=1, label="First note duration value")
|
|
|
48 |
def mode_pitch(seq):
|
49 |
return statistics.mode([t % 128 for t in seq if 256 < t < 512])
|
50 |
|
51 |
+
sections_dict = sorted(set([str_strip(s[2]).rstrip('-') for s in melody_chords_f]))
|
52 |
|
53 |
train_data = []
|
54 |
|
|
|
58 |
|
59 |
for tv in range(-3, 3):
|
60 |
|
61 |
+
section = str_strip(m[2])
|
62 |
+
section_tok = sections_dict.index(part)
|
63 |
|
64 |
score = [t+tv if 256 < t < 512 else t for t in m[5]]
|
65 |
|
66 |
+
seq = [916] + [section_tok+512, mode_time(score)+532, mode_dur(score)+660, mode_pitch(score)+tv+788]
|
67 |
|
68 |
seq += score
|
69 |
|
|
|
144 |
|
145 |
#===============================================================================
|
146 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
147 |
|
148 |
#===============================================================================
|
149 |
|
|
|
189 |
song_f = []
|
190 |
|
191 |
time = 0
|
192 |
+
dur = 0
|
193 |
vel = 90
|
194 |
+
pitch = 0
|
195 |
channel = 0
|
196 |
+
|
197 |
for ss in song:
|
198 |
+
|
199 |
+
if 0 <= ss < 128:
|
200 |
+
|
201 |
+
time += ss * 32
|
202 |
+
|
203 |
+
if 128 <= ss < 256:
|
204 |
+
|
205 |
+
dur = (ss-128)* 32
|
206 |
+
|
207 |
+
if 256 <= ss < 512:
|
208 |
+
|
209 |
+
pitch = (ss-256) % 128
|
210 |
+
cha = (ss-256) // 128
|
211 |
+
|
212 |
+
if cha == 0:
|
213 |
+
channel = 3
|
214 |
+
vel = 110
|
215 |
+
patch = 40
|
216 |
+
|
217 |
+
else:
|
218 |
+
channel = 0
|
219 |
+
vel = 80
|
220 |
+
patch = 0
|
221 |
+
|
222 |
+
song_f.append(['note', time, dur, channel, pitch, vel, patch ])
|
223 |
|
224 |
+
fn1 = 'Popular-Hook-Transformer-Composition'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
|
226 |
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
|
227 |
+
output_signature = 'Popular Hook Transformer',
|
228 |
output_file_name = fn1,
|
229 |
track_name='Project Los Angeles'
|
230 |
)
|
|
|
280 |
|
281 |
with app:
|
282 |
|
283 |
+
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Popular Hook Transformer</h1>")
|
284 |
+
gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Generate unique POP music sections</h1>")
|
285 |
gr.Markdown(
|
286 |
+
"This is a demo for popular-hook MIDI Dataset\n\n"
|
287 |
+
"Check out [popular-hook](https://huggingface.co/datasets/NEXTLab-ZJU/popular-hook) on Hugging Face!\n\n"
|
|
|
288 |
)
|
289 |
|
290 |
+
gr.Markdown("## Select generation options:")
|
291 |
|
292 |
input_parsons_code = gr.Textbox(label="Parsons code",
|
293 |
info="Make sure your Parsons code starts with *",
|
|
|
295 |
value="*"
|
296 |
)
|
297 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
298 |
gr.Markdown("## Select generation options:")
|
299 |
|
300 |
input_first_note_duration = gr.Slider(1, 127, value=15, step=1, label="First note duration value")
|